AgentSkillsCN

brainstorm

根据Git差异生成合适的提交消息

SKILL.md
--- frontmatter
name: brainstorm
description: "Workflow-first design through collaborative dialogue. Use before creating features, building components, or modifying behavior."

Feature Tree Brainstorming

Turn ideas into fully formed plans through collaborative dialogue.

Four phases: Discovery → Product → Design → Specification


CRITICAL: How to Use This Skill

ONE QUESTION AT A TIME.

Do NOT dump all questions at once. That produces shallow, useless answers.

For each question:

  1. Present the question
  2. ULTRATHINK — deeply consider it, write your thinking
  3. Ask user: "Does this seem right?" or "Is this the right read?"
  4. Wait for user response
  5. Only then move to next question

DO NOT MOVE TO NEXT PHASE UNLESS USER EXPLICITLY AGREES.

At the end of each phase, ask: "Phase N complete. Ready for Phase N+1?"

Wait for "ok" or "yes" before continuing.

This is collaborative dialogue, not a checklist dump.


Phase 1: Discovery

Push yourself to think. Don't skip these.

First Principle — Find Actual Intention

Users communicate in solutions, not problems. "Add a spinner" might mean "page feels slow."

Ask yourself:

  • What did they say? (surface)
  • Why did they say it? (intention)
  • Is there a better approach?

Crux — Core Assumption

Every project has ONE assumption that must be true or everything falls apart.

Ask yourself:

  • What's the ONE thing that must be true?
  • How do we test it before building everything else?

Pre-Mortem — How This Fails

Imagine it failed. Why?

Ask yourself:

  • "It failed because..." (top 3 reasons)
  • What are early warning signs?
  • What can we do to prevent each?

Scope Fence — What This Is NOT

Scope creep kills projects. Define the boundaries.

Ask yourself:

  • This IS: [one sentence]
  • This is NOT: [explicit exclusions]
  • Features we're saying no to?

User Day-In-Life — Who Specifically

"Users" is too vague. Pick a real person.

Ask yourself:

  • Who specifically has this problem?
  • When in their day does it appear?
  • What do they do now? What's annoying about it?

Phase 2: Product

Think like you're explaining to a YC partner. Capture the essence.

What IS This Product?

Not the features. The VALUE.

Ask yourself:

  • If I had 30 seconds to explain this, what would I say?
  • What's the ONE thing that makes this valuable?
  • What's the core experience?

Core Workflows

Identify the workflows that ARE the product (not supporting stuff like auth).

For each core workflow:

  • Write the Description (YC partner level)
  • Write the Steps (the actual user experience)

This is human thinking. No technical details yet.


Phase 3: Design

Technical thinking. Consider alternatives, find simplicity, identify risks.

Approaches & Trade-offs

Don't jump to first solution.

Ask yourself:

  • What are 2-3 ways to solve this?
  • What are the trade-offs of each?
  • Which fits this context best? Why?

Simplest Thing That Works

Complexity is the enemy.

Ask yourself:

  • What's the minimal solution?
  • What can we NOT build?
  • What complexity are we avoiding?

Hard Parts / Where It Breaks

Every design has weak points.

Ask yourself:

  • What's the technically risky part?
  • Where will this fail first?
  • What needs extra attention during implementation?

Phase 4: Specification

Output the plan. Workflows first, always.

Output Format

markdown
# [Topic] Plan

## Summary
[What we're building, why, for whom — 2-3 sentences]

## Workflows

### WORKFLOW.id — Workflow Name
**Status:** planned
**Description:** [YC partner explanation — what this journey IS, why it matters]
**Steps:**
1. [Detailed step — what user does, what system does]
2. [Next step...]
3. [...]
**Depends on:** FEATURE.id, FEATURE.id

---

### WORKFLOW.id — Another Workflow
...

---

## Features

### FEATURE.id — Feature Name
**Status:** planned
**Description:** [YC partner explanation — what it does, user-facing]
**Technical Notes:** [How it works, gotchas, implementation details — enough for Claude to implement without asking questions]
**Uses:** INFRA.id, FEATURE.id

---

### INFRA.id — Infrastructure Feature
**Status:** planned
**Description:** [What it provides]
**Technical Notes:** [Technical details]
**Uses:** —

---

## Implementation Order

Group by commit. Distribute effort evenly — big features get own commit, small ones batch together.

### Commit 1: [Group Name]
- [ ] INFRA.database — setup connection pool
- [ ] INFRA.config — env-based config
(small, related → batch together)

### Commit 2: [Big Feature Name]
- [ ] AUTH.login — validate credentials, create session, handle errors
(complex feature → own commit)

### Commit 3: [Group Name]
- [ ] AUTH.logout — destroy session
- [ ] AUTH.refresh — refresh token
(small, related → batch together)

### Commit 4: [Integration]
- [ ] USER.login_flow — end-to-end test
(workflow integration test)

**Grouping rules:**
- Big/complex feature → own commit
- Small/simple features → batch with related ones
- Each commit should be test-able with REAL testing
- If a group feels too big to test confidently → split it

## Decisions
- [Decision]: [Why we chose this over alternatives]
- [Decision]: [Why]

Key Points for Specification

  • Workflows FIRST — Always
  • Self-contained detail — Description + Technical Notes + Steps should be complete enough that Claude can implement without further questions
  • Status markers — planned / in-progress / done
  • Commit grouping — Group by effort: big features alone, small features batched
  • Each commit must be testable — If you can't test it REAL, the group is wrong

After Plan Approval

1. Save Plan

Write to docs/plans/YYYY-MM-DD-<topic>.md

2. Create in Feature Tree

Create the workflows and features in Feature Tree:

code
add_workflow(id="...", name="...", description="...", purpose="...", steps=[...], depends_on=[...])
add_feature(id="...", name="...", description="...", technical_notes="...", uses=[...])

3. Sync to Memory

Use sub-skill: ft-mem:brainstorm-sync

Syncs discoveries to CONTEXT.md and memories.

4. Implementation

Use sub-skill: feature-tree:executing-plans

Implements the plan task-by-task with commits between.


Principles

  • Workflow-first, always — Broad context before details
  • Think, don't checkbox — Mind tools push real thinking
  • Self-contained output — Plan needs no further explanation
  • Simplest thing that works — YAGNI ruthlessly
  • Capture WHY — Decisions include rationale