parallel-brainstorming
/parallel-brainstorming
Use when requirements are vague or the solution space is open before a plan exists. Prefer over request-plan when two or more distinct architectural approaches are in play.
- Author
- j0hanz ↗
- Repository
- squad-agents ↗
- Licence
- MIT
- Last repository change
- 18.07.2026
Assessment and security
- Files in this package
- 3
- Scripts
- 2
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 18.07.2026, without references to retired models or APIs.
MIT, open: use and modify while retaining author attribution.
- 2 scripts in this package. Review them before installing, as Claude may execute them.
Scripts: scripts/compress_report.py, scripts/scan_context.py
What it does and when to use it
Use when requirements are vague or the solution space is open before a plan exists. Prefer over request-plan when two or more distinct architectural approaches are in play.
Original author description
Use when requirements are vague or the solution space is open before a plan exists. Prefer over request-plan when two or more distinct architectural approaches are in play.
How it works
- Process Flow
- Phase 1: Framing & Discovery
- Phase 2: Clarification
- Creative Checkpoint (Pre-Ideation)
- Phase 3: Multi-Lane Divergent Ideation
- Phase 4: Convergence & Synthesis
- Phase 5: Persona Critique
- Phase 6: Design Brief
- Worked Example
- Strict Rules
SKILL.md preview
---
name: parallel-brainstorming
description: 'Use when requirements are vague or the solution space is open before a plan exists. Prefer over request-plan when two or more distinct architectural approaches are in play.'
disable-model-invocation: false
allowed-tools: AskUserQuestion, Write, Bash(python *), Bash(python3 *), Read, Grep, Glob
---
# parallel-brainstorming
<HARD-GATE>
Do not propose code, file changes, or a concrete implementation plan for a new feature or ambiguous
request until Phase 6 produces an approved Design Brief. Sketching the approach in a doc is still
design work — it still needs Phase 1 Discovery first. Does not apply to bug fixes, typos, or one-line
config changes with no design space.
</HARD-GATE>
## Process Flow
1 → 2 → 3 → 4 → 5 → 6
## Phase 1: Framing & Discovery
- **No Silent Skips:** If a task needs zero discovery, name the exact step skipped (Probe, Scan, or Understanding Lock) and explain why — never skip silently.
- **Probe:** Identify target users; ask clarifying questions if the request is ambiguous.
- **Scan:** Run `python ${CLAUDE_SKILL_DIR}/scripts/scan_context.py '<nouns>' --cwd '<root>' | python ${CLAUDE_SKILL_DIR}/scripts/compress_report.py` (fallback to `Grep` if it fails).
- **Report:** Extract Related Files (with recent commits and test coverage), Interface Shapes, Design Docs, Analogous Features, Constraints, Scope (S/M/L/XL) with reasoning, and Unknowns.
- **Zero-Code Check:** Stop and offer exit if existing code/config already solves this.
- **Understanding Lock:** Summarize the problem and your understanding. Only ask the user (via `AskUserQuestion`) if you have genuine doubts; otherwise proceed directly to Phase 3.
- **Routing:**
- Scope XL → Offer to break it down.
- Ambiguous → Go to Phase 2.
- High Risk / Scope L+ → Set Phase 5 Flag.
**Done when:** a Context Report lists Related Files, Interface Shapes, Design Docs, Analogous Features, Constraints, Scope (S/M/L/XL), and Unknowns, and the zero-code check has been answered.
## Phase 2: Clarification
- **Resolve with the user:** clarify the ambiguous terms via `AskUserQuestion` (max 4 per batch), keeping questions to a couple of options each.
- **Glossary:** Save resolved definitions to `glossary.md` (never `CONTEXT.md`).
- **Visuals:** Offer a diagram only if layout or data flow requires it. Wait for a reply.
**Done when:** ambiguous terms are resolved with the user (max 4 per batch) and saved to `glossary.md`.
## Creative Checkpoint (Pre-Ideation)
- **Evaluate:** Look for a 10x simpler or zero-code solution.
- **Seed:** If found, use it as "Approach A" (Minimalist lane) in Phase 3.
**Done when:** a 10x/zero-code candidate is seeded as Approach A, or you've confirmed none exists and are proceeding to Phase 3 unseeded.
## Phase 3: Multi-Lane Divergent Ideation
- **Single-Shot Generation:** Generate 2-3 distinct approaches in one response. **Do not spawn subagents.**
- **Context:** Use the feature description and Context Report to inform all perspectives.
- **Lenses (Assign 1 per agent):**
1. _Conventional:_ Use existing codebase patterns.
2. _Radical:_ Best outcome, ignoring legacy constraints.
3. _Minimalist:_ Smallest working change (Seeded by Checkpoint).
4. _Constraint-First:_ Optimize for the hardest non-functional constraint (e.g., speed, scale).
5. _Analogous:_ Copy and adapt a similar existing feature.
- **Output (Per Agent):** Idea, core mechanism, winning factor, key risk, first step.
**Done when:** 2-3 distinct approaches are generated in one response, each with idea, core mechanism, winning factor, key risk, and first step.
## Phase 4: Convergence & Synthesis
- **Synthesize:** Group similar ideas. Combine strong mechanisms with risk-mitigations from other lanes.
- **Distill:** Present 2-3 distinct approaches. Approach A must be Minimalist. For each: What, Gains, Costs, Fit, First Step.
- **Approval Lock:** Present the 2-3 distilled approaches to the user via `AskUserQuestion` to lock one — this is the hard-to-reverse decision that commits Phase 6's Design Brief. **Await their decision. Do not guess.**
- **Routing:** If Phase 5 flag is set → Phase 5. Otherwise → Phase 6.
