Launchling was the first product I coded myself with AI tools, and I think it's more useful to me now as a record of what building with AI was like in mid-2025 than as a product. I'm not relaunching it. I'd rather explain honestly what it was, how far I got, what went wrong, and why the gap it filled has since closed, because each of those taught me something I now use every week.
What it was for
Launchling (originally TinyLaunch) helped people with an early idea, particularly non-technical solo founders and people the startup world doesn't usually make room for, turn it into a clear, personalised plan: what to build, for whom, what to test first, and which tools suited their skills, budget and confidence. It also had a Risk Mapper to surface the assumptions most likely to sink an idea. The point wasn't to generate documents faster; it was to take away some of the advantage enjoyed by people who already know how product development works.
How the stack grew
- Prove it: a Tally form and Zapier, with AI generating the plan. No code.
- Make it feel real: Framer for a branded experience, until I hit the limits of what it could do dynamically.
- Own it: a React, Vite and TypeScript front end, with Tailwind tried first and CSS Modules at launch (July 2025), Vercel serverless functions in TypeScript and JavaScript, GPT-4o in two structured steps (the plan, then the tool stack), and Airtable for storage. Built mostly in Cursor.
- Make it trustworthy: a separate Python evaluation system (pytest, Pydantic schemas, Jinja2 prompt templates, versioned prompts and rubrics, real and synthetic test cases), writing its results to Airtable and local files.
Every step was a reasonable decision for a solo builder in 2025. The problem is that each one added software I then had to understand and maintain, and by the fourth I wasn't running a React app that called GPT-4o; I was coordinating two languages, several places that each held their own idea of what a plan looked like, versioned prompts, two kinds of test data, logging and several external services.
What worked
I think it's worth saying clearly, because the ending can make it look as though nothing did. The core flow worked: a founder could answer the onboarding questions, GPT-4o generated a plan and tool recommendations, and the result was saved. I tested it with deliberately absurd ideas, French pirate founders included, and designed it to handle them gracefully. Early versions of the evaluation framework ran and scored plans, in two early runs in July 2025: 11 real plans on 13 July (published at the time: personalisation 4.36, tool fit 4.09, help alignment 3.73), then a larger set of scores, committed on 19 July, that included those 11 (overall 3.98, tool appropriateness 4.08, personalisation 4.08, help alignment 3.94). Neither came from the later full system, which never ran end to end again. I never checked the judge against my own scoring, which is now the first thing I'd do. The thinking behind the evaluation, that AI output needs judging on whether it's actually useful to this person and not just on whether the API returned something valid, is the part I'm proudest of, and it still holds.
Where it broke
As the evaluation system grew, it stopped saving its results properly to Airtable, even though the logs showed it running. Over five days at the end of July and start of August 2025 I tried to get a complete, repeatable evaluation run working again, and couldn't. Between 31 July and 3 August it made 258 run folders and saved a result in none of them (the code made a folder whenever it was loaded, so that isn't a count of runs), and its last summary, on 4 August, records no plans evaluated. I had 33 real plans and 96 synthetic ones to evaluate, but every fix exposed the next problem: scripts that couldn't find each other, configuration one part expected and another didn't supply, duplicate modules and circular imports that stopped the tests loading, mocks of Airtable and OpenAI that no longer matched the real thing, and metadata that existed in the logs but vanished before it reached the results.
Looking back, almost none of these were problems with the AI's output; they were broken agreements between parts of my own system. Each AI-suggested fix was locally sensible without anyone, me or the tool, holding a picture of the whole, so the next missing piece surfaced straight away. The pattern I'd recognise now is that I'd built a sophisticated platform for evaluation before I'd proved the simplest version of it: one plan, one rubric, one result, saved somewhere I could read it.
The uncomfortable part is what I was measuring by the end. I'd started out wanting to know whether Launchling's advice was good, and by August I was counting passing tests. In a product team I'd call that substituting activity for evidence of the outcome, and I did it to myself.
I couldn't see what it cost
Separately, and just as stubbornly, Launchling's AI token usage reached Airtable, but never reliably. I spent days on it with both ChatGPT and Cursor, and it never logged reliably. It mattered more than it sounds: I was paying for every generation and every evaluation run myself, and without reliable usage records I couldn't say what a plan cost to produce, what a batch evaluation would cost, or whether a prompt change had made things more expensive. I was building an AI product without dependable sight of either of the two things that decide whether one works: the quality of what it produces and the cost of producing it.
I still don't know the cause. The records that did arrive miscalculated and often restarted from zero, and when a response carried no usage data the logger wrote a stand-in record (2,000 tokens, one cent), so the table mixed real usage with placeholders. It made a second problem worse: Cursor occasionally ran OpenAI calls I hadn't approved, until I stopped it running scripts at all, and I couldn't see what those cost either.
Why it has no place now
The real reason I'm retiring it is that the gap it filled has closed. The capability Launchling packaged, turning a vague idea into a structured, personalised plan, is now something general-purpose models do well on their own. The people I built it for can get most of what it offered by asking Claude or ChatGPT directly. I think the defensible version of the idea would be something quite different, helping a founder establish whether an assessment is true rather than generating one, and that would be a new product rather than a rebuild. I've written about what makes an AI product durable when the models keep improving, in The Moat, Not the Feature; Launchling is my own worked example of the opposite.
What carried forward
Quite a lot. The instinct that AI output needs judging on usefulness, not just validity, became my guide to evaluating AI output, rewritten a year later with everything Launchling taught me. The failure to see cost became a rule for Fold: it must never be loss-making, and it has a dedicated cost reviewer, which first ran on 1 November. The broken agreements between parts of the system are why Fold's reviewers each start from a fresh context, and why each of them has to say what it didn't check.
If you built something with AI in 2025 that you'd now retire, I'd be glad to compare notes. Get in touch.