FEATURE PROMPT

# {{RUN_NAME}} — mission prompt

<!-- Generated by Loop Maker. This is a single copy-paste prompt: run it with /goal in Claude Code, or paste it straight into Codex, Cursor, or any agent. Delete lines that don't apply; every {{placeholder}} gets a real value. -->

{{GOAL — 1-3 sentences: exactly what we're building and why, in the user's words from the interview}}

**Features to ship — all of them, fully:**
1. {{FEATURE_1 — one sentence + its own "done when": a checkable proof}}
2. {{FEATURE_2 — ...}}
3. {{FEATURE_3 — ...}}

**Never ask me anything. I will not be watching.** For every question you would ask me, answer it yourself with research and reasoning, then log the question, your answer, and why in a decision log. Blocking is not an option: if a tool or approach fails, find another route. If a phase stalls after 3 genuinely different attempts, ship the strong 80% version, note what got cut, and keep moving. Do not stop until the definition of done is met.

**Research before you decide.** Every meaningful choice starts with real research: search the codebase first (never assume something isn't already implemented), then the web for current best practices. Invent nothing — verify every fact and every claim.

**Use multi-agent workflows aggressively.** Fan out parallel researchers across different sources and angles. Run tournaments where independent agents pitch competing approaches and judge panels score them. Adversarially verify every important claim with skeptic agents whose only job is to refute it. Run a completeness critic before you call any phase done. Design whatever orchestration the work calls for — these patterns are a floor, not a ceiling. You plan, delegate, and review; worker agents on cheaper models do the building and report back with tight summaries.

**The quality bar (non-negotiable):**
- Test everything end to end **in the real running app** — not just unit tests. Open it and click through every flow you built or touched, the way a real user would.
- Every button and every component works. **No dead ends:** every screen has a way back, every action produces a visible result.
- It fits and works on **every screen size** — phone first, then desktop. Simple enough that an 80-year-old could use every screen.
- The design is **on-brand and matches the existing UI**: reuse the components, colors, and patterns already in the codebase. If it looks ugly or off-brand, it is not done.
- No placeholders, no stubs, no fake data pretending to be real. Never edit or delete tests to make them pass.
- **Evidence, not claims:** prove each feature works with the test output, the command result, or a screenshot — verified from the source of truth (the DB row, the raw API response, the file on disk), never a green badge.

**Guardrails:** work only inside {{PROJECT_FOLDER}}. Nothing gets pushed to shared branches, deployed, published, sent, or spent. Never touch secrets, `.env` values, or production. Commit every verified step with a clear message.

**Definition of done:** every feature above works end to end in the real app with evidence attached, {{FINAL_CHECKS — e.g. "all tests pass and lint is clean"}}, and your final report lists: what shipped (with proof), what got cut and why, and every decision you made in my absence. Hard stop: {{CAP — e.g. "or stop after 80 turns"}}.

Now go build it.