CodeTrain

Said I'd come back with real numbers, so here they are.

First, the promise I can't keep: I said I'd report what each channel converts. I can't. I never instrumented attribution before launch, so I have placements and I have signups, and no honest line connecting the two. That's the most expensive mistake of the two and a half weeks and it cost me nothing to avoid. If you're pre-launch reading this, that's the free lesson: an hour of attribution work before launch day might be worth more than anything you'll do on launch day. I got so busy just trying to get content online I skipped some steps.

What I do have is the activation funnel, which turned out to be the more interesting half anyway.

- 31 signups since launch

- 18 of them opened a lesson

- 8 of those at least wrote a line of code in it

- 4 actually finished a lesson

- 0 paying

So 13 people signed up and never opened anything, and another 10 opened a lesson and stopped before typing. The whole business is downstream of that second group. I spent two and a half weeks optimizing the top of the funnel while the leak was in the first five minutes of the product, which I could have measured on day one for free.

Costs: $0 on ads. Total inference cost of every lesson anyone has run since launch, on my own key, is 31 cents. The free tier is capped at 10 lessons a month and nobody has come close to hitting it, which kills the "raise the limits" theories.

Channel placements, for whatever they're worth without conversion data: Show HN got 4 points. Product Hunt finished 83rd. A directory called Uneed that I hadn't heard of a month ago put us 3rd for the day. The pattern I did not expect is that the twenty-minute directory listings did more for me than the two launches every founder optimizes for. I'm not a social media native and it shows, so take that as one founder's data point rather than advice.

The other honest thing: on the Team tier question I asked in this thread on day 3, I still don't have an answer, because I have no upgrades of either kind to learn from. What I have instead is a hypothesis I'm now going to test directly. Through September I'm taking a few engineering teams as onboarding pilots, hands-on setup from me, in exchange for real feedback that isn't filtered through politeness. If self-serve was never going to produce Team accounts at this volume anyway, better to find out from ten real teams than from another month of waiting.

Next 30 days: fix the first lesson, instrument activation properly, and stop treating launch days as the job.

Question for anyone further along: how did you instrument activation early without building an analytics project instead of a product? I keep debating myself about a week of tooling I don't think I can afford right now.

About two weeks ago I posted CodeTrain here: the AI tutor that refuses to write your code for you. Today it's on Product Hunt!

The short version, if you missed it: AI writes most of our code now, and teams are quietly leaking understanding of the code. CodeTrain is basically the opposite of autocomplete. It breaks a task into small steps, makes you write every line, then reviews your work Socratically and never hands you the answer. Free tier, no card. The core tutor is an open-source skill you can run today.

I'm in the comments all day if you want to poke holes in it: https://www.producthunt.com/products/codetrain?launch=codetrain

(Founder here — happy to answer anything, including the skeptical stuff.)

A few days ago I launched CodeTrain publicly (codetrain.ai). The pitch is contrarian to the current agent coding culture: it's an AI coding tutor that will not write code for you. It plans a lesson from your own codebase in tiny steps, you type every line into an editor with a run/test loop, and it grades each step. Beg for the answer and the step gets smaller, not handed to you.

Why I think there's a business here: AI assistants made producing code cheap, and most engineering managers I've talked to have noticed the same side effect. Juniors ship more and understand less. Teams are paying for that gap in review load and bus factor, and not many solutions are available for the countermeasure. Codecademy teaches someone else's codebase. Copilot writes yours for you. The empty quadrant is "learn your own codebase by writing it." A brand new industry with the rising of AI agentic coding.

Where the money is: free browser tier (10 sessions/mo, capped because inference costs real money and I don't want a rug-pull free tier), Pro at $24/mo with repo mode, Team at $32/seat with a ramp dashboard. My bet is the Team tier carries it; individual devs are the funnel and the proof.

Stack is FastAPI on Fly, static site from Cloudflare Pages, auth via Clerk, billing from Stripe, LLM is mainly Anthropic models with BYOK support. (we do have some test models available for usage through OpenRouter as well)

The launch plan is deliberately boring: owned channels this week, Show HN and Product Hunt next week once my accounts have enough history not to look like a bot(not particularly a huge social media fan unfortunately), then a month of Reddit and newsletters. Zero paid anything. I'll report what each channel actually converts.

Question for people who've sold to eng teams: did you land Team plans from self-serve individual users upgrading, or from founder-led outbound? I've budgeted for outbound but I'd love to be wrong.