Brainstorm

ADHD-friendly ideation with two decision points (not four) and one escape hatch (not a "switch depth" tree). Transforms a topic + project/git context into a BRAINSTORM-<topic>-<date>.md or, with the save action, a docs/specs/SPEC-<topic>-<date>.md.

Boundary With Adjacent Skills

Typical chain: brainstorm → SPEC → project-planner (breakdown) → adhd-workflow (spec-review/done).

Inputs

• Topic — explicit argument, or inferred from conversation / .STATUS / git branch / recent commits.
• Depth (quick | default | deep) — controls question count. No "max" tier — see "Going deeper" below for why.
• Focus (feat | arch | ux | api | ui | ops) — shapes output sections.
• Action (optional save) — capture as SPEC-<topic>-<date>.md.

Depth Budgets

Two decision points total per session: (1) depth+focus, asked together in one AskUserQuestion call when not provided as arguments; (2) one optional follow-up — "anything else before I generate this?" — offered once, after the expert questions, regardless of depth. No per-depth escape-hatch menus, no "switch depth mid-flow," no milestone re-prompting every N questions. If the user wants more, they ask; don't build a state machine for it.

Flow

1. Parse args / detect topic. If no topic, scan conversation, .STATUS, git branch, recent commits. 1 topic → use it; 2+ → one AskUserQuestion; none → ask free-form.
2. Pick depth + focus (if not provided as arguments) — single AskUserQuestion call combining both, not two sequential menus.
3. Context scan. Check for an existing SPEC or prior brainstorm on the topic, project type, .STATUS version, recent test failures. Pre-fill answers where project state already answers a question.
4. Expert questions. Per-depth count (0/2/6), from the question bank (full text and selection logic). 8 categories × 2 questions = 16 base questions, plus 6 project-type extensions (auto-detected via utils/claude_md_detector.py). Use AskUserQuestion.

Category override (-C / --categories)

Default category selection is focus-driven (see table below). Override with -C <list> / --categories <list> (comma-separated, matching commands/brainstorm.md frontmatter):

Default categories by focus:

Example: /brainstorm "caching" -C tech,risks limits expert questions to the technical and risk categories regardless of focus.

5. One follow-up offer. After the questions: "Anything else to add, or ready to generate?" — single yes/no-shaped prompt, not a menu of depth switches.
6. Generate output. Focus-specific sections (user stories for feat, Mermaid diagram for arch, API endpoints for api, etc.). Save to BRAINSTORM-<topic>-<date>.md.
7. Spec capture (action=save, or prompted for feat/arch/api). Ask user-story + acceptance criteria, render full SPEC template, write to docs/specs/SPEC-<topic>-<date>.md.
8. Going-deeper offer (optional, post-spec only) — see below.

Going Deeper: No In-Skill Agent Delegation

This skill does not spawn subagents itself. The pre-redesign version had a "max" depth that launched 1-2 background agents directly from inside brainstorm — duplicating what orchestrator-v2 already does (wave checkpoints, file-based results, explicit model routing, AskUserQuestion confirmation before spawning), but without those safeguards. It also referenced agent type names that didn't exist (backend-architect, product-strategist, etc.), which silently fell back to general-purpose with no model pin — see agents/orchestrator-v2.md BEHAVIOR 2.

Instead: after a spec is captured (Step 7), offer to hand off to the orchestrator via the existing --orch flag / plan-orchestrator skill / /craft:plan <spec-path>. One delegation mechanism, already hardened, instead of two.

Mandatory Interactive Steps

Even when all arguments are provided as flags, the per-depth expert-question count (0/2/6) is never skipped. Skipping silently is a known regression pattern. The depth/focus menu may be skipped when args supply them — the questions themselves may not.

Removed: Per-Invocation Question-Count Override

The pre-redesign version supported colon notation (d:5, m:12, q:0) to override the per-depth expert-question count on a single invocation. This is intentionally not carried forward — the fixed per-depth counts (0/2/6) are part of the same two-decision-point simplification described above ("Going Deeper"); a per-call override reintroduces the kind of escape-hatch menu this redesign removed. If a specific topic genuinely needs more than 6 expert questions, run /brainstorm at deep, then use the "anything else before I generate this?" follow-up offer (Flow step 5) to add more — don't request a count override.

Outputs

• BRAINSTORM-<topic>-<date>.md (always — even when also saving a spec).
• docs/specs/SPEC-<topic>-<date>.md (when save action selected).
• Terminal summary with paths and suggested next command.

Test-Plan Scaffolding (default-on)

When this skill emits a BRAINSTORM or SPEC artifact, it also emits a test-plan scaffold by default. Pass --no-tests to suppress the section.

Tier-inference rule

Infer tiers from the shape of the artifact/change being scoped:

Unselected tiers print as N/A — <reason> (never empty stubs).

Emission rules

• Emit test stubs red-first (failing placeholder, not passing no-op).
• Each stub carries # TODO(author): delete if not contract-bearing until the author confirms the contract.
• Scaffold templates live in ../brainstorm-insights/references/scaffold-templates.md (shared with orchestrate) — point to that file, do not duplicate templates inline.

--yes non-suppression

--yes auto-accepts prompts only; the test-plan and Documentation sections are content and are still emitted under --yes. Only --no-tests / --no-docs remove them.

Documentation Scaffolding (default-on)

When this skill emits a BRAINSTORM or SPEC artifact, it also emits a Documentation section by default. Pass --no-docs to suppress the section.

Which docs to emit

Derive which documentation artifacts are needed by running the existing doc-scorer rubric from skills/orchestration/references/doc-impact-rubric.md (threshold ≥3). Do not invent a new rubric — reuse the scorer as the single source of truth. For each doc type the scorer evaluates (guide, refcard, demo, mermaid), pre-check ([x]) boxes that meet the threshold, mark the rest N/A — score <N>.

Count-cascade exclusion

Auto-docs emission touches only semantic docs — CHANGELOG [Unreleased] mirror, guide/refcard/tutorial prose. It never touches version or count lines; those are excluded and stay in bump-version.sh.

Related Skills

• brainstorm-insights — friction/goals reporting from session history (was bundled with this skill; now separate, directory name kept for path stability).
• project-planner — breakdown after a SPEC exists.
• plan-orchestrator / --orch — multi-agent implementation, post-spec.