Think through what something should become — collaboratively, one question at a time — then write a right-sized requirements-only unified plan.
ce-brainstorm is the definition skill. It's a thinking partner that asks one question at a time, pressure-tests your premises against named gap lenses, explores 2-3 concrete approaches before recommending one, and, for software work, produces a requirements-only unified plan strong enough that planning never has to invent product behavior.
It runs equally well on software features, on entirely non-software topics (event planning, business decisions, personal-project framing, travel itineraries, naming briefs), and anywhere in between. The same one-question-at-a-time discipline applies everywhere. Software brainstorms write the requirements-only unified plan artifact; non-software brainstorms stay in facilitation mode and can hand their synthesis to ce-plan for a domain-appropriate plan.
This is the middle step in the compound-engineering ideation chain:
/ce-ideate /ce-brainstorm /ce-plan /ce-work
"What's worth "What does this "What's needed "Build it."
exploring?" need to be?" to accomplish
this?"
It's also the most common standalone entry point — for any feature, decision, or project where the question isn't "how do I do it?" but "what am I really doing, and is it the right shape?"
One thing it deliberately does not do is render a verdict. When a request is really a whether-to-adopt decision on a specific external candidate — a named technology, library, pattern, platform, or architecture, judged against this project — ce-brainstorm recognizes the shape (at intake or when a brainstorm later narrows to it) and offers to hand off to /ce-pov for a decisive, project-grounded verdict, rather than scope something the user hasn't decided to commit to. It's an offer, not a silent