Ten weeks of Just Started ends with this post.

The Substack was the journal of a specific stretch — the last leg of building the first working version of The Retirement Strategy with AI, and the launch. This is the retrospective on that. Not on the product. On what the practice taught me about becoming an AI power user.

Most people I talk to use AI regularly and would not call themselves powerful with it. They are the audience for this piece.

The gap between using AI and being powerful with it is not a gap of hours logged. Some people log a lot of hours and stay in the same place. The gap is what you do with the hours.

The short answer: you have to play with the tool, and you have to play with it differently every week.

Just Started, and why it isn’t actionable

The Substack is called Just Started because that is what I did. Sat down with the tool. Asked it something. Kept going. That is the true story of how it started for me.

It is not actionable advice for most people.

I have told plenty of people to just start, and most of them don’t. Not because they are lazy. Because just start does not tell them what to do. Sit down and ask it what? Try it on which piece of your work? Open the empty prompt box and stare?

The actionable version is closer to this: pick one thing you were going to do this week — an email, a memo, a decision, a spreadsheet — and try doing it with the tool instead of the way you were going to. Not to prove anything. Just to see what shows up.

That specific first try is what just start means when it means something. And it is only the beginning. What determines whether you actually get anywhere is what happens after.

The hump

There is one real climb to becoming a power user. I will call it the hump. It has three specific moves inside it, and if you do not make them, you never really start playing at all — you look at the tool from the outside.

Talk to it, don’t Google it. Most people ask AI things the way they’d ask a search engine. Short queries. Short answers. The tool answers back the way you talk to it. If you want a colleague, you have to write to it like a colleague. Paragraphs. Context. What you have already tried. Where you actually want to end up. Give it that, and it gives you back a different kind of output.

Bring context, then bring more. AI without context is a smart stranger. AI with context is a smart collaborator who has been in the work with you for months. Projects, memory, custom instructions, uploaded files, prior threads — all of these turn the stranger into the collaborator. Setting it up feels like friction the first few times. Then it disappears.

The code barrier is gone. This one stops more people than the other two combined. The belief that you need to be technical to build things with AI. You don’t. AI writes the code. You describe what you want. If you can articulate what you want clearly, you can build. That fact hasn’t been said clearly enough or often enough to most people.

Get past those three and you have crossed the hump. Now the path is flat.

The flat path

Past the hump, the terrain does not get steeper. It gets wider. There is no next climb. Just a long, flat plane with things to try all over it.

What determines what happens next is curiosity.

The specific voice of it is two questions. I wonder if I could…? What if I tried…? Asked constantly. Followed up on. Those two questions, aimed at the tool most weeks, are the difference between someone who becomes a power user and someone who plateaus at competent.

They are not techniques. They are a habit of mind. Most people, once they are past the hump, stop asking them — usually without noticing it. They settle into three or four use cases they return to. Same tool. Same prompts. Same shape of task. Comfortable competence.

That is fine, and it is not what makes a power user.

What makes a power user is asking those two questions week after week, and letting the answers pull you off the groove. Four kinds of variety show up most:

Know what you want, sharply. AI amplifies the direction you point it. Fuzzy aim, polished nothing. Sharp aim, real acceleration. Play with sharpening your own aim before you play with prompting the tool.

Try many things fast, kill fast. The single biggest habit of power users is that they iterate through many maybes rather than executing on one right answer. AI collapses the cost of prototyping. Use that. Do not defend an approach past the moment it should be defended.

Build new things, not just smarter versions of old things. Most people use AI to build faster versions of what already exists — a slicker form, a nicer PDF, a quicker email. That is fine. It is not the frontier. The frontier is what could only exist because AI is embedded inside it. Every so often, force yourself to try building something that could not have existed a year ago.

Teach it back. Your work generates data. Feed that data back to the tool — through custom instructions, through updated context, through review of what worked and what didn’t. Close the loop. Every week you don’t is a week the tool stays the same.

These are four variations on one theme. Try something you haven’t tried before. Do it deliberately. Do it often.

What blocks the curiosity

The biggest block is the mood.

Pew Research published a survey in June that puts a number on it. Only 16% of Americans think AI will have a positive impact on society over the next 20 years. 40% expect it to be negative. Two-thirds think the technology is moving too fast. Younger adults are more skeptical than older ones — the reverse of what you might expect.

That is the mood, and the mood is a curiosity killer. When most of the people around you think a tool is bad, or reckless, or moving too fast, I wonder if I could…? and What if I tried…? get quieter. They feel less permitted. Less serious. Skepticism does not stop AI from being powerful. It stops the user from getting curious enough to become powerful with it.

The individual blocks I see most often are posture, not skill:

• Fear of looking dumb. There is nobody to look dumb in front of.

• Not knowing what to try. The tool is your work — start there.

• Sticking with the interface you first used. Try a different model. Try a different mode.

• Waiting for a use case to arrive. Use cases don’t arrive. You find them by trying things.

• Waiting for permission. Nobody is giving it.

What actually changes

After months of playing with variety, the biggest change is not what I built. It is how I think.

My relationship with information, with problem-solving, with what is possible to make in an afternoon — all of it has shifted. Some questions I would have spent a week on take an hour. Some artifacts I would have hired someone to make take a morning. Some problems that would have been too big now feel obviously doable.

The product is the artifact. The change is the point.

Where the work goes next

Ten weeks of writing Just Started has changed how I think more than any of the specific tools or techniques I picked up along the way.

The mood is against being curious about AI right now. That is exactly when being curious pays the most.

The next thread of writing starts later this summer. Different question. Same posture.

Dennis

Step 7 ended with the realization that embedded AI gets sharper in two ways.

The first way happens on its own. Every few months a more capable reasoning model comes out, and the AI inside the experience steps up with it — better at conversation, better at reasoning, better at picking up what someone is actually saying. That improvement is real and it requires nothing from me.

The second way is the one this post is about. The AI inside the apps could get sharper about this specifically — about retirement readiness, about this cohort, about the questions that keep coming up and the responses that have been working and the ones that have not. The intelligence the model brought was general. The work it was doing was specific. There was a way to close that gap.

This is the step I am still in the middle of, so I want to be honest about what I have done and what I have not.

What I have done is build the first version of the loop. The cohort’s conversations with the assistant generate transcripts. The transcripts are read — not by the live assistant in real time, but by a separate process that runs weekly. That process clusters the conversations by topic, identifies the recurring patterns, and produces a brief on what the cohort is actually using the assistant for and where the assistant is falling short. The brief is the input to the next round of system-prompt updates and surface refinements.

The AI is not training itself in any technical sense. The model weights are not changing. The model is doing what it has always done. What’s changing is the context the model is operating inside — what it knows about this cohort, what patterns it has been pointed at, what responses have been tuned in and out.

But its behavior is changing, week over week, because the data it is generating is being used to make it sharper.

That is the loop. It is crude. It is manual. It is real.

The Sunday the loop closed for the first time

This was a Sunday in spring of 2026. I had been getting cohort feedback through a combination of email replies, occasional calls, and the assistant transcripts I had started reading by hand. Reading by hand was not going to scale. The cohort was small but already producing thousands of messages per week.

I asked the AI to read its own transcripts.

Specifically: given the past week of cohort conversations with the Plan assistant, produce a brief on what the cohort is using the assistant for, what is working, and what is not. Be specific. Quote where useful. Cluster by topic. Tell me where the assistant is being asked to do things it cannot do well. Tell me where the assistant is producing responses the user clearly did not find useful.

It identified four patterns I had not seen, two of which were obvious in retrospect, two of which I would not have noticed on my own. It quoted specific user phrasings that showed up across multiple conversations. It flagged three places where the assistant was producing wrong-shaped responses. It pointed to one entire category of question — about the timing of retirement decisions — that the assistant kept punting on, when it should have been engaging.

I wrote the next round of system-prompt updates against what the brief had told me. The next week, the same analysis ran again. The patterns from the prior week showed up less. New ones appeared. The brief had pointed me at things I needed pointed out, and the assistant improved inside one week.

This was the first time the work I was producing was being used by the system the work was producing. The AI was not autonomously improving. I was the bridge. But the bridge could be walked, repeatedly, faster than I could improve the system by guessing.

The artifact

The artifact is not a single thing. It is a weekly process. The transcripts feed a brief. The brief feeds the prompt updates. The prompt updates change the assistant. The next week’s transcripts reflect the change. The week after that, the next round of brief, prompt, change.

It is not technically recursive — the AI is not modifying its own weights, and given the model architecture available to me, it cannot. But it is functionally recursive at the level that matters: the product improves because of how it is being used, on a cadence shorter than any feedback loop I have run in my career.

If the artifact has to be named, it is the brief itself. The thing that did not exist before — a weekly written reflection on how the AI inside the apps was being used, what was working, what was not — is now an artifact the AI produces about its own behavior. That artifact is what makes the loop walkable.

