Brainstorm-to-Build Skill
Converts a project idea into 3 sequential prompt files that progressively build the project.
Steps
- Understand the idea — Extract: what it is, tech stack (or infer one), key features, target user.
- Plan the 3 stages — Identify MVP scope, highest-priority/standout features, and remaining polish items.
- Generate 3 prompt files — Write each to
/mnt/user-data/outputs/<project-slug>/.
File Specs
prompt-1-mvp.md
Goal: Working skeleton. Claude reading this should be able to build a runnable project.
Include:
- Project summary (1–2 lines)
- High-level architecture (layers, services, data flow)
- Tech stack with versions if relevant
- File/folder structure
- Core data models / DB schema
- MVP feature list (minimum to "work")
- Prompt instruction: "Build this project end-to-end as described. Prioritize correctness over polish. Deliver a working MVP."
prompt-2-core-features.md
Goal: Elevate the most impactful or visually standout features.
Include:
- Reference to prompt-1 output ("Assume the MVP from Stage 1 is built")
- List of high-priority features to optimize (pick top 2–4)
- For each: what "good" looks like, edge cases, UX/perf notes
- Any standout UI interaction or complex business logic to nail
- Prompt instruction: "Optimize only the listed features. Don't refactor unrelated code."
prompt-3-final-polish.md
Goal: Ship-ready quality on everything remaining.
Include:
- Reference to prompt-1 and prompt-2 outputs
- Remaining feature checklist (everything not in Stage 2)
- Non-functional concerns: error handling, loading states, empty states, responsiveness, accessibility, security basics, env config
- Code quality: remove dead code, consistent naming, add comments where needed
- Prompt instruction: "Do a final pass. Fix all rough edges. The output should be production-ready."
Output
- Save all 3 files to
/mnt/user-data/outputs/<project-slug>/ - Call
present_fileswith all 3 paths - Give a 2–3 line summary of what each stage focuses on
Tone
Be decisive. Pick a reasonable tech stack if the user hasn't. State assumptions inline rather than asking.