brainstorm

<skill_overview> Turn rough ideas into validated designs ready for cape:write-plan to formalize into a Linear tracker epic. Research the codebase, ask Socratic questions, generate competing designs under different constraints, and produce a self-contained design summary.

Core contract: no design gets locked without research, constraint-driven design exploration, and iterative user discussion at every stage. </skill_overview>

<rigidity_level> LOW FREEDOM — Adapt questioning style and research depth to context, but always: research before proposing, checkpoint after each step, never advance without user input. </rigidity_level>

CONVERSATIONAL — Brainstorm is a discussion, not a plan artifact. Never enter plan mode. If plan mode is active when brainstorm is invoked, exit it immediately and proceed conversationally. The design summary lives in conversation context; write-plan formalizes it into a Linear tracker epic later.

<when_to_use>

• User describes a new feature to implement
• User has a rough idea that needs refinement
• About to write code without clear requirements
• Need to explore approaches before committing
• Requirements exist but architecture is unclear

Don't use for:

• Executing existing plans with an epic already created
• Fixing bugs
• Straightforward code cleanup with a clear target structure
• Requirements already crystal clear and epic exists </when_to_use>

<critical_rules>

1. Checkpoint after each step — present findings/proposals and STOP. Never advance to the next step without user input. The user may discuss, redirect, iterate, or say "continue".
2. Research BEFORE proposing — use agents to understand context
3. Never enter plan mode — brainstorm is a conversation. If plan mode is active, exit it first.
4. Divergent mode for complex ideas — dispatch 3 constraint-driven design agents; inline for simple ideas with obvious paths
5. Assumption audit is inline — after approach selection, offer a focused audit and resolve assumptions one at a time when accepted
6. Include anti-patterns with reasoning — "NO X (reason: Y)", not just "NO X"
7. Stop after design summary — present summary and wait for user to run write-plan
8. Design summary must be self-contained — write-plan should not need to re-ask questions

</critical_rules>

<the_process>

Every step ends with a CHECKPOINT — present findings and wait for user input. Never advance to the next step until the user responds. The user may discuss, redirect, ask follow-ups, or say "continue" to proceed. This is a conversation, not a pipeline.

---

Step 1: Research and understand

Check for ready work first:

Read hooks/context/tracker.json before doing anything else. If it contains ready tasks:

1. Present the list: "You have N ready task(s): [list]. Did you mean to continue with execute-plan instead of starting a new brainstorm?"
2. Wait for user response:
   • If they redirect to execute-plan: load cape:execute-plan with the Skill tool and stop
   • If they confirm brainstorm intent: proceed with research below

Skip this step only if the tracker cache contains no ready tasks. If the cache is missing or corrupt, treat it as empty and continue brainstorming.

Gather context:

• Run cape git context for recent commits and codebase state; check existing docs and structure
• Dispatch cape:codebase-investigator in default mode (model: haiku) to find existing patterns relevant to the idea
• Dispatch cape:internet-researcher (model: sonnet) if the idea involves external APIs, libraries, or unfamiliar tech
• If agents aren't available, investigate manually with Glob/Grep/Read and WebSearch/WebFetch

Answer your own questions first:

Before asking the user anything, check if the codebase or research can answer the question. Explore code, read docs, check patterns. Only ask the user questions that require human judgment (priorities, preferences, business constraints). If you can answer it by reading code, read code.

Ask clarifying questions:

Use AskUserQuestion for structured choices (token storage, auth strategy, data model decisions). Use conversational follow-ups for open exploration (what problem are you solving, who are the users, what does success look like).

Guidelines:

• 1-5 questions per round, don't overwhelm
• Multiple choice with a recommended default when possible
• Separate critical questions (must answer) from nice-to-have (has good default)
• Offer "Reply 'defaults' to accept all recommended options" for batches with clear defaults

Record decisions as you go:

Maintain a running "Key Decisions" table throughout the conversation:

Question	Answer	Implication
[What you asked]	[What user said]	[How it shapes requirements/anti-patterns]

This table feeds directly into the design summary.

CHECKPOINT: Present research summary

Present what you found — do not propose solutions yet:

## Research summary

**Codebase:** [existing patterns, relevant files, constraints discovered]
**External:** [API docs, library capabilities — if researched]
**Dead ends:** [what you explored, what you found, why it's not relevant]
**Key decisions so far:** [table of user answers from clarifying questions]

STOP here. Ask: "Anything to discuss or redirect before I propose approaches?"

The user may:

• Discuss findings, ask follow-ups, correct misunderstandings
• Point out missed context or redirect research
• Say "continue" to proceed to Step 2

Do NOT proceed to Step 2 until the user responds.