What it unlocked

The pace of improvement changed. Before the loop existed, I would notice things in conversations I happened to read, hold them in my head, and remember to make changes when I sat down to update the system. The lag between noticing and shipping was weeks, sometimes longer. With the loop in place, the lag is days. The version the cohort uses today is meaningfully different from the version they used three weeks ago, and it is different in directions their own usage pointed at.

The cohort is teaching the system without knowing they are teaching it. They are using it. The use leaves a trail. The trail is now being read. The reading changes the next version. They are not aware of any of this, and that is how it should be — the use should feel the same, only sharper.

Something else shifted: my confidence in my own judgment. When I update the system now, I am not working from intuition about what users probably want. I am working from a brief that quotes the cohort back to me. The brief is not always right, and I do not always agree with it. But the decisions are happening on a foundation of evidence I did not previously have. The work is more grounded.

Where it does not work yet

The honest answer is that the loop is small and slow and limited to the surfaces I have built. The brief reads conversations; it does not yet read the deeper signal — what users do after a conversation, whether the actions in their Plan actually got executed, whether the dimensions they scored low on shifted over months. The behavioral telemetry exists. Most of it is not yet feeding the loop.

A bigger limit: the AI inside the loop is the same AI inside the apps. There is no separation between the model that does the work and the model that learns from the work. In a more developed version of this, the learning function would be its own thing, with its own data, its own audit, its own permissions. I do not have that yet. What I have is a process that depends on me sitting between the data and the system, doing the synthesis and authoring the updates.

There is also the question of what happens when the brief sees something I do not. If the brief tells me a pattern is showing up, and I disagree with the pattern, whose read of the data is right? I am operating on the assumption that I am, for now, because my context is broader than the assistant’s. That assumption gets weaker every month the system runs.

The frontier is not just technical. It is what the relationship between the maker and the system becomes when the system is being shaped by its own use.

This was the eighth step. I don’t know what the ninth is. The eight-step staircase is what I climbed. From here, the work is no longer a staircase. It is a loop, and the loop is running.

A question for anyone reading this. What would your work learn from you if you let it? What data are you generating that no system is reading?

Dennis

Step 6 ended with a working system. The cascade of maybes had produced something a person could open, walk through, and come out the other side with a profile across six dimensions and the beginnings of a plan. It was real. It worked. I could have shipped it.

Step 7 is when I noticed I had been about to ship the wrong thing.

This was March 2026. The launch date I had picked for the first version of The Retirement Strategy was March 31.

The first week of March I had been away on an international business trip, a full week of meetings and travel that pulled me entirely off the build. I came back with distance I had not had before. A week of not looking at it had been enough to let me see it again as a stranger would.

What I saw was a competent, decent, soon-to-ship retirement-readiness product.

It was also, if I was honest, not particularly different in kind from the products already in the market.

There are several retirement-planning tools out there today. Some are pretty. Some are automated. Some are very detailed. They walk a user through a sequence of dimensions, or questions, or considerations. They produce a status read. They suggest a set of next steps.

The underlying structure is the same in every case.

They are forms. Multiple choice, check boxes, the occasional open-response field. Fundamentally a list of questions.

A glorified form is still a form. The artifact is a record of the answers. What it does not capture is the conversation around the answering.

I had spent weeks of nights and weekends building a smarter form. The AI had been the builder — it had helped me design the surface, write the interview, score the dimensions, draft the report. But the artifact I had built was, structurally, the same kind of artifact those other tools had built.

That was when I had the next maybe.

The maybe that changed the shape

Maybe the AI didn’t have to just live in the toolchain. Maybe it could also be embedded in the product itself. Maybe the AI wasn’t only the builder. Maybe it was also the surface.

The difference is harder to see from outside the work than it is from inside it. Built with AI meant the AI was in the toolchain — helping me make the thing. Built with embedded AI meant the AI was also in the experience itself — the user talking to it, the user’s words part of the artifact. The toolchain placement didn’t go away. The new placement was where the product became a different kind of thing.

Two things follow from that decision that did not follow from the smarter-form decision.

The first is that today is the worst the product will ever be. A form released today is approximately as good as it will ever be — the questions are fixed, the flow is fixed, the math is fixed. A product with AI embedded inside it gets sharper as the models underneath it get sharper, as the prompts get refined, as the use accumulates. The slope of the product over time is different.

The second is that the system captures something a form cannot capture. When a user finishes a dimension on a form, the record is: user said 7 out of 10 on Health. When a user finishes a dimension inside a conversation, the record is: user said 7 out of 10 on Health, and here is what they said about why, here is what they pushed back on, here is the specific worry they named when the AI asked what they were actually afraid of. The score is one number. The conversation around the score is everything else.

I made the call.

The remainder of March was the work of rebuilding what I had so the AI sat at the center of the experience, not just inside the toolchain that produced it. The interview, the scoring moment, the closing screen — all of it had to be reworked to put the AI in the room with the user rather than behind a curtain.

The product that launched on Tuesday, March 31 — version one of TRS — was not the product I had been about to launch when I left for the trip three weeks earlier.

April 1 was my last day at Dell.

What it became

The artifact was no longer a smarter form. It was a conversation. The user was not filling something out. The user was talking to something. The data the system was collecting was not a column of scores. It was language — in the user’s own words, about what they were actually wrestling with.

That is what embedded AI gave the product that built with AI never could.

The next maybe

Once embedded AI was running, there were two ways the product could keep getting sharper.

The first was the model itself. Every few months, a more capable reasoning model would come out, and the AI inside the experience would just become more capable with it — better at conversation, better at reasoning, better at picking up on what someone was actually saying. That improvement happened on its own, without me touching anything.

The second was specific to this domain. A general-purpose model could have an excellent conversation about retirement on day one. What I started to want was a way for it to get sharper at this in particular — at the patterns this cohort was showing, at the questions that kept coming up, at the responses that had been working and the ones that had not. The intelligence the model brought was general. The work it was doing was specific. There was a way to close that gap.

Maybe the AI could get smarter — not just in general, but about this.

That is Step 8. I will get to it next week.

There is a version of this story that is about AI features inside products. That is true and it is not the part that matters.

The part that matters is the realization that the AI didn’t have to just live in the toolchain that built the product. It could also live inside the product itself. That is a different product. That product has a different slope, a different data shape, and a different ceiling.

Denni

s

By Step 5 the AI was producing strategies, briefs, calendars, decks. All of it better than I would have written it, faster than I would have produced it. All of it still paper. The categories were the categories I had specified my whole executive career.

Step 6 is where the categories opened. Not in a flash. In a sequence of small maybes.

For some time before this, I had been working with Claude on my own retirement readiness plan. Not as a product. As a personal exercise. I was still at Dell, still running the day job, and April 1 — the day I would actually leave — was on the horizon. I was preparing for the transition the way I knew how: by getting clear on where I actually stood before I had to figure it out in real time.

The arrangement was simple. I had pulled together the research and popular thought around happiness and longevity across the six dimensions — Health, Wealth, People, Place, Passion, Purpose — and used that material to take an honest look at where I was on each one against where I wanted to be a year or two ahead. Claude was the research partner. I asked it to summarize the literature. I asked it to challenge my self-assessments. I asked it to push back when I was being soft on myself. That was the whole arrangement. It was useful, but it was still me running the show. The AI was a librarian. I was the questioner.

Then I had a thought.

Maybe Claude could use that knowledge to interview me, instead of me interviewing Claude. Start with Health. Hand it the research. Tell it the dimension. Have it ask me where I actually am — directly, conversationally — and use what I say to assess where I sit against what the literature would call ready versus exposed.

That was the first maybe. It was not a leap. It was an inversion. Same knowledge, same dimension, same AI — just facing the other way. Instead of me being the questioner, the AI was. Instead of the AI being a librarian, it was an interviewer.

We tried it. It worked. The interview produced a more honest read on Health than the self-assessment had. That made the next thought thinkable.

So I had the next thought.

Maybe I could standardize this. Build the same interview process across all six dimensions. One conversation per dimension. One way to come out the other end with a coherent read on where I am.

That was the second maybe. Now Claude was not just an interviewer for Health. It was an interviewer for a system of six. The AI’s role had shifted again.

Then the next.

Maybe I could automate the whole thing. Have Claude run the six interviews, score each one against the literature, and pull the six scores together into a high-quality status report. Not a spreadsheet. A document a person could read and recognize themselves in.

Third maybe. The AI was now an interviewer, a scorer, and a writer of a coherent multi-dimensional report. None of those roles had been on the table when this started.

Then the last one.

Maybe the report doesn’t have to be a document at all. Maybe the whole thing could be an app. Something a person opens, walks through, and comes out the other side with their own profile — and eventually, the beginnings of a real plan.

That was the maybe that took the work off paper.

What had actually changed

Read in reverse, none of these maybes was a flash of insight. Each one was a reasonable next thought from the position the last one had put me in. The knowledge base made the interview thinkable. The interview made the score thinkable. The score made the multi-dimensional report thinkable. The report made the app thinkable.

