Milestone map
Milestone map
3 milestones
Deploy a working prototype to a URL
2–4 weeks
Pick the specific problem your AI product solves. Build a prototype and deploy it to a URL anyone can open. It needn't be polished, but it must work without a local setup. Name the core AI part: a model call, retrieval pipeline, classifier or agent loop. Make sure it carries the product. Remove it, and the main value should disappear.
Proof required
Submit three things. A live URL anyone can open now. Two paragraphs on the problem and the AI part driving it. Name the retrieval design, prompt chain or model choice, and why. 'I used GPT-4' isn't enough. One screenshot of the product in use.
What gets checked
- The URL is live, needs no login, and the AI part is interactive. A repo link doesn't count.
- The problem names who has it and what they do today. Say why AI does it better.
- The AI part carries the product. You can say what changes when the model answers differently.
Common mistakes
- Weeks of architecture before deploying anything. Localhost isn't shipped; deploy early.
- The AI is incidental. It sits in one optional step and the product works without it.
- 'Works on my machine' isn't deployed. Vercel, Render, Hugging Face Spaces and Replit host free.
Resources
Foundationstart here
Depthgo deeper
Masteryfor the dedicated
What a verifier looks for
- Open the URL now. If it doesn't load, this milestone isn't met.
- Ask what happens without the AI part. If 'not much', the AI is decorative.
- Ask them to explain the model or retrieval choice in one sentence. 'I used the API' fails.
You'll sign in first, then come straight back here.
Get five real users and their feedback
1–3 weeks
Reach at least five people who really have the problem. Don't count people doing you a favour. Have each try a real task in the product. Record what they said. Feedback from people with the problem is evidence; friends being supportive is noise. It shows whether the AI part is useful or just impressive.
Proof required
Submit a feedback log with five or more entries. Each shows how you found them and why they have the problem, not just 'friend'. Include the task you set and what they said, word for word or close. Include one thing you changed or will change. A spreadsheet, doc or notes all work.
What gets checked
- Each entry has real recruitment context, like 'found in a Reddit community for Y'.
- Each shows a task they tried, not 'I showed them the product'.
- At least one entry shows confusion or frustration, with your concrete response.
Common mistakes
- Counting demo viewers as users. They must try a task themselves.
- Recruiting only people likely to be kind. The point is to find what breaks.
- A list of quotes with no 'here's what I changed'.
Resources
Foundationstart here
Depthgo deeper
What a verifier looks for
- Ask how each user was found. Probe whether they really had the problem.
- Look for friction. If everyone found it clear and delightful, recruiting was too friendly.
- Ask what they changed because of one piece of feedback.
You'll sign in first, then come straight back here.
Prove 10 users return over 4 weeks
4–6 weeks (runs concurrently with M2 user-recruitment period)
A product used once and dropped is a demo. Show at least 10 distinct users who came back more than once. Cover at least four calendar weeks. The proof is analytics, API logs or a usage record with timestamps. A self-reported count doesn't pass. The AI part must be why people return, not a feature they skip.
Proof required
Submit usage data covering four weeks or more: an analytics screenshot, log excerpt or dashboard. It must show 10 distinct users with at least one return each. Add a short note on which AI feature brings people back, and how you know. Add one user quote about using it in real work.
What gets checked
- Ten distinct users, not 10 visits from three people. The data must tell users apart.
- The dates span at least 28 days. A 48-hour burst doesn't count.
- You can explain why users return for the AI feature specifically.
Common mistakes
- Treating page views as users. Use sessions, logins or API keys to tell people apart.
- People return, but for a non-AI feature. That doesn't prove the AI product works.
- Four weeks pass, but use clusters in one week. Use should spread across the weeks.
Resources
Foundationstart here
Depthgo deeper
What a verifier looks for
- Ask for the raw data source, not a summary they wrote.
- Ask how the data tells one user from another. Page views aren't enough.
- Ask which AI feature shows up in return visits. If none, retention may be for another reason.
You'll sign in first, then come straight back here.
Part of