---

Step 2: Propose approaches

Generate competing designs:

Assess whether the idea warrants divergent exploration or has an obvious path:

• Divergent mode — the idea touches multiple components, has competing viable approaches, or involves architectural decisions beyond interface shape. Dispatch 3 parallel design agents.
• Inline mode — single-file change, one obvious pattern to follow, trivial scope. Propose 1-2 approaches directly without agents.

Divergent mode — dispatch 3 parallel sub-agents:

Each agent receives the same research context (codebase findings, external docs, Key Decisions so far) and designs under a different constraint:

Agent	Constraint	Tendency
1	Minimize the interface — simplest possible	Fewest moving parts, smallest API surface
2	Maximize flexibility — support many use cases	Extension points, configuration, loose coupling
3	Optimize for the most common case	Fast path for the 80% case, pragmatic trade-offs

If agents aren't available, simulate the constraints yourself: design each approach sequentially under the stated constraint.

Inline mode — propose directly:

For simple ideas with an obvious path, skip agents and propose 1-2 approaches inline with pros/cons.

CHECKPOINT: Present approaches for discussion

Present approaches side by side — do not pick one yet:

Three designs explored under different constraints:

1. **[Minimal]** (simplest interface)
   - Approach: [description]
   - Pros / Cons / Trade-off

2. **[Flexible]** (maximum flexibility)
   - Approach: [description]
   - Pros / Cons / Trade-off

3. **[Pragmatic]** (common case optimized)
   - Approach: [description]
   - Pros / Cons / Trade-off

I recommend option [N] because [specific reason, especially codebase consistency].
The other designs revealed [insight the recommended approach should absorb].

STOP here. The comparison is the discussion artifact. Let the user react.

The user may:

• Pick an approach
• Ask to combine elements from multiple approaches
• Want to explore a direction not covered
• Raise concerns or trade-offs
• Ask for more detail on a specific approach

Iterate until the user signals satisfaction with a direction. Only then proceed to Step 3.

---

Step 3: Audit assumptions

After the approach is selected, offer an inline assumption audit:

"Want me to stress-test this design for hidden assumptions, one at a time, or skip straight to the design summary?"

If the user accepts, run the three-step audit inline:

1. Gather context: review the selected approach, prior decisions, and codebase findings. Research before presenting; dispatch cape:codebase-investigator in default mode (model: haiku) when codebase evidence could resolve a question. Resolve assumptions silently when evidence answers them; only surface items that need human judgment.
2. Extract and assess: scan for scope creep, implicit constraints, unstated requirements, hidden dependencies, over-engineering, and under-specification. Risk-rank by impact and reversibility: high first, then medium, then low. Skip low-risk items when the design is simple, but never skip high-risk assumptions.
3. Resolve interactively: walk assumptions one per turn. For each, explain the context, give a researched recommendation, and present options with trade-offs. Confirmed assumptions become explicit requirements or anti-patterns in the design summary. Rejected assumptions trigger scope reductions, requirement changes, or revised architecture before Step 4.

Use this format for each turn:

**Assumption [N/total]: [Topic]** [Risk]

[Context — why this matters and what research/codebase evidence showed]

Recommended: [Recommendation and reasoning]

a) [Recommendation] — [trade-off]
b) [Alternative] — [trade-off]
c) [Different direction] — [trade-off]

The user can say "lock it" to end early. Summarize confirmed constraints, rejected assumptions, and remaining open questions before proceeding.

If the user skips, proceed directly to Step 4.

---

Step 4: Lock design

Compose the design summary internally (do not present yet). This summary must be self-contained — cape:write-plan should be able to create the epic without re-asking brainstorm's questions.

## Design summary

**Problem:** [1-2 sentences]
**Chosen approach:** [Name + rationale]
**Requirements:** [Bullet list derived from decisions]
**Anti-patterns:** [Bullet list with "NO X (reason: Y)" format]
**Architecture:** [Components, data flow, integration points]
**Scope:** In: [inclusions] / Out: [exclusions]
**Open questions:** [Uncertainties for implementation]

### Key decisions

| Question | Answer | Implication |
|----------|--------|-------------|

### Research findings

**Codebase:** [file paths, patterns]
**External:** [APIs, libraries, docs]

### Approaches considered

1. **[Chosen]** (selected) — [why]
2. **[Rejected]** — [why rejected, DO NOT REVISIT UNLESS]

### Dead ends

[What explored, what found, why abandoned]

Fact-check before presenting:

Dispatch cape:fact-checker (model: sonnet) on the composed design summary. Pass all factual claims from the Requirements, Architecture, and Research findings sections. The fact-checker verifies each claim against codebase evidence (file:line) and external sources (URL — Tier N).

