BrainstormSAFE

The Agent Harness for AI-Human Collaboration, inspired by the AI-DLC (AI-Driven Development Lifecycle)

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: brainstorm
description: Optional divergent-then-convergent dialogue for fuzzy ideas. Invoked from the idea skill as a prelude to structured elaboration; produces one ElaborationRound of decision-point Q&A and returns control. Never writes files, never posts comments, never validates elaboration.
license: AGPL-3.0
metadata:
  author: chorus
  version: "0.22.1"
  category: project-management
  mcp_server: chorus
---

# Brainstorm Skill

A divergent-then-convergent dialogue cadence for ideas whose direction is still being formed. Compresses the conversation into one `ElaborationRound` of decision-point Q&A — same shape as a structured elaboration round, but the questions, options and answers are synthesized at the end of the conversation rather than asked up front.

This skill is a **producer** of one elaboration round; the **scheduler** decision (validate vs. follow-up) belongs to the calling idea skill.

> **Tool namespace:** Chorus tools are exposed by the connected MCP server under a `chorus__` prefix on OpenClaw (e.g. `chorus__chorus_get_idea`). Bare names are used below for readability — prepend `chorus__` when invoking. See `/chorus` for the full rule.

> **OpenClaw interaction model:** OpenClaw has no `AskUserQuestion` primitive. Every "ask the user" step below is a **plain-text prompt** — write the question (and options) in your message and wait for the user's free-text reply before continuing. The one-question-at-a-time cadence still holds: send one question, read the answer, then send the next.

---

## When invoked

Only as a sub-step of the idea skill, only after the user has explicitly opted in (via the idea skill's plain-text Brainstorm prompt). Never run standalone, never run without user opt-in. The expected entry point is the idea skill's "Step 4.5: Brainstorm Mode (Optional Prelude)" — see the idea skill for the surrounding flow.

---

## Research boundary

Reuse the calling Idea's findings and follow `/research` ([shared rules](../research/SKILL.md)); switching into brainstorm does not grant a second investigation. If the goal needed focusing first, after the user selects a direction and before synthesis, apply the caller's optional research decision once with that focus. Return any findings alongside the synthesized round for the Idea caller to persist. Research does not choose the user's direction, remove the opt-in/selection gate, write files/comments, or resolve elaboration.

## Hard rules

1. **One question at a time.** Each prompt MUST contain exactly one question. Wait for the user's reply before asking the next.
2. **Multi-choice preferred.** Frame each question as 2-4 options where possible (offer them as a lettered list in your plain-text prompt). Open-ended is acceptable when options would be premature, but lean toward concrete choices.
3. **Propose 2-3 directions before stopping divergence.** Once the requirement direction is clear enough to enumerate, present 2-3 distinct approaches in a single prompt. Mark exactly one as the recommended option and say why (the dominant tradeoff).
4. **Explicit user approval required to exit divergence.** Do NOT proceed to synthesis until the user has selected one of the proposed directions in their reply.
5. **No files written.** Do NOT write any markdown, design doc, scratch file, or any other file to disk. The conversation produces an `ElaborationRound` and nothing else on disk.
6. **No comments posted.** Do NOT call `chorus_add_comment` from this skill. Comments belong to the idea skill or the user, not to the brainstorm step.
7. **No design-doc handoff.** Do NOT invoke `writing-plans`, `writing-skills`, or any skill whose purpose is to produce a design document. The brainstorm output is the synthesized round — there is no separate doc.
8. **No `validate_elaboration` call.** Do NOT call `chorus_pm_validate_elaboration` from this skill. Whether to resolve the elaboration or open a follow-up round (`chorus_pm_start_elaboration` again) is the calling idea skill's decision, not this skill's.

---

## Step-by-step

### 1. Gather context

Before asking the first divergent question, read the idea and surrounding project state. Mirror the idea skill's gather-context list:

```
chorus_get_idea({ ideaUuid })
chorus_get_documents({ projectUuid })
chorus_get_document({ documentUuid })   # for any document worth reading in full
chorus_get_proposals({ projectUuid, status: "approved" })   # to understand patterns
chorus_list_tasks({ projectUuid })   # to avoid duplicating existing work
chorus_get_comments({ targetType: "idea", targetUuid: ideaUuid })
```

Skim each result for: stated background, stated requirements, stated constraints, and what is conspicuously NOT stated. The gaps are the questions worth asking.

### 2. Divergent Q&A

Ask one question at a time as a plain-text prompt. Aim to surface:

- The **goal** the idea is trying to serve (often more abstract than the idea statement).
- The **constraints** that exclude entire branches of solution space (deadlines, compatibility, scope).
- The **success criteria** — how will the user know this is done.

Keep each question single-purpose. If you need to ask three things, that is three prompts, not one combined message.

### 3. Propose 2-3 directions

When the goal, constraints, and success criteria are clear enough that you can name distinct approaches, present them in a single plain-text prompt:

```
Based on what you've told me, here are three directions. Which do you want? (reply A / B / C, or describe your own)

A) <Option A — RECOMMENDED> — <what + tradeoff>
B) <Option B> — <what + tradeoff>
C) <Option C> — <what + tradeoff>

I recommend A because <one sentence about the dominant tradeoff>.
```

State **why** you recommend one option — usually a sentence about the dominant tradeoff.

### 4. Wait for explicit approval

Do not proceed to synthesis if the user has not selected one of the options in their reply. If the user picks "Other" with free text, treat that as a n

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (1)

.agents/skills

Gates applied: no_behavioural_pass.