How we use AI to ship features faster — without cutting corners. Read this once before you start your first feature.
7 stages from raw feature idea to live in production. Each stage has a clear input, a clear output, and a done checklist.
You paste the prompt into Claude, GPT-4o, or any LLM. It generates a draft. You review it. You own the result.
Each stage produces a .md file. No lost Slack decisions. The whole feature is traceable.
Each stage has its own .md file: stage-01-feature-clarification.md, stage-02-check-existing-flow.md, and so on up to stage 07. Find the one matching your current stage.
Inside the file, find the section labelled 🤖 AI Prompt. Copy everything inside the code block — that's the prompt you'll paste into the AI.
Every prompt has clearly marked spots like [PASTE STAGE 01 RESULT HERE] or [FEATURE REQUEST]. Replace each one with the real information before sending to the AI.
Open Claude (claude.ai) or ChatGPT or any LLM. Start a new chat. Paste the filled-in prompt. Send it. Wait for the full response.
Read every line. The AI drafts, you decide. Fix anything that doesn't match reality. Add anything it missed. This is not optional — AI output must be reviewed by a human who knows the context.
Create a folder for the feature: features/2025-01-15-pet-reminders/. Save the output as the appropriate file, e.g. 01-feature-definition.md.
Each stage file ends with a ✅ Done When checklist. Every box must be checked before proceeding to the next stage. If a box can't be checked, resolve the blocker first.
| Stage | Name | Primary Owner | AI Involvement | Key Output |
|---|---|---|---|---|
| 01 | Feature Clarification | Product | Drafting aid | Feature definition + user stories |
| 02 | Check Existing Flow | Tech Lead | Analysis aid | Flow diagram + dependency map |
| 03 | Impact Analysis | Tech Lead | AI Risk identification | Impact list + risk & mitigation plan |
| 04 | Feature Design | Senior Eng | AI Architecture drafting | Backend + DB + frontend design doc |
| 05 | Development | Engineer | AI Code generation + tests | Working code + PR summary |
| 06 | Testing & Regression | QA / Eng | AI Test case generation | Test report + security audit |
| 07 | Release & Deployment | Tech Lead | AI Release notes + runbook | Release notes + rollback plan |
Every prompt asks for prior stage results. Include them. The AI has no memory — without context it guesses.
Going from idea to code without Impact Analysis leads to rework. Every stage catches something the next one shouldn't have to fix.
AI drafts, you decide. Read it line by line. Fix inaccuracies. Add what it missed. You own the output.
AI doesn't know your codebase, your customers, or your constraints. It gives a head start — not a finished answer.
These files are the project memory. Name them consistently and keep them in the feature folder.
The done checklist is a gate, not a suggestion. Unresolved risks in stage 03 will become bugs in stage 05.
Any instruction-following model works — Claude, ChatGPT, Gemini. For design stages (03, 04) Claude tends to produce more structured output. For code generation (05) Claude or GPT-4o both work well. Use what you have access to. The prompts are model-agnostic.
Fix it. The AI is a fast first draft, not a source of truth. If you notice a gap, add it to the file before moving on. The done checklist will catch most missing items.
No. Use judgment. A one-line bug fix doesn't need a risk register. A new user-facing feature or anything touching shared infrastructure should go through all 7. If in doubt, start at stage 03 (Impact Analysis) as a minimum.
Create a folder in the repo under features/ named with the date and feature slug, e.g. features/2025-01-15-pet-reminders/. Save each stage output there with the matching filename from the master guide.
Document the blocker in the output file under "Open Questions" and tag the owner in your project tracker. Do not proceed to the next stage until it's resolved — the whole point of early stages is to catch blockers before they become expensive.