| name | manual-brainstorming |
| description | Use only when the user explicitly asks to brainstorm, clarify, design, or spec a change before implementation, or explicitly names this skill. Do not use automatically for ordinary build, fix, implementation, or code-edit requests. |
Manual Brainstorming
Overview
Use this skill as an explicit design conversation before implementation. Keep it lightweight enough for Codex: clarify intent, compare approaches, validate the design, then produce an implementation plan only after approval.
Trigger Discipline
- Use this skill only because the user explicitly requested brainstorming, design/spec work, or this skill by name.
- Do not infer this skill from a normal request to build, fix, refactor, or edit code.
- Do not write code, scaffold files, commit, push, or open a PR during this skill unless the user explicitly changes the task.
Workflow
-
Explore project context.
- Read existing project instructions such as
AGENTS.md, relevant docs, package/config files, and recent diffs when a repository exists.
- Identify existing ownership boundaries, validation commands, documentation rules, and whether the request touches sensitive areas such as persistence, auth, routing, background jobs, or UI state.
-
Scope the request.
- If the request includes multiple independent subsystems, recommend splitting it into separate design/plan cycles.
- Keep the current cycle focused on one deliverable that can be implemented and verified independently.
-
Ask clarifying questions one at a time.
- Focus on purpose, constraints, non-goals, success criteria, affected users, and acceptable tradeoffs.
- Prefer concise multiple-choice questions when they reduce ambiguity.
- Stop asking when there is enough information to propose credible approaches.
-
Propose 2-3 approaches.
- Lead with the recommended approach and explain why.
- Include tradeoffs: implementation risk, testability, maintainability, user impact, and scope.
- Call out what should be deliberately excluded by YAGNI.
-
Present the design in reviewable sections.
- Cover architecture, affected modules, data/control flow, edge cases, error handling, test strategy, and documentation impact.
- Scale detail to risk: short for small changes, deeper for stateful or cross-module changes.
- Ask the user to confirm or revise each meaningful section before proceeding.
-
Use the user's .agent-work/PLAN.md plan flow instead of Superpowers docs.
- Do not create
docs/superpowers/specs/ or docs/superpowers/plans/.
- Do not commit the design or plan unless the user explicitly asks for a commit.
- Keep the design in chat until the user approves it.
- When an implementation plan is requested and approved, write the complete Markdown plan to
.agent-work/PLAN.md according to the user's global instructions.
- If
.agent-work/PLAN.md already exists, read it before continuing and avoid overwriting user work without explaining the replacement.
-
Self-review before handoff.
- Check for placeholders, contradictions, vague requirements, missing error cases, untestable acceptance criteria, and scope creep.
- Verify the plan maps each requirement to concrete tasks and validation steps.
- Ensure the plan does not introduce unnecessary shared state, broad refactors, or duplicated logic.
-
User review gate.
- Ask the user to review the final design or
.agent-work/PLAN.md before implementation.
- If the user requests changes, revise and repeat the self-review.
- Do not implement until the user approves the design/plan or explicitly asks to proceed without a plan.
-
Transition to implementation.
- After approval, follow repository instructions and the user's normal implementation workflow.
- If context was compacted or plan memory may be incomplete, reread
.agent-work/PLAN.md before continuing.