What had changed across the four maybes was not the AI. It was what I was willing to ask the AI to be.

In the first arrangement, the AI was a research assistant. That role was familiar. Every executive I had ever known had worked with somebody like that. Easy to imagine. Easy to specify.

By the fourth maybe, the AI was the interviewer, the scorer, the multi-dimensional analyst, and the surface a real person would eventually meet. None of those roles were ones I would have thought to specify on day one. I had not arrived at them through vision. I had arrived at them by letting each successful step show me what the next step could be.

I want you to notice what that is and what it isn’t. It isn’t a moment of breakthrough. It is a willingness to keep asking, after each maybe worked, what the next maybe could be. The cascade is the thing. The cascade is what “anything” actually meant. Not anything in a flash — anything across a sequence.

What it changed

The cascade I had just walked through was the pattern. Once I had seen it work, I could not unsee it. Every time I sat down to build the next piece of TRS, the question I started with shifted: not what should this artifact be, but what is the next maybe?

What I had been missing was not the technology. The technology was waiting. What I had been missing was the next maybe. I had been thinking inside the categories I knew — documents, decks, sheets — and stopping there.

Where it broke down

The apps work. The interviews work. The scoring works. The reports read well. But the apps still need me to know what is supposed to happen on the inside of them. The AI builds the surface. I tell the AI what the surface is for.

The next maybe, the one I am inside now, is whether the AI can make some of those calls itself — not just to build the experience, but to participate in it. To notice things. To recommend. To be a participant in the conversation rather than the architect of an empty room.

That is Step 7. I will get to it next week.

There is a version of this story that is about technical capability — the AI can build apps now, look at what is possible. That is true and it is not the part that matters.

The part that matters is what the cascade did to the question I was asking. The first maybe sounded ambitious to me. By the fourth, it sounded obvious. The size of the question I was willing to ask had moved. The categories of things I was willing to imagine had grown. The constraint had never really been the tool. It had been my willingness to ask whether the tool could be the next thing.

A question for anyone reading this. What is the maybe you have already had once and never let yourself say out loud? What is the next role you have not yet asked the AI to play?

That is the maybe you have not yet tried.

Dennis

Step 4 ended where last week’s post ended: clean artifacts on demand, from sharp briefs. Faster output. Better drafts. Less laundering.

What I didn’t notice for a few weeks was that the work hadn’t really gotten lighter. I was producing more, but I was also specifying more, and the specifying — every brief, every angle, every form decision — was its own piece of writing. The drafting work had moved out of my head. The strategy work had stayed in.

The plan was carrying me, and I was the only one carrying the plan.

The morning I asked for the strategy instead of the post

This was still late 2025, last operating role. There was a topic I cared about — a perspective on our industry I’d been putting on LinkedIn one post at a time, in isolation. Some posts were landing. Some weren’t. I had a sense of where I was trying to go but no document that captured it.

I went into the project for the executive communications consultant — same specialist, same persona, same knowledge files including the three high-performing posts I’d added in Step 4. But this time I didn’t ask for a post.

I asked for a comms strategy.

Specifically: given my role, my audience, the perspective I’ve been building, and the posts in the project knowledge, draft a complete communications strategy for the next quarter. Treat LinkedIn as one channel among several. Tell me what the goal is, who we’re trying to reach, what the through-line of the messaging is, and how the channels work together.

What came back was a real strategy document. Not a post outline. A page and a half of structured thinking — audience segments, message pillars, channel role definitions, sequencing principles, success metrics. Some of it was right. Some of it was generic. The audience analysis was sharper than I would have expected; the metrics were soft and needed work.

I didn’t accept it. I worked it. I told it where I disagreed, what I wanted to add, what I thought it had wrong. It revised. I pushed back again. After three or four passes, what was on the screen was something I would have signed off on if a real comms team had brought it to me.

I saved it into the project’s knowledge files. I called it the strategy. I closed the conversation.

The next morning, the part that surprised me

I came back the next day with a different ask. I told the AI to use the strategy in the project knowledge as the reference. Then I said: develop the LinkedIn portion of this strategy. Treat the strategy as the brief. Produce what you think the LinkedIn part of executing this strategy should look like.

I assumed I would get posts. That’s what the LinkedIn portion of the strategy had said it would include. Posts, scheduled, in a reasonable cadence.

What I got was not just posts.

What I got was:

• A calendar for the next several weeks, laying out what would publish when.

• Narrative arcs — not a list of post topics, but threaded story arcs that each spanned multiple posts and built toward specific ideas I wanted to land.

• Post sequences, grouped by arc, each one positioned in the calendar.

• Specific post timing, by day and by time window, based on when my audience was most likely to read.

• A posting rhythm that balanced frequency against fatigue, with deliberate pauses and tonal variation.

• Bylined article companions for the strongest arcs — extending the ideas into long-form where the LinkedIn format couldn’t hold them.

• A target publication list for those articles, ranked by fit and by reach within my intended audience.

• Outreach emails to the editors at those publications, written in the register an editor would expect.

I had asked for the LinkedIn portion of a strategy. I had assumed I’d get posts.

It blew me away.

I had not asked for a calendar. I had not asked for narrative arcs. I had not asked for timing windows or posting rhythm. I had not asked for bylined articles or a target publication list or the outreach emails that would go with them.

The AI had read the strategy and decided, on its own, that those were the pieces a competent comms execution would produce. Not as disconnected deliverables — as a coherent plan, structured the way a real comms team would have structured it, with the through-lines between the pieces already drawn.

It was right on most of it. I pushed back on some of the timing, adjusted a couple of arc framings, and cut one of the target publications as a bad fit. Everything else landed as proposed.

What had actually changed

Nothing technical had changed about the AI. It was the same model, in the same project, with the same persona instructions and the same knowledge files. What had changed was how I was specifying the work.

In Step 4, I was giving the AI deliverables — a post, this length, this audience, this angle.In Step 5, I gave it an objective — execute the LinkedIn portion of this strategy. I didn’t specify the form. The AI chose the forms it thought would best serve the objective. The cognitive load shift in Step 4 was drafting → editing. In Step 5 it became editing → directing.

Step 5’s artifact is the strategy document itself — the page-and-a-half I’d worked through with the AI and saved into project knowledge before any of the deliverables got produced. That artifact is what made everything else possible. What you make at one step becomes the brief for the work at the next.

Where it broke down

Here is the new friction at the end of Step 5, the thing that pulled me toward Step 6.

The package was good. The strategy was real. The artifacts conformed. But everything I’d produced — even the surprises — was still in a format I would have produced anyway. A calendar, in a doc. Posts, as text. A presentation, as a deck. Articles, as drafts. Emails, as emails.

What I started to want, especially as I got more comfortable describing objectives instead of deliverables, was for the AI to make things in categories I hadn’t asked for at all. Not just to produce a calendar — to produce something that behaved like a calendar but was alive, updatable, queryable. Not just to produce a doc — to produce something that did what the doc was trying to do, but in a form that didn’t have the limits of a doc.

I had been working with the AI as if its only outputs were the kinds of artifacts that fit on my desk. The friction was that some of what I now wanted to make didn’t fit on a desk.

That’s Step 6. I’ll get to it next week.

There is a version of this story that’s about delegation. He learned to delegate to the AI.That’s true, but it’s not the part that matters.

The part that matters is what got delegated. It wasn’t the work. It was the choice of what work to do. For my entire career, the people I worked with executed against deliverables I had specified — even when the deliverables were sound, the act of specifying them was mine. The strategy, the sequencing, the choice of what artifacts to produce in service of which objectives — that was the part of the work I had always carried alone.

Step 5 was the first time I handed any of that to a system that wasn’t a person.

A question for anyone reading this. What instruction have you been giving the AI that you could replace with a goal? What deliverable have you been specifying when what you really care about is the outcome the deliverable was supposed to serve?

That’s the brief you haven’t written yet.

Dennis

The specialists were sharp. The output was not yet share-worthy.

I’d built six friends in Step 3, each with its own job. The strategist talked like a strategist. The CIO talked like a CIO. The executive communications consultant could mark up a draft, tell me what was working, hand back a rewrite of a clunky paragraph. The conversations were the best I’d ever had at a desk by myself.

But what came out of those conversations almost always needed work before it was usable. Not the ideas — the ideas were often good. The artifacts.

Anyone who has asked an AI for a formatted document in the last two years knows what I’m talking about. You ask for a memo and get back something that reads like it was written in a markdown editor and pasted straight into the window. Headers set in hash marks. Bullets that render differently depending on where you paste them. Tables that look like tables in the chat and like nothing at all the second you move them into Word or Google Docs. You ask for a deck outline and get back a document. You ask for an email and get back an essay with “Subject:” on the first line and no sign-off. You ask for it in a specific format and the AI says it can do that format, and then produces something that is technically in that format and visually an embarrassment.

