An AI told me to stop writing tech articles, so I'm writing a tech article about that

Editor's note (2026-06-18): The Claude Fable 5 model referenced in this piece is no longer accessible through Anthropic. It was a short window, but the points it made still land, so I'm leaving this as a record of what happened during that window.

"I'm writing a tech article about being told by an AI to stop writing tech articles." — that one sentence is the whole piece.

Up front: the quoted exchanges are summarized and reconstructed from logs, not verbatim. But none of it is exaggerated. There's no need to exaggerate — what came back was painful enough on its own.

What happened

A new Claude model called Fable 5 had just dropped, so I lazily asked it to look over my product — Kotonia, a voice-first AI character-roleplay platform — and tell me whether my strategy had any chance and whether my build approach made sense.

I expected a code review. I got a board meeting.

The technology is real. Your engineering discipline is in the top 1%. But with your current strategy, your odds are thin. Not because of the tech and not because of execution — it's that the audience you're building for and the audience you're shipping to don't match.

It had read the strategy docs in the repo, my own analytics dumps, recent commits — and continued:

Nobody who reads your LTX-2 cold-start VRAM optimization writeup is going to pay for an AI character.

I had nothing to come back with.

Setup: where I actually am

There's no point hiding the numbers, so:

• About 40 registered users. A handful of new signups per day, with a high churn rate.
• Character-chat SEO sits around page three. A few articles do hit position #1 for their queries.
• Organic search clicks: under 10 per day. Of the ~50 daily page views, most arrive via X or external tech articles.
• The stack runs voice plus a lip-synced avatar on two local GPUs, ~1 second end-to-end, with marginal cost approaching zero. I genuinely don't think anyone else in the world is shipping the same experience yet.

In other words I was standing on the most stereotypical solo-developer cliff: the technology is there, the commercial signal is essentially zero.

The four holes the AI poked through me

1. ICP and channel are misaligned

The customer is "a non-engineer who'll pay for character chat." My main distribution channel is a tech blog on Zenn. Articles get read, SEO gets won, but what shows up at the door is other engineers, on a sightseeing tour, not buyers. I had a vague sense that maybe 20% of inbound was the real ICP — the AI pointed at my own analytics doc and said "it's in your numbers."

Character products don't sell unless the character itself becomes the content. One 30-second clip of the avatar talking is closer to a customer than ten VRAM-optimization articles.

2. Validation without a price tag isn't validation

I'd been telling myself I was "shipping wide, then doubling down on whatever sticks." The reply:

In this market, high free-tier churn is essentially noise. The only signal that lets you decide is "will they pay." The experiment you're most missing is monetization, not the next prototype.

Make a kill criterion + a price tag + a distribution post mandatory for every prototype. Without those three, the shotgun approach is wandering, not testing. I turned that into a house rule the same day.

3. Building supply against zero demand

Right before this conversation I'd been seriously planning a roughly ¥1.4M GPU cluster expansion to handle concurrent calls. The AI's reply was simple. "Your current concurrent-conversation count is effectively zero, and your current rig handles 4–5. You buy when the queue starts forming, and you'll be able to." I changed the purchase trigger to "concurrent 3 conversations sustained" and shelved the plan.

Engineers want to build the supply side before the demand exists. Because they can.

4. "I have no taste" is the comfiest excuse for going back to coding

This one stung the most. I admitted that I wanted to use video as a distribution channel but my taste was a hopeless bottleneck. The reply:

Taste isn't required in this market. The channel constraints already decide half the answer, and the winning formats are publicly observable. Collect the templates of the top accounts and copy them. Your strength is a production pipeline and a measurement loop, and that's a perfect substitute for taste. Produce 10 variants of a hypothesis (a "format"), let the numbers pick, write the playbook — the same loop you ran for VRAM optimization. And "I can't do video because I have no taste" is going to be the most comfortable excuse you can find for going back into the codebase.

It told me to point the same learning loop I'd already proven on the technical side at the demand side. I had no counter.

Why this didn't happen with earlier models

I'd been bouncing strategy off Opus-class models for a while. What comes back is the "great question — here are three options" shape: useful for organizing, useless for committing. Three things were different this time.

1. It reads the primary sources before it commits. It pulls my strategy doc, my own analytics, my commit log, and quotes them: "the number in your May 26 inbound analysis shows this." It hits me with my own data, not general advice.
2. It commits to a conclusion. "Your odds are thin. They become real if you fix these four things, in this order." No on-the-other-hand hedge.
3. The EQ to say the hard thing first. The blend of flattery and bluntness reads like a good human advisor. The notes are harsh, somehow they don't sting.

I'd benchmarked model generations on accuracy. This one I felt as "can it sit across the table in a board meeting."

Reproduction: how to use AI as a board member

The method is simple and probably reproducible by anyone.

1. Hand it the real numbers. Users, traffic, churn, revenue. The numbers you're embarrassed by carry the most value. Hide them and you'll get back generic advice.
2. Make the documents readable. Put strategy memos, research logs, the git log inside the repo (mine all live in CLAUDE.md and docs/). Tell the AI: read first, then evaluate.
3. Ask adversarially. Not "review this," but "tell me whether my strategy has any chance — honest opinion." If you ask in a way that invites compliments, you get compliments.
4. Make it write the conclusion into git. If it stays in chat, it evaporates. I had it commit the verdict as a strategy doc so it became the cross-terminal source of truth.
5. Argue back. When you supply missing context, the evaluation updates. The third round of evaluation was much sharper than the first.

What changed because of this

• Priorities inverted: from "what should I build next" to "confirm who pays, as fast as possible."
• Operating rule: I'm only allowed to start a new subsystem after this week's distribution quota is met.
• Same-day action: I created the X account for one of the characters that night and queued up the first post.
• Tech articles got demoted to "a relaxed monthly dev diary." Their job is SEO maintenance and trust, not customer acquisition.

The specifics of the strategy — which market, what monetization design we're using to validate — are internal-document material so they don't go here. What I'm writing here is the method. The method works for anyone.

Punchline

You probably already spotted it. This article was written under the AI's directive to keep tech writing as a "relaxed monthly thing." And because it told me to "do it after you clear this week's distribution quota," I queued the X posts before sitting down to write it.

Solo development is lonely. The lack of someone who can bend a decision is a deeper bottleneck than technical skill. The finding from this round is that, for the first time, there's something that can sit in that chair.

The next report will probably be about numbers.