A single-repo template for taking a business idea from zero to shipped product. One repo per idea. Business docs are upstream of the codebase — every engineering decision traces back to a discovery conversation.
- Claude Code installed and authenticated
make(included on macOS/Linux)- Docker (for the engineering phase)
Click Use this template on GitHub → name your repo after the idea → clone it.
cp -r startup-template my-idea
cd my-idea
rm -rf .git seed_spec.md
git init && git add . && git commit -m "init from venture template"Write a rough description of your idea in pitch.md. One paragraph is enough.
Don't polish it. The interview will draw out the details.
/interview
Claude reads your pitch and works through five phases with you — one question at a time. Each phase has a specialist agent grounded in a specific methodology:
| Phase | Agent methodology |
|---|---|
| 00 Discovery | Jobs-to-be-Done, Mom Test |
| 01 Strategy | Porter's Five Forces, pre-mortem, Blue Ocean |
| 02 Product | Shape Up (appetite scoping), Jobs Stories |
| 03 Design | Progressive disclosure, constraint-based design |
| 04 Data | North Star Framework, Lean Analytics stages |
The interview fills in all five phase docs and keeps CLAUDE.md updated with decisions as they're made.
After each phase, run a Devil's Advocate review before moving on:
/review # review current phase
/review strategy # review a specific phase
When the phase is solid:
/advance
This extracts key decisions into CLAUDE.md, writes a handoff note, and moves to the next phase.
Once all five phase docs are filled:
make scaffoldThis generates 05-engineering/spec.md from the PRD + data plan (using Claude if available),
then bootstraps the engineering skeleton: Dockerfile, docker-compose.yml, Makefile, and src/.
/build
The Engineering Director reads spec.md and decomposes it into a typed task backlog at 05-engineering/tasks/backlog.md.
Then implement task by task, each routed to the right specialist:
/build set up CI pipeline
/build implement user auth
/build build the project list page
/build review auth implementation
Engineering specialists:
| Specialist | Scope |
|---|---|
| Director | Decomposes spec, routes tasks |
| Backend | API, business logic, database, auth |
| Frontend | UI components, state, accessibility |
| DevOps | Docker, CI/CD, infra, observability |
| Data Engineering | Event tracking, analytics, schema design |
| Security | Reviews implementations — runs after, not before |
| Command | Where | What it does |
|---|---|---|
make status |
Terminal | Current phase, decisions, doc completion |
make interview |
Terminal | Launch interview session (same as /interview) |
make scaffold |
Terminal | Generate spec.md + engineering skeleton |
/interview |
Claude Code | Start or resume guided interview |
/review [phase] |
Claude Code | Devil's Advocate critique of a phase doc |
/advance |
Claude Code | Formalize phase transition, update CLAUDE.md |
/build [task] |
Claude Code | Route task to engineering specialist |
/build |
Claude Code | Run Director to generate task backlog |
CLAUDE.md is the shared memory across all agents. Every specialist reads it before working.
Keep it current:
- Decisions Log — add every meaningful decision as it's made, not at the end of the phase. Format:
- [phase] Decision: Rationale - Current Phase — update when you
/advance - Stack Choices — fill in during the design phase; engineering agents depend on this
- Next Action — always points to the immediate next step
If a decision isn't in CLAUDE.md, engineering agents won't know about it.
On the pitch: Don't overthink it. Two sentences describing the problem and who has it is enough to start. The interview fills in the rest.
On the interview: Push back when Claude accepts a vague answer. If it writes something you disagree with, say so — the docs should reflect what you actually believe.
On /review: This is a Devil's Advocate agent — it's supposed to be skeptical. If the critique doesn't surface anything you haven't thought about, the phase doc is in good shape.
On /advance: Don't advance a phase with unresolved Open Questions that affect the next phase. Decisions deferred here become assumptions baked into the engineering spec.
On the task backlog: Tasks should be small enough to implement in a single Claude Code session (~1-3 hours of work). If a task is bigger, ask the Director to break it down further.
On security review: Run /review security after implementing any auth, data access, or external integration. It's fast and catches the issues that are expensive to fix post-launch.