Every artifact needed a laundering step. Copy from the chat. Clean up the markdown. Reformat the tables. Fix the headers. Rewrite the parts that sounded like an AI had written them even though what you wanted was something that didn’t. By the time the thing was share-worthy, I had done half the work myself.

So I kept treating the AI like a conversation partner, not a production tool. The artifacts I cared about — the ones that went to other people — I still wrote myself, with the conversation as input. It was faster than writing cold. But it wasn’t the shift I could sense was possible if only the output could be trusted.

The morning the comms consultant produced something I didn’t have to clean up

This was late 2025, still in my last operating role. I had a topic I wanted to put on LinkedIn — a perspective on part of our industry I’d been thinking about for months.

I went into the project I had built for the executive communications consultant. Same specialist as before, same persona brief, same knowledge files about my voice and my context. But this time I asked for something specific: a finished post, in my voice, roughly 250 words, no preamble, no sign-off, formatted as a single block of prose with line breaks where I’d want them on LinkedIn. Ready to paste.

What came back was not ready to paste. The voice was wrong. The structure was generic. The opening line read like every other AI-written LinkedIn post I’d ever seen. I started typing a critique and stopped halfway through, because I realized I was about to spend ten minutes describing what my voice was supposed to sound like to an AI that had no examples to learn from.

So I did something I hadn’t tried before. I went and pulled my three highest-performing LinkedIn posts from the past year, copied them verbatim, and dropped them into the project’s knowledge files. Three posts I had written myself, in my actual voice, that had actually worked. Then I went back and asked for the new post again, with a brief addition: use the three posts in the project knowledge as the reference for voice, structure, and pacing.

What came back the second time was 80% right. There were two clunky sentences and the opening was still a touch generic. I told it what to fix. The third draft was 95%. I edited two words and pasted it.

Start to finish, maybe twenty minutes — most of it spent finding and copying my own posts. No markdown. No laundering. No opening a separate document to clean things up.

That had never happened before.

It was not that the AI had suddenly gotten smarter. It was that I had given it the right material. With a sharp brief, a specific format request, and — most importantly — examples from my own work, the output could come back share-worthy. Not perfect. Not finished. But clean enough that my job was editing, not translating.

Once I had that experience, I started looking for it everywhere. What other artifacts could I get back clean?

The answer turned out to be: most prose artifacts, with enough briefing. A memo. A pre-read. A one-page summary. A follow-up email. A meeting agenda. These were the formats the AI handled well, because they were structurally prose — paragraphs, maybe a list or two, some headers. The markdown didn’t get in the way of prose.

What I still had to launder — decks, anything visual, anything with real tables or layout — I left for later. Those were the artifacts where the AI would insist it could produce what I’d asked for and then produce something that looked like a violation of basic design taste. I went back to building those myself. For a while.

What changed when the artifact came back clean

For the next several weeks I kept doing the same thing with prose artifacts. Different topics, same pattern: brief the specialist sharply, ask for the finished artifact, iterate two or three times, paste. The cognitive load shifted. Writing an artifact required holding its structure in my head until I got it out of it. Editing a clean draft from the AI is a smaller, lighter, faster kind of thinking — every editor knows it’s easier to fix a draft than to write one. What I had stumbled into was a way to give myself a clean draft, on demand, in the formats I most often needed. From there, the work was editing, and editing is fast.

What the AI still couldn’t make cleanly

I should be honest about the boundary as I experienced it at this stage.

The artifacts the specialists produced cleanly at Step 4 were the prose-shaped ones. A post. A memo. An email. A one-pager. A narrative pre-read. The form didn’t change — I was still producing the same kinds of documents I’d been producing for years. What changed was who made the first draft and how much work it took to get the draft into share-worthy shape.

The artifacts that still required me to do the laundering were everything else. Slides. Visual dashboards. Anything with a real table rather than a markdown approximation of one. Anything where the format itself was the medium rather than a wrapper around prose. The AI would say it could produce them. It produced them. They looked bad. I went back to making those myself.

Step 4 is the small but consequential move from the AI helps me make a thing to the AI makes the thing, cleanly enough to share, in the formats where the AI can do that today. Those formats happened to be mostly prose. The rest came later, and it’s further up the staircase.

Step 4’s artifact is the clean draft itself — produced by the AI, owned by me, available the moment the conversation ended, usable without a laundering step in between. There’s a ceiling on what that gets you, and I started to bump into it within a week of working this way.

Where it broke down

Here is the new friction at the end of Step 4, the thing that pulled me toward Step 5.

I was producing more, and producing better, and producing cleanly. But I was still asking for one artifact at a time. A post on Monday. A memo on Tuesday. A pre-read on Wednesday. Each request was a small specification: here’s the topic, here’s the audience, here’s the form, here’s the angle.

And while I was doing that, I was carrying something else in my head that the AI had no access to: the strategy the artifacts were serving.

The post on Monday was meant to land an idea I’d been working out for weeks. The memo on Tuesday was meant to set up a conversation I knew was coming on Friday. The pre-read on Wednesday was meant to nudge a decision in a direction I had spent months thinking about. None of that was in the AI’s hands. All of it was in mine. So I had to walk into every session and brief the artifact in isolation, even though the artifact was part of a larger plan.

What I started to want was for the AI to have the plan. Not just the artifact’s shape. The strategy.

If the AI had the strategy, I wouldn’t have to spec each artifact one at a time. I could ask for a series of artifacts that serve the strategy — and trust the AI to figure out what each one should be.

That’s Step 5. I’ll get to it next week.

Faster drafts, more output, fewer hours. That’s the productivity argument, and all of it is true. But what I think about, when I think back to those first weeks of getting share-worthy artifacts out the door without a laundering step, is something quieter.

I had been treating the drafting as if it were the valuable work. The thinking, the framing, the choices about what mattered — those were the valuable work. The drafting was the place where the thinking became visible to me. As soon as I could get a clean draft from somewhere else, the thinking happened in editing, where it had probably belonged all along.

The labor I had been doing was not the labor I had thought I was doing. The actual labor was the editorial judgment, the yes-this-no-not-that, the call about what the audience needed to hear and how. That work was still mine. It was always mine. The drafting had been getting in its way.

A question for anyone reading this.

What artifact have you been making by hand — because the AI’s first pass always needs laundering — that might come back clean if you briefed it sharply enough? What share-worthy thing are you still carrying from conversation to document and cleaning up yourself?

That’s the artifact you haven’t asked for yet.

Dennis

The friend was good. The problem was that one friend was being asked to be everything.

In the operational role I had at the time, I was using my new profile-equipped AI for almost everything that crossed my desk. Drafting a memo. Working through a strategic question. Editing something for a colleague. Talking through a personal financial decision after work. Thinking about a health question on a Sunday afternoon.

Same friend. Same context file. Same paragraph at the top of every conversation about who I was and what I cared about.

The friend was patient. The friend was smart. The friend would do whatever I asked.

But the friend was a generalist. Asked to be a strategist on Tuesday morning, an editor on Tuesday afternoon, a financial advisor on Tuesday evening, and a doctor’s-office-translator on Tuesday night. The same overhead at the start of every conversation: here’s the new role I want you to play today.

I started to notice it most when the work got harder. The strategy work got vaguer when it followed an editing session, because some part of the friend was still in line-edit mode. The financial work hedged in a way that felt like consultant-speak rather than CIO directness. The medical questions got the bedside-manner answer instead of the here’s what the data says answer. Different jobs wanted different voices. The single friend couldn’t be all of them at once without dilution.

The morning I built six friends

This was still pre-product, well before The Retirement Strategy. I was at my last operating role. The different tools all had a version of the same feature — projects, tasks, gems, workspaces, call them what you will. The idea is the same: set up a specific instance of the AI with its own instructions, its own background files, its own focus. A profile per role, on top of the global profile that described me.

One morning I built six.

Three for work:

• An executive communications consultant. Drafts and edits to my voice. Knows the difference between a pre-read for a small group and a kickoff for a large one. Hard on filler. Strict about what gets cut.

• A chief of staff. Operational sequencing, pre-reads, follow-through. Asks the second-order question. Schedules the audit I would have skipped.

• A chief strategy officer. Sees the chessboard. Pushes back when I’m describing a tactic as if it were a strategy.

Three for personal:

• A chief investment officer. Reads my financial situation and the broader market with the same eye. No hedging. No “you should probably consult a professional” at the end of every answer.

• A general counsel. Explains the risk in a contract or a decision the way a real GC explains it to a CEO — plainly, with implications named.

• A chief medical officer. Translates a doctor’s note, a lab result, a treatment option. Not bedside-manner reassurance. Real interpretation.

Each one got a persona-level briefing — who they were, what their job was, the standard they were holding to. Each one inherited the broader profile about me — my situation, my role, my goals. So every conversation started in the same shared world, but with a different specialist in the chair across from me.

Here’s the part that still surprises me when I think about it: I didn’t write the briefings from scratch. I asked the generalist to help. I want an assistant that plays the role of a chief strategy officer in the context of my work. Draft a project name, a description, an instruction set, and suggest what files I should make available to it.

