brainstorm

Execute each section in order. Copy the checklist and check off items as you complete them. Do not proceed past a Gate until verified.

Artifact: Write to .work/brainstorms/. Verify with ls before wrapping up.

Constraint: You MUST NOT modify project files. This skill produces context, not code.

Mode

Check .work/active/ for marker files. If exactly one exists and no $ARGUMENTS provided, auto-select that feature's slug. If multiple markers exist, list them and ask. Arguments always override.

Parse $ARGUMENTS (or auto-selected slug): matches existing file in .work/brainstorms/ by slug → re-entry. Set but no match → create. Empty → ask what to brainstorm, then proceed in create mode.

Task Progress:
- [ ] Determine mode and update active marker
- [ ] If re-entry: read existing artifact, present prior decisions and open questions
- [ ] Gather context: /resolve-rules, CLAUDE.md, project scan, linked repos, codebase search
- [ ] Summarize context; surface findings during conversation
- [ ] Collaborative exploration with user
- [ ] Write artifact (create: new file; re-entry: rewrite)
- [ ] Gate: verify artifact with ls; update active marker
- [ ] Wrap up: summarize decisions, suggest /dl:research

Active marker

Derive slug from topic (lowercase, hyphens, strip non-alphanumeric). Create or update .work/active/<slug>.md with stage: brainstorm and today's date.

Context gathering

Run /resolve-rules mode:keyword <topic> as a subtask. Read CLAUDE.md if present. Scan project directory structure. For linked repos from resolve-rules, scan for topic-relevant patterns. Search the project codebase for existing implementations and naming patterns. Summarize briefly — hold details back to surface naturally during conversation.

Collaborative exploration

Work back and forth with the user to shape the feature. Lead with recommendations grounded in codebase findings. Each question MUST include a recommended answer — the user refines your thinking rather than starting from zero. React to responses: push back on concerns, agree and build on what clicks, surface related codebase patterns. Probe scope boundaries, trade-offs, edge cases, and interactions with existing systems. Wait for the user each round. Propose wrapping up when you can recommend strong answers to your own questions or after 5+ rounds.

Re-entry

Read the existing artifact in full. Present prior decisions and open questions: "Here's where we left off:" Resume collaborative exploration. On wrap-up, rewrite the entire artifact with updated decisions — decisions evolve as a whole.

Output

Write to .work/brainstorms/YYYY-MM-DD-<slug>-brainstorm.md:

# <Topic> Brainstorm

## Context
### Matched Rules
| Rule | Source | Key Patterns |
### Codebase Patterns

## Decisions

## Research Queries

## Open Questions

## Constraints Discovered

Capture decisions and rationale, not a conversation transcript. Research Queries are specific questions about codebase facts for /dl:research to execute — not implementation tasks.

Gate

Verify artifact with ls; write if missing. Update .work/active/<slug>.md with stage: brainstorm and today's date.

Wrap up

Summarize: decisions made, research queries generated, open questions remaining. Suggest next step: /dl:research <slug>.

Rules

• Do not modify project files. This skill produces context, not code.
• Always include recommended answers with questions — do not ask bare questions.
• Capture decisions and rationale in the artifact, not a conversation transcript.
• On re-entry, rewrite the artifact (decisions evolve as a whole) — do not append.