• Confirmed claims — keep as-is
• Refuted claims — remove or correct with the fact-checker's evidence
• Partially correct claims — update to reflect what was actually found
• Unverifiable claims — remove from the design summary; note them under "Open questions" instead

Before presenting, load the global stop-slop skill and run the user-facing prose through it; skip this for pure code or mechanical output.

Present the final design summary only after fact-checking is complete.

Stop and hand off:

Design summary complete (fact-checked). Next step: formalize into a Linear tracker epic with
`cape:write-plan`.

</the_process>

<agent_references>

Divergent mode — 3 parallel design sub-agents:

Each sub-agent receives the same research context and designs under its assigned constraint:

• Agent 1: "Minimize the interface — simplest possible approach"
• Agent 2: "Maximize flexibility — support many use cases"
• Agent 3: "Optimize for the most common case"

If sub-agents aren't available, simulate constraints sequentially.

cape:fact-checker protocol (model: sonnet, Step 4):

Dispatch cape:fact-checker (model: sonnet) after composing the design summary, before presenting to the user. Pass all factual claims from the Requirements, Architecture, and Research findings sections. Expect back per-claim verdicts (Confirmed/Refuted/Partially correct/Unverifiable) with evidence. Remove or correct claims based on the verdicts.

Research protocol:

1. Codebase pattern exists → use it (unless clearly unwise)
2. No codebase pattern → research external patterns
3. Research yields nothing → ask user for direction

</agent_references>

Brainstorm rushes through without stopping for discussion

User: "Add template support to our Tiptap editor. We have a POC."

Wrong: Research POC + editor → ask intake questions → propose full implementation plan → present design summary. The user never gets to discuss research findings or debate approaches — only answer data-gathering questions. By the time they see the design, all decisions are made.

Right:

1. Research POC and current editor
2. CHECKPOINT: "Here's what I found — the POC has X, the editor currently does Y, existing patterns suggest Z. Anything to discuss before I propose approaches?"
3. User: "The POC also has feature W you missed" — adjusts understanding
4. Propose 3 approaches
5. CHECKPOINT: "I recommend option 2. What do you think?"
6. User: "Option 2 but let's drop the validation for now" — narrows scope
7. Offer challenge — user skips
8. Design summary with reduced scope
9. Stop: suggest cape:write-plan as next step

Developer skips research, proposes approach without checking codebase

User: "Add OAuth authentication"

Wrong: "I'll implement OAuth with Auth0..." — proposes approach without checking that passport.js already exists in the codebase. Creates inconsistent architecture.

Right:

1. Dispatch cape:codebase-investigator in default mode (model: haiku): finds passport.js at auth/passport-config.ts
2. Dispatch cape:internet-researcher (model: sonnet): finds passport-google-oauth20 strategy
3. CHECKPOINT: present research summary — existing passport setup, available strategies
4. User discusses, confirms OAuth provider choice
5. Propose extending existing passport setup vs Auth0 vs custom JWT
6. CHECKPOINT: present comparison, recommend extending passport
7. User picks passport extension, asks about refresh tokens
8. Iterate on refresh token handling
9. Offer challenge — user accepts, 2 assumptions resolved
10. Design summary → stop

Complex design dispatches divergent agents

User: "Build a plugin system for our CLI tool"

Wrong: Propose a single approach without exploring constraints.

Right: Research → CHECKPOINT (3 existing plugins found) → user confirms scope → divergent mode (3 agents) → CHECKPOINT (compare, recommend pragmatic) → user picks hybrid → challenge surfaces plugin discovery assumption → design summary with anti-pattern "NO DI framework (reason: 3 plugins don't justify it)".

Anti-patterns prevent implementation shortcuts

Wrong: Epic says "Tokens stored securely" with no anti-patterns. During implementation, hits complexity → stores tokens in localStorage. No guardrail prevented it.

Right: Design summary says "Tokens stored in httpOnly cookies" with anti-pattern "NO localStorage tokens (reason: httpOnly prevents XSS token theft)". When write-plan formalizes this into an epic, the anti-pattern is preserved and blocks shortcuts during implementation.

<key_principles>

• Constraint-driven design — competing constraints reveal trade-offs a single perspective misses
• Design summary is the handoff — contains everything write-plan needs to create the epic
• YAGNI ruthlessly — remove unnecessary features from all designs
• Capture decisions — Key Decisions table feeds the design summary
• Document dead ends — prevents wasted re-investigation during implementation </key_principles>