The generalist I’d built in Step 2 became the scaffolding for the six specialists I built in Step 3. Each step made the next step available. That’s the pattern I keep noticing as I climb this staircase — each step isn’t a leap, it’s a small move that the previous step made possible.

The first time I used them, the work felt different in a way I hadn’t anticipated.

It wasn’t that the answers got smarter. It was that the answers got cleaner. The strategist sounded like a strategist. The CIO sounded like a CIO. The general counsel didn’t soften the bad news. Each conversation stayed in its lane, and the lanes were narrower and deeper than the single-friend setup had ever produced.

The thing I didn’t expect: parallel conversations

The unexpected gift of having six friends instead of one was that I could talk to several of them at the same time.

Not literally simultaneous — but close. I could have a strategy question open with the CSO in one tab, a draft with the executive communications consultant in another, and a financial check-in with the CIO in a third. Each conversation persisted independently. Each one had its own thread, its own working context, its own ongoing project.

I was no longer doing one piece of work at a time. I was doing three or four pieces of work in parallel, with three or four specialists, and switching between them the way you’d switch between a 1:1 with your strategy lead, a chat with your communications head, and a ten-minute check with your CFO. Different conversations, different paces, different stakes.

That’s not how I had ever worked before. Not at any company, at any level. The teamworked that way — multiple specialists with multiple workstreams running in parallel, coordinated by me. But I had never had access to that kind of working pattern as a single person, sitting at a single desk.

This is the thing I want to be specific about. The leverage wasn’t in any single answer. The leverage was in the architecture. Six narrow specialists, each with persistent context, working on different problems in parallel, each available the moment I had a question.

What an artifact actually is, again

Step 2 introduced the idea of an artifact — something the model produced that lived outside the conversation. The profile was the first one.

Step 3’s artifact is a different kind of thing: not a single document, but a role. Each specialist friend is, structurally, an artifact. They persist between sessions. They carry their own background. They have a stable identity. You can hand them a piece of work and they remember who they are while doing it.

A role is an artifact in the same way a job description is an artifact. It outlasts the moment of its creation. It shapes the work. It tells the inhabitant — human or otherwise — what they are supposed to be doing and why.

This is what I think the staircase is really about, by the way. Each step is a way of making the AI into a more durable, more usable, more structural piece of how you work. Conversation persists for the length of a session. A profile persists for the length of a relationship. A role persists for the length of a project. Each step is more durable than the last.

We are not yet at the end of that progression.

Where it broke down

Here is the new friction at the end of Step 3, the thing that pulled me toward Step 4.

The friends were good. The friends were specialized. The friends could work in parallel. But the friends were still mostly talking to me. The output of a conversation with the strategist was advice. The output of a conversation with the CIO was a recommendation. The output of a conversation with the executive communications consultant was a marked-up paragraph or a suggested rewrite.

All of that was useful. Most of it lived in the conversation window.

What I started to want, especially as the work got bigger, was for the friends to actually make things. Not say things — make things. A finished memo, not a critique of my draft. A scheduled plan, not a suggested approach. A document I could use as-is, not a transcript I had to mine for the parts that were useful.

The conversation was sharp. The output was not yet portable.

That’s Step 4. I’ll get to it next week.

There is a version of this story that’s about software features. Projects. Persona instructions. Role-level context. Those are the technical names for what I built that morning.

But when I think back on the change, what I think about is something more human — and something I want to be careful not to overstate.

I have had decision-ready analysis in my career before. We all have. It shows up when you find, or develop, a person who can give it to you: the strategist who returns a question in the shape of a real answer, the CFO who has already considered the counterfactual, the chief of staff who brings you the sequenced plan and not the discussion of whether to plan. When you find one of those people, you cherish them. You hold onto them.

What was new that morning wasn’t that I had access to decision-ready analysis. It was that I could create people who produced it. Six of them, in a morning, faster than I could have developed a single human in months. Better? I don’t know. But faster and cheaper, for sure.

That changes the math. Not on any single question, but on the whole day. When decision-ready analysis is rare, you save it for the most important questions. When you can spin it up at will, more of your work gets that treatment. Decision-making velocity goes up. Productivity goes up. Not because anyone typed faster, but because the input to every decision arrived closer to ready than it used to.

That’s the shift, at Step 3. Not access to specialists — I’d had that before. The ability to make them.

A question for anyone reading this.

Who are the specialists you’ve cherished in your career — the ones whose analysis arrived already useful? What would it mean to be able to build one of them this afternoon, for a job you’re doing badly or partially by yourself right now?

That’s the friend you haven’t built yet.

Dennis

The conversations were smart but impersonal.

I’d describe a problem in detail, get a genuinely useful answer, and then have to describe the same problem again in the next conversation. And the one after that. The model had no memory of me. Every session started from zero.

For a while I worked around it. I’d paste in background at the top of each session. A paragraph about the product. A sentence about what stage I was in. It helped, but it was manual, and it meant every conversation started with me doing the same administrative work before I could get to the actual work.

This is the friction I described at the end of last week’s post. The model was smart. It just didn’t know me.

The fix I stumbled onto wasn’t something I planned. It was something I got tired of not having.

The afternoon I stopped re-explaining myself

This was still in my last operating role — well before the product, well before any of the work I do now. I’d been deep into something I was wrestling with, making real progress, when I had to step away. When I came back and opened a new session, I spent the first fifteen minutes re-establishing what we’d already established. The context. The constraints. The decisions we’d already made.

Those fifteen minutes made me angry in a productive way.

I didn’t want to carry the context. I wanted the model to carry it. I needed a way to install the background once — who I was, what I was working on, what I’d already decided — so that every conversation started from a shared foundation.

Here’s what I didn’t know at that point: the tool had already built this feature. There was a field, in settings, where I could put a profile that traveled into every conversation. A master prompt. It had been there the whole time. I just hadn’t noticed it, because I was still treating the tool like a search engine that happened to talk back.

I spent an afternoon filling in that field. Not a paragraph — a document. My name, my background, the work I was doing and why, the voice I write in, the decisions already locked, the things I was still working out. A profile, effectively. Not unlike a client intake form. A way of saying: here is who you’re talking to, so we don’t have to spend the first ten minutes on introductions.

The difference was immediate. The first conversation after I filled in that field felt like picking up a thread with someone who had been working on the same problem all week. The model referenced things from the profile without my prompting it to. It gave me specific recommendations instead of general frameworks because it knew my specific situation. When I pushed back on something, it pushed back with knowledge of what I’d already decided and why, instead of just accommodating me.

The AI could remember me now. Not in a magical way — in a structural one. The remembering lived in a settings field, but the effect was something I hadn’t felt from a tool before: I didn’t have to be a stranger every time I came back.

That’s when the word that came to mind changed. The model stopped being a tool. It started being something closer to a friend. A friend knows who you are without being reminded. A friend picks up where you left off. A friend has context.

That’s the artifact at this step. Not a document or a report — a profile. A file that lives outside the conversation and travels into every conversation. The first thing I made that changed how I worked rather than just what I produced.

What an artifact actually is

I should say something about the word.

An artifact, in this context, is anything the model produces that you can hold, share, use, or build on top of. A document. A spreadsheet. A plan. A script. A profile. An outline. A piece of code. Anything that persists beyond the conversation.

The jump from Step 1 to Step 2 — from conversation to artifact — sounds small. It isn’t.

When you’re having a conversation, you’re in the moment. When you have an artifact, you’re outside the moment. You can look at it. Revise it. Hand it to someone else. Bring it back to a different conversation. Use it as the starting point for something that didn’t exist before.

This is where the thought partner framing from Post 2 starts to become real. A thought partner who only lives in the session is limited. A thought partner who produces work that outlasts the session is categorically different.

The profile was my first artifact. Once I understood the pattern, I started making others: a voice guide that described how I write; a decision log of things I’d already figured out; a structure document for whatever I was working on. Each one was a way of externalizing knowledge I had been holding in my head, so that the model could carry some of it with me.

A thing that the better writers about AI keep saying is that the model is a mirror, not an oracle. I think that’s right but incomplete. When you start making artifacts, the model is a mirror that remembers. The conversation becomes a collaboration across time — between the Dennis of this session and the artifacts produced by the Dennis of last Tuesday’s session.

That’s a different kind of tool. And it changes what you reach for.

What it unlocked

Two things, immediately.

First: my working sessions got shorter and sharper. The re-establishment at the top of every conversation shrank to nothing. The work itself took up more of the time.

Second: the outputs got specific. Generic answers happen when the model is compensating for context it doesn’t have. When it has the context, it doesn’t need to hedge. It knows what kind of recommendation I want, what constraints apply, what I’ve already tried. The answers felt like they were for me, not for a category of person I loosely fit.

Neither of these is dramatic. They compound over time. After a few weeks of working this way, I realized I was going faster not because any single session was faster, but because the cumulative overhead of re-establishing context every time had quietly disappeared.