**Done when:** the user locks one of the 2-3 distilled approaches (not guessed).
## Phase 5: Persona Critique
- **Trigger:** Phase 5 flag is set, or user requested a stress test.
- **Simulated Review:** Adopt 3 personas in your thought process to evaluate the chosen design:
1. _Skeptic:_ Finds edge cases and failure modes.
2. _Constraint Guardian:_ Enforces scale, performance, and security rules.
3. _User Advocate:_ Evaluates usability and cognitive load.
- **Severity Rating:** High (Blocks deployment), Med (Worse outcome), Low (Minor). Ignore styling/naming.
- **Resolution:** Record objections. For all High/Med issues, you must "Accept & Revise" or "Reject with technical rationale."
- **Self-Arbitration:** Resolve any debates yourself. Mark the design `APPROVED`, `REVISE`, or `REJECT`.
**Done when:** every High/Med objection is "Accept & Revise" or "Reject with technical rationale", and the design is marked `APPROVED`, `REVISE`, or `REJECT`.
## Phase 6: Design Brief
- **Self-Review:** Fix any contradictions or scope creep in the chosen design before writing.
- **Format:** Write a strict `markdown-kv` brief containing: Approach, Why, Scope, Constraints, Interface, Architecture, Risks, First Step.
- **Save:** Present in chat, then write to `docs/design/YYYY-MM-DD-<topic>-design.md`.
- **Commit Guard:** Do not commit as part of brainstorming. If the user wants to commit (and optionally push / open a PR), do it directly with git/gh once the Design Brief is approved.
**Done when:** the markdown-kv Design Brief (Approach, Why, Scope, Constraints, Interface, Architecture, Risks, First Step) is written to `docs/design/YYYY-MM-DD-<topic>-design.md`.
## Worked Example
Request: "add a way for users to save and re-run searches."
1. **Phase 1:** Scan finds an existing `Filter` model and a one-off "recent searches" list in `localStorage`. Scope: M. No flag (not high-risk, not L/XL).
2. **Creative Checkpoint:** Minimalist seed found — extend `Filter` with a `name` + `saved: boolean` column instead of a new table.
3. **Phase 3 (Multi-lane generation):** Conventional — new `SavedSearch` table + CRUD API, mirrors `Bookmark` feature. Minimalist — reuse `Filter` + 2 columns, no new endpoints (piggyback on existing filter-list endpoint). Constraint-First — same as Minimalist but adds a per-user cap (20 saved searches) to bound query cost.
4. **Phase 4:** Synthesize 2 approaches — Approach A (Minimalist + cap, cheapest) and Approach B (Conventional, more flexible but a new table + endpoints). User picks A. Not flagged → skip Phase 5.
5. **Phase 6:** Design Brief written to `docs/design/2026-06-29-saved-searches-design.md`: Approach (extend `Filter`), Why (reuses existing model, smallest diff), Scope (M), Constraints (cap 20/user), Interface (`Filter.saved`, `Filter.name`), Architecture (no new table), Risks (cap needs a migration default), First Step (`ALTER TABLE filters ADD COLUMN saved boolean DEFAULT false`).
6. Commit Guard: user declines auto-commit → brief left in chat + on disk; handoff to `request-plan` to formalize tasks.
## Strict Rules
- **Blend Ideation:** Keep Phase 3 perspectives distinct; do not bleed them into each other until Phase 4 synthesis.
- **Ship Raw Ideas:** Phase 4 synthesis is mandatory. Never present raw brainstormed ideas as the final answer.
- **Accept Empty Rejections:** Require a technical reason for any rejected High-severity issue during Phase 5 critique.
- **No Agent tool for Phase 3 or 5.**
## Next Skills
| Skill | Use Case |
| :--------------------------------------------- | :----------------------------------------------------------------------- |
| [request-plan](../request-plan/SKILL.md) | Formalize the Design Brief into a task plan when auto-commit is declined |
| [dispatch-agents](../dispatch-agents/SKILL.md) | Execute the resulting plan once tasks are formalized |
Files in this package (3)
- SKILL.md 8 KB
- scripts/compress_report.py 6 KB
- scripts/scan_context.py 16 KB
Versions
The database has no other skills named parallel-brainstorming.
Database snapshot: 13.11.2026 · package version 4f7f0b6cb21b (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 90brainstorm
Use when requirements are vague, the solution space is open, AND no deliverable shape has been chosen yet, before a plan exists. Not for a named deliverable (plan/spec/doc) — use plan. - 100plan
Start here — use when the user wants to build something new or plan a feature. Collaborative brainstorming that explores requirements, constraints, and approach through natural dialogue before any spec is written. - 100phx-brainstorm
Brainstorm Elixir/Phoenix features — explore ideas, compare approaches, gather requirements. Use when vague idea, not sure how to approach, or want to discuss before plan. - 96brainstorming
在任何创造性工作之前必须使用此技能——创建功能、构建组件、添加功能或修改行为。在实现之前先探索用户意图、需求和设计。 - 100brainstorming
当用户明确要求"使用 brainstorming"或"使用 awesome-code"时使用。⚠️ 不适用:用户只是想优化/改进某个功能(应直接修改)、只是询问技能问题(应直接回答)、没有明确使用 brainstorming/awesome-code 的一般性开发。 - 100request-plan
Use when a new feature, change, or repository modification requires a plan or specification. Not for unproven problems with open solution space — use parallel-brainstorming.