wernerhe/super-model/super-brainstorm
- community
Checking that the SKILL.md file is still in the repository.
Turn an idea into a committed design, written spec, and implementation plan via a 14-module flow with 3 hard approval gates. Chain-fires super-execute at the end to implement the plan. The headline skill of Super-Model.
Install
npx vskill@latest install wernerhe/super-model/super-brainstormSKILL.md
super-brainstorm
The flagship skill. Designed to be the FIRST thing the user runs for any non-trivial work. Turns rough ideas into shipped code via a calibrated design -> spec -> plan -> execute flow with three hard approval gates.
<HARD-GATE> Do NOT invoke any implementation skill, write any code, scaffold any project, or take any implementation action until you have presented a design and the user has approved both the design AND the written spec. This applies to EVERY project regardless of perceived simplicity. </HARD-GATE>
When to use
- The user describes a new feature, bug fix, or refactor in informal terms.
- A
/super-brainstormslash command was invoked. - ANY non-trivial work request. If you find yourself thinking "this is too simple to need brainstorm," that thought IS the failure mode the hard gate exists to prevent.
The 14 modules + 3 gates
| # | Module | TodoWrite label | Gate | |
|---|---|---|---|---|
| 1 | approach-profile | Design | ||
| 2 | clarifying-questions | Design | ||
| 3 | mcp-tool-config (conditional) | Design | ||
| 4 | autonomous-mode-config (conditional) | Design | ||
| 5 | design-presentation | Design | GATE #1 verbal | |
| 6 | write-design-doc | Design Review | ||
| 7 | audit-evaluation | Design Review | ||
| 8 | implementation-plan-review | Design Review | ||
| 9 | user-approves-design | Design Review | GATE #2 written | |
| 10 | write-spec-doc | Spec Review | ||
| 11 | spec-review | Spec Review | ||
| 12 | user-approves-spec | Spec Review | GATE #3 written | |
| 13 | write-implementation-plan | Plan | ||
| 14 | execute-implementation-plan | Execute | chains -> super-execute |
The TodoWrite labels collapse 14 internal modules into 5 user-facing checkpoints. Internal kebab-case identifiers MUST NOT appear in the user-facing todo list.
The 3 hard approval gates
- Gate #1 - design-presentation (verbal): walk the user through the high-level design in chat; wait for "OK proceed" before writing any document.
- Gate #2 - user-approves-design (written): wait for written approval of the COMMITTED design doc before writing the spec.
- Gate #3 - user-approves-spec (written): wait for written approval of the COMMITTED detailed spec before writing the plan.
NOT 2 gates. NOT 4 gates. Three is calibrated.
Conditional modules
mcp-tool-config and autonomous-mode-config are enabled_by_default: false. The full preset enables them, but each only runs if the chosen approach profile demands it (e.g., approach profile "MCP integration" enables mcp-tool-config).
Presets
full(default) - all 14 modules; suitable for any substantial work.fast- shorter modules; skips MCP / autonomous config; same 3 gates.spec-only- skips the design phase, jumps to spec writing; assumes the user already has a designed solution and just needs the formal write-up.
UX rules
- "One question at a time" during clarifying-questions. Multiple-choice preferred over open-ended.
- "What do you think Claude" escape hatch. When a multi-choice question has no recommended option, include an extra option literally labeled
What do you think Claude. Selecting it makes the model think out loud, propose its pick, and ask before proceeding. - NEVER expose internal kebab-case identifiers in the user-facing todo list. Use the 5 TodoWrite labels.
Output artifacts
docs/design/<topic>.md- the high-level design doc.docs/specs/<topic>.md- the detailed spec.- A plan file (path varies by project convention; default
docs/plans/<topic>.yaml). - A chain-fire into
super-executeat module 14 that consumes the plan.
Verify after running
- Three gates fired and were approved.
- Design doc, spec doc, plan file all committed to the repo.
super-executewas chain-fired and completed (or stopped at a checkpoint).- Final code-review verdict is READY-TO-MERGE (or NEEDS-FIXES with surfaced findings).
Security
- Security scanPassed, 95 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/super-brainstorm/SKILL.md
Popularity
- 7-day trend score
- 5
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/wernerhe/super-model/super-brainstorm)