Where it broke down

Here is the new friction that came with Step 2, the thing that eventually pushed me toward Step 3.

A friend who knows who I am is good. A friend who knows my whole life is better. The profile I’d built was static — a single document about me, in a single field, applied to every conversation. It didn’t matter what I was working on. It didn’t matter who I was talking to inside the tool. The same paragraph showed up. The same intake form. One friend, who knew me in general, but didn’t know what I was up to today.

What I started to want was a different kind of friend for different kinds of work. The friend who knew the product I was building. The friend who knew my writing voice. The friend who knew the technical problem I was wrestling with on Tuesdays. Each one with its own background, its own focus, its own memory of what we’d been doing together.

A single profile couldn’t be all of those at once. What I needed was the ability to have many friends — each one specialized, each one tailored, each one ready to pick up exactly where we’d last left off.

That’s Step 3. I’ll get to it next week.

There’s a version of this story that’s about technology. But when I think back on that afternoon — the angry fifteen minutes, the profile I built in a single sitting, the session the next morning that felt so different — what I think about is something simpler.

I’d been treating a remarkable tool like a stranger I had to introduce myself to every morning.

That’s not a technology insight. That’s a human one. And I think it’s available to anyone who gets tired enough of re-explaining themselves to tools that should, by now, remember who they are.

A question for anyone reading this.

What do you re-explain about yourself every time you open a new window? Your role. Your voice. The thing you’re trying to do. The decisions you’ve already made and don’t want to defend again.

That’s the profile you haven’t filled in yet. And it’s the difference between a tool that doesn’t know you and a friend who does.

Dennis

Just Started is a weekly Monday essay by Dennis Hoffman about what it actually looks like to build with AI tools after a long career in something else. The Retirement Strategy (theretirementstrategy.ai) is the ongoing experiment this publication documents.

When I realized I could keep talking.

Step 1 of 8: conversation, not search.

Published: Monday, May 11, 2026, 7:00 AM MT
Cover: juststarted_post03_step1_cover.png (staircase graphic, Step 1 lit teal, Steps 2–8 muted)

The first thing I used it for was the thing everyone uses it for.

I had a question. I typed the question. I read the answer. I closed the tab.

That’s the search habit. I’d been running it since 1996. Google changed the form slightly — you typed keywords instead of sentences, you sorted through ten links instead of reading one answer — but the underlying move was the same. Question in, answer out, close the window, move on.

What I didn’t notice for the first several months of using Claude is that the window didn’t have to close.

The thing that’s different about a conversation

In the spring of 2025, I was working through a specific operational problem in my last operating role. Not strategic — operational. We were trying to figure out how to automate a process that was eating significant team time, and I was using Claude to think through it the way I might have used a consultant call. I typed the problem. I got a framework. I closed the tab.

Then, a few weeks later, working on a different version of the same problem, I typed a follow-up question. Not a new question — a follow-up. I wanted to go one level deeper on something the answer had implied.

The model remembered. Not in some mysterious way — I was still in the same conversation window — but it built on what had been said rather than starting from scratch. It treated my follow-up as a continuation, not a new query. And then it asked me a clarifying question I hadn’t thought to ask myself.

That was the moment.

Not the answer it gave. The question it asked.

It asked because I had kept talking. Because I hadn’t closed the tab. Because I was in a conversation now, not a search.

What conversation actually does

When you search, you retrieve.

When you converse, you think.

The difference sounds semantic. It isn’t. When I was searching — question in, answer out — the model was serving me what it knew. When I was conversing, the model was surfacing what I was already half-knowing but hadn’t yet said out loud.

The thing that conversation does, that no search ever does, is hold the thread.

A conversation remembers that I said, twelve messages ago, that the problem was really about two things, not one. It remembers that I agreed with a point and disagreed with another. It builds on the context I’ve given it rather than starting fresh with each exchange. And because it’s holding the thread, it can notice when something I say in message fourteen is actually a more interesting version of what I was trying to say in message two.

This is what I mean when I say the artifact of Step 1 wasn’t a document or a plan or a product. The artifact was my own thinking, made legible. Conversation produced something I couldn’t have produced alone — not because the model is smarter than me, but because it forced me to articulate what I thought until the articulation got clear enough to be useful.

Questioning makes you find what you think you know. Conversing makes you find what you actually think.

The experiment

That operational problem I was working on in spring 2025 took several conversations across several weeks. By the third conversation, the model had enough context that it was asking me questions I hadn’t asked myself. By the fifth, it was pushing back on an assumption I had baked in from the beginning — one I’d never examined because I’d just inherited it.

The assumption turned out to be wrong. Or at least incomplete. We ended up solving a slightly different problem than the one I had originally typed.

That would never have happened in a search. A search gives you what you already know to look for. A conversation gives you what you didn’t know you needed to find.

My team started doing the same thing, independently. I’d walk into a meeting and someone would have worked through a problem in Claude before the call — not to get an answer, but to figure out what question they were actually asking. The prep had changed. The questions in the room had gotten sharper.

We weren’t talking about it. We were just doing it. Conversations that didn’t close when we got an answer.

What it unlocked

The tool stopped being a reference and started being a relationship.

Not a relationship in the sentimental sense. A relationship in the working sense — the way you develop with a colleague over time. You stop explaining the backstory. You reference things that have already been established. You push back more confidently because you have a shared history of pushback-and-revision. The conversation has a texture that individual queries don’t.

Once I understood this, I stopped using single-shot questions for anything that mattered. Every real problem got a conversation. Every piece of work got a thread.

Where it broke down

Here’s the friction that comes with Step 1, the thing that pushed me toward the next step.

The conversations kept being too general.

I’d get a framework when I needed a recommendation. A set of options when I needed a decision. A thorough answer to the wrong version of my question because the model had no way to know which of a hundred contexts I was actually in.

The model was smart. It just didn’t know me.

What I needed wasn’t more conversation. It was a way to give the conversation context — to tell the model who I was, what I was working on, what I’d already decided. A profile. A set of instructions. A project.

That’s Step 2. I’ll get to it next week.

The thing the advice almost gets right

“Just start” is terrible advice. I’ve said that twice already in this publication, and I’ll keep saying it, because handing someone that phrase and walking away is the laziest kind of mentorship.

But here’s what the phrase gets right, almost by accident.

Once you change the first verb — once you stop searching and start conversing — you don’t have to plan the next step. The next step finds you. The conversation keeps going until the generality of it frustrates you, and the frustration is the thing that pulls you toward Step 2. You don’t sit down one morning and decide to add context. You get tired of repeating yourself, and context shows up as the obvious fix.

That’s the part that’s true. The ball does start to roll. You do just have to follow your nose.

What’s false is the idea that the ball starts rolling on its own. It doesn’t. Something has to tip it. For me, that something was the day I didn’t close the tab.

A question for anyone reading this.

When did you realize you could keep talking instead of closing the window? What happened in that conversation?

I’m asking because I think this is the most underdiscussed part of the whole AI-fluency thing. Everyone talks about prompting — the right words to use in the question. Nobody talks about the conversation that follows the question. But that’s where the actual work happens.

Tell me what you found on the other side of the tab you almost closed.

Dennis

Just Started is a weekly Monday essay by Dennis Hoffman about what it actually looks like to build with AI tools after a long career in something else. The Retirement Strategy (theretirementstrategy.ai) is the ongoing experiment this publication documents.

Last week I told you the story of the Saturday morning I opened Claude Code for the first time and typed a paragraph that turned into a product. I argued that “just started” is useless as an answer because it hides three ingredients: a year of small asks with the tool, an idea I’d been quietly carrying, and a specific kind of meta-question.

This week I want to go back to the first of those ingredients — the year of small asks — and tell you what was actually happening. Because when I look at it now, I wasn’t doing one thing. I was climbing a pyramid. I didn’t know it at the time. Nobody named the rungs for me. I want to name them for you, because once you can see the pyramid you can see where you’re standing on it, and that alone changes what you do next.

The seven rungs

Here is the pyramid. Read it from the bottom up.

1. Unaware. You don’t know what’s there. This is most of the world, still. Every time I say “Claude” to someone my age they say “Claude what?” That’s rung one. No judgment. You can’t use a tool you haven’t heard of.

2. Aware, not using. You know it exists. You’ve seen the headlines. Maybe a colleague mentioned it. You haven’t touched it. Something — time, fear, the sense that it’s not for you — keeps you off. Most educated professionals over fifty are on this rung right now.

3. Party trick. You’ve tried it. You asked it for a haiku about SPAM. You asked it to rewrite something in the voice of a pirate. You laughed. You showed someone. You closed the tab. You have the anecdote — “oh yeah I’ve played with it” — but not the habit.

4. Writing helper. This is the first rung where it starts to matter. You use it to fix an email. Tighten a paragraph. Summarize a long document. Draft a first pass of something you didn’t feel like drafting. It saves you time in small doses. It hasn’t changed how you think. It’s a better spellcheck.

5. Web browser replacement. You stop opening Google for a whole category of questions and you open the AI tool instead. “What’s the capital of Bhutan.” “How do I unclog this valve.” “What’s a 1031 exchange.” You get an answer in a paragraph instead of a page of blue links. You don’t cross-check every time. You notice, if you’re honest, that the results feel better than Google’s.

6. Thought partner. This is the rung where the tool becomes a relationship, not a utility. You bring a messy problem. You say here’s what I’m wrestling with, here’s what I think is going on, here’s what I’m not sure about, help me think through this. The exchange takes twenty minutes, not twenty seconds. You push back when the tool is off. You acknowledge when you’re lost. You end up with something better than you would have arrived at alone. You start noticing that some of your best thinking of the week happened in that window.

7. Builder. The single-threaded questions stop. Prompts stop starting with who, what, when, where. They start starting with build, create, draft, design, make. You stop asking the tool for information and you start commissioning it to produce something. A product. A plan. A document. An app. A website. A series. A whole thing.

The shift between 6 and 7 is grammatical

That’s the part I missed for most of the year.

The jump from rung six to rung seven isn’t about using the tool more. It’s not about a better prompt. It’s not about a subscription tier. It’s about the verb at the start of your sentence.

Can we discuss this. What should I think about here. Help me work through this. Those are thought-partner verbs. They assume you are the one doing the work, and the tool is there to think with you.

Build this. Draft this. Create this. Design this. Make this. Those are builder verbs. They assume the tool is doing the work, and you are there to direct it.

Most people don’t cross that line, and they don’t know the line exists. They stay on rung six forever, getting smarter conversations out of the tool, and they conclude that they’ve figured out what AI is for. They have not. They have figured out what rung six is for, which is a lot, but it’s not rung seven.

The people who are building the wave of products you’re starting to see coming out of solo operators — retired executives, former teachers, post-career professionals in their fifties and sixties — are people who crossed from six to seven. They didn’t do it by becoming more technical. They did it by changing the verb.

The Saturday was my crossing

I’ve been telling you, for two emails now, the story of a February Saturday when I opened Claude Code and typed a paragraph asking whether my retirement framework was a good use of the tool’s capabilities. That paragraph is a rung-six move. Can we discuss and refine a detailed approach. It’s a thought-partner question.

But the two hours after that paragraph — the back-and-forth that turned the discussion into an outline, the outline into a specification, the specification into code running on my laptop — was a rung-seven experience. Somewhere in the middle of that Saturday the verbs changed. Let’s build. Draft this. Create this piece. I wasn’t interviewing the tool anymore. I was directing it.

I had been sitting on rung six for most of 2025 without knowing it. I didn’t have a word for it. I just knew that my conversations with Claude had gotten longer and more substantive and that my team had started doing the same thing independently. When I look at that year now, I see six people together, on rung six, quietly getting fluent, none of us knowing that rung seven was an option.

The Saturday was the day I took a step up without planning to. The step wasn’t technical. It was grammatical.

The proof is in the second prompt

I didn’t tell you the whole story last week. There was a second prompt that same Saturday.

The first prompt, the one I walked you through last week, was the paragraph asking whether building a retirement advisor was a good use of Claude’s capabilities. Claude spent the next stretch of that morning interviewing me through the framework, teasing out where I was strong and where I was thin on my own retirement readiness. It was the best intake conversation I’d ever had about my own life. I was still in thought-partner mode — rung six — and I was getting a lot out of it.

A few hours later, still the same Saturday, I typed this second prompt:

I see a market opportunity and would like to work through it with you. Observation: an internet search for “Retirement Strategy” yields 99% “Wealth Management” information. So I built a retirement strategy framework with an ultimate objective of happiness (attached). This came from backsolving from unhappy retirees I know. After the first six months of “every day is Saturday” and “I’m on permanent vacation”, they seemed to get pretty depressed. I think that’s because the people, place, purpose part of their world revolved around work and largely disappeared when they retired. Moreover, they neglected their health and subjugated their passion — all for the pursuit of “enough money to retire”. I would like to find a modern, AI-first way to build a little side business helping people avoid that by putting the “same” amount of time, energy and money into the other five elements of my framework that they have in the wealth part. Thoughts about the idea? Ways to build a modern business around it?

Read those two prompts next to each other. Same Saturday. Same person. Same tool.

The first one starts with can we discuss. The second one starts with I see a market opportunity. The first one asks about implementation for me. The second one asks about a business for other people. The first one is thought partner. The second one is builder.

I didn’t notice at the time. I was just typing what came next. But when I go back and look at the two paragraphs side by side, the grammatical shift between them is the whole subject of this essay, happening to me, in real time, over the course of an afternoon I thought was just me playing with an AI on a Saturday.

And notice what the second prompt already contained. A market observation I’d done the search to justify. A customer segment I’d backsolved from. A product thesis — put the same time, energy, and money into the five neglected dimensions. A business model ask — ways to build a modern business around it. All of that came out of me in one paragraph because once you start typing builder verbs, the tool has permission to pull the rest of the thesis out with it.

I called it “a little side business.” The thing I was about to start has become something other than little. The prompt was smaller than the thing it started.

That is what rung seven looks like, up close. Not drama. A paragraph. A change of verbs. Claude doing the work you authorize it to do, instead of the work you keep asking it to think about with you.

Where you are right now

Look at the list again and locate yourself honestly. Nobody’s watching.

If you’re on rung one or two, the move is to get on rung three. Go try it. Ask for a haiku about SPAM. I mean that literally. The party trick is not a waste of time. It’s the first touch. Without it you can’t get to four.

If you’re on rung three or four, the move is to use the tool for something real this week. Not a haiku. An actual task from your actual life. Draft the email you were dreading. Summarize the document you didn’t read. The point is not the output. The point is the groove.

If you’re on rung five, the move is to bring a real problem. Something you’re actually wrestling with. Something you don’t have a clean answer to. Spend twenty minutes on it. Let the conversation go where it goes. See if you end up somewhere you wouldn’t have gotten to alone.

If you’re on rung six — and I suspect many of you are, because the subscribers to this publication are the kind of people who would be — the move is different. The move is to notice, one of these days, that you want to ask the tool for a thing instead of a conversation. When that moment comes, don’t flinch. Change the verb. Say build this or draft this or create this and see what happens.

You don’t have to know what you’re doing. The whole point of rung seven is that you find out by doing.

One thing to try before next Monday

Open your AI tool. Ask yourself: what’s a thing I’ve been meaning to make but keep not making?

Doesn’t matter what it is. A toast for a wedding. A business plan for an idea you half-have. A speech. A spreadsheet. A letter you owe someone. A website for something you’ve been thinking about. A chapter. A curriculum. A policy doc. A game. An app.

Then type this, or something like it:

I want to build [the thing]. Before we start, can you walk me through how you’d approach this, what the options are, and what questions I should answer before we begin? Then let’s build a first draft together.

Notice the verbs. Build. Walk me through. Begin. Build a first draft together. That is a rung-seven sentence. Most of you haven’t typed one before.

Type it. See what happens. The point is not the artifact. The point is the step up.

I’ll see you next Monday.

Dennis

Just Started is a weekly essay by Dennis Hoffman about what it actually looks like to build with AI tools after a long career in something else. The Retirement Strategy (theretirementstrategy.ai) is the ongoing experiment this publication documents.

The phrase I’ve said too many times in the last six months is “I just started.”

Someone asks how I went, in a little over a year, from running a large team at Dell to using AI to build software and run a small business with no employees. How does an executive who spent 22 years coordinating people start writing apps? The honest answer is that I just started.

It’s true. It’s also useless.

If you’re the person asking, “I just started” isn’t a story. It’s a wall. You nod, you smile, you say something polite, and inside you think okay, but what does “just start” actually mean when I don’t know where to begin, when I’ve never written a line of code, when the AI thing sounds like it’s for twenty-five-year-olds in San Francisco, when my career was in a different field, when I’ve watched three younger colleagues try this and quietly fail, when I’m not even sure what I’d build, when I’m quietly worried I’m already too late?

You don’t say any of that. Instead you say something like “that’s amazing, I really should look into it,” and you change the subject, and you go home and you don’t look into it. Looking into it means sitting in front of a blank screen with no idea what to type. Nobody tells you what to do at that moment. The moment passes.

I want to tell you what it actually looked like. Because once I say it out loud, “just started” doesn’t sound the way I’ve been using it.

What I was really doing during the year before I “started”

The part of my story I’ve been glossing over is that I didn’t just start. I spent a year getting ready to start without realizing I was getting ready.

It was the spring of 2025. I was still at Dell, still running the telecom business, still working through the specific operational problem of modernizing the division I was responsible for — how to simplify, how to standardize, how to automate, how to apply AI to the key processes that determined whether the business unit worked. How to build better products for bigger customer problems. How to transform telecom with operational efficiency. That kind of work.

Those problems taught me to talk with Claude.

Not as an experiment. Not as a prompt-engineering exercise. As a working relationship. I learned, across that spring and through the year, how to actually have conversations with an AI model. I learned to acknowledge when I was clueless. I learned to ask for help. I learned to tell Claude when it was off track and push back when the response was thin. I learned what it could do well and where it fell over.

By the middle of the year Claude wasn’t just mine. It had become a thought partner for my exec team — a member of the team, in a strange way, not anyone’s assistant but a resource we talked to together. We ran problems through it before we ran them through each other. By the end of the year most of the senior people on my team were doing the same thing independently. None of us announced it. It just happened, as the tool got better at what we needed and we got better at the tool.

In parallel, underneath the work drumbeat, I was hearing a second one. The entrepreneurs and thought leaders I was reading were saying that agentic AI was going to enable something new — the one-person startup. A company without employees. A product without a team. An entrepreneur with an idea and a set of tools.

I was curious. I didn’t have an idea.

The thing I actually had

What I had, before I had a framework, was a sentence running in my head.

There has to be a better way.

I was watching my own retirement coming at me and I could feel that the standard playbook wasn’t going to be enough. At some point I googled “retirement strategy” expecting to find it. What I got instead was buried in articles and videos about when to take social security and how to build a retirement portfolio. Financial planning, dressed up as retirement planning. That was the whole internet on this subject.

My first idea, if you can call it an idea, was just the complaint. Retirement planning is broken. There has to be a better way. I need that better way.

That’s it. That was the whole thing, for a while.

I want you to notice what that is and what it isn’t. It isn’t a product. It isn’t a business plan. It isn’t an insight. It’s a person noticing a gap between what exists and what he needs, and feeling irritated enough about it to keep noticing. Most people have a sentence like that running somewhere in their lives. Most people don’t think of it as an idea. They think of it as a frustration, or a complaint, or a nothing — the thing they grumble about in the car.

The frustration is the idea. You just haven’t given it that name yet.

Mine eventually turned into a framework. Across 2024 I kept sketching and re-sketching a way of thinking about my own retirement, and in the back half of the year I wrote it up as a blog post — a hand-drawn image of a house, with Happiness as the roof, People, Place, Passion, and Purpose as the four pillars, Health and Wealth as the foundation. Six dimensions. The argument was that most retirement planning is actually financial planning — just Health and Wealth — and the four pillars that determine whether retirement is a good life get almost no attention.

A few dozen people read it. A few replied with thoughtful notes. I thought about it, off and on, for another year. I didn’t turn it into anything. I had a framework looking for a form.

That is the part of the story that matters. The thing I eventually handed to Claude wasn’t a blank idea I came up with that morning. It was an argument I had been carrying around for more than a year, that had itself come out of a frustration I had been carrying for longer than that, waiting for me to figure out what to do with it.

Most people who say they “just started” did this too. They had a sentence running in their head — there has to be a better way to do this thing — that they had never named as an idea. The moment of starting wasn’t the moment they had a flash of insight. The moment of starting was the moment they finally picked up the complaint they had been carrying and decided to treat it like a question worth answering.

If you’ve been telling yourself you don’t have an idea to start with, I want you to consider that you may be wrong. You may have a sentence running somewhere in your life right now — there has to be a better way to X, why is Y still this hard, why has no one fixed Z — that you’ve stopped recognizing as an idea because you’ve gotten used to carrying it.

The Saturday

On a Saturday in early February 2026, I opened Claude Code for the first time. I took a photograph of the framework image. And I typed this:

Attached is my philosophy/framework for Happiness (the goal) in retirement. I would like your help implementing this framework in my life. Specifically, I’m curious how we might create a retirement advisor that enables me to chat with a single point of contact but parses the work to a parallel team of experts on each part of the framework. I would like to be able to use this as often as I want over the next 20+ years as I age and the world changes, so there needs to be long-term memory. Before setting anything up, can we discuss and refine a detailed approach for implementation? What are the options, pros and cons? Is this a good use of your capabilities?

That’s what I wrote.

I want you to notice what it is and what it isn’t.

It isn’t a polished incantation. It isn’t a demonstration of prompt-engineering expertise. It isn’t a request to build a product. It’s a paragraph asking for a discussion. I wasn’t asking Claude to build anything yet. I was asking Claude whether this was worth building, what the options were, what the pros and cons were, and whether this was even a good use of the tool in front of me.

That last sentence is the one I want you to read again. Is this a good use of your capabilities?

That’s the question I had been too intimidated to ask for most of the previous year. I had been using Claude for discrete tasks — help me think through this problem, help me draft this note, help me understand this thing. I had not, until that Saturday, asked the bigger question. What could this actually be? Is what I’m imagining a reasonable use of what’s in front of me?

The fact that I eventually asked it — after a year of smaller asks — is the actual beginning of The Retirement Strategy. Not the code, not the product, not the launch. The Saturday I asked the big question of the tool I had been quietly getting to know for a year.

The answer Claude gave me, over the next two hours that Saturday, turned into the outline that turned into the product I’ve spent the last three months building.

What “just started” actually means

Read in reverse, my “just started” story has three ingredients that all matter:

One: a year of small asks. The spring-of-2025 Dell work taught me how to work with AI. Without that year I would not have had the fluency or the nerve to ask the big question on that Saturday. The small asks compound, quietly, into something that feels like confidence when the moment comes.

Two: an idea I had been carrying for a long time — first as a frustration, then as a framework. There has to be a better way came before the blueprint. The blueprint came before the Saturday. I hadn’t generated either one in the moment. They were ready because I had been carrying the itch, and then the shape of an answer to it, for longer than I had noticed.

Three: a specific kind of prompt. Not a request to do the thing. A request to discuss whether the thing was worth doing, what the options were, and whether the tool could actually do it. The meta-move. The question most people skip because asking it requires admitting you don’t know what’s possible.

Most people who are stuck are stuck because they have one or two of these ingredients, not all three. They have the carried idea but no fluency with the tool. They have the fluency but no carried idea. They have both but they skip the meta-question and go straight to “build this for me,” and the tool builds something half-right, and they conclude it doesn’t work.

The sequence matters. Fluency first. Then recognition of what you’ve already been carrying — even if what you’ve been carrying is still just a complaint. Then the meta-question.

“Just started” means all three of those, compressed into a sentence, after the fact. That’s why it’s useless when people say it to you. The version that’s actually useful is the version I just told you.

What this publication is

If you’ve been hearing your own version of the voice that told me I was too old, not technical enough, too late to matter — this publication is for me to stop giving you the useless answer and start giving you a useful one.

Every Monday I’ll narrate what it actually looked like to build The Retirement Strategy as a sixty-two-year-old retiree with a year of professional AI experience but no coding background, using tools that didn’t exist three years ago, as a solo operator with no team and no investors. Then I’ll narrate whatever comes next.

The posts will be shorter than you expect. The point is not to make this publication into a thing. The point is to give you, in enough specificity that you can locate yourself inside it, the view of what it looks like when a person like me decided to pick up the idea he had been carrying and ask whether the tool in front of him could help him finally do something with it.

You’ll hear about the pricing decisions I can’t land. The partnership conversations that might change the whole venture. The moments I thought I should quit. The things I keep getting wrong. The tools I use, with specificity — not “AI” in general, but this tool, for this task, in this way, here’s what broke.

I’m writing this for the person who reads everything and doesn’t start. The person who’s bought three books on AI and hasn’t opened them. The person who has the tab open and doesn’t know what to type. The person whose spouse or kids keep asking what they’re going to do next and who doesn’t have a good answer.

You don’t have to build a product. You don’t have to start a business. You don’t have to retire to do this. What you have to do is recognize the idea you’ve been carrying, put your hands on a tool you’ve been quietly getting to know, and ask the meta-question first.

One thing to try before Monday

Two small things, actually.

First: write down, in one or two sentences, the idea you’ve been carrying. Not pitched. Not polished. Just what it is. Most people reading this have one. If you don’t have one, write down the problem you keep noticing — the thing that bothers you when you see it, the gap you keep thinking about. That’s the same thing in a different form.

Second: open the AI tool of your choice. Paste the sentence. Then ask this, or something like it:

I’ve been thinking about this for a while. Before we do anything, can we discuss whether this is a good use of your capabilities? What are the options, pros and cons? Is this actually worth doing?

That’s the meta-question. It takes you sixty seconds. It puts your hands on the keyboard. It doesn’t commit you to anything.

Watch what comes back. See where it goes. Tell me about it, if you want — I read the replies.

Next week I’ll tell you about the pyramid I didn’t know I was climbing — the seven rungs of AI use, and the one grammatical shift that separates the people who use these tools from the people who build with them.

I’ll see you Monday.

Dennis

Just Started is a weekly essay by Dennis Hoffman about what it actually looks like to build with AI tools after a long career in something else. The Retirement Strategy (theretirementstrategy.ai) is the ongoing experiment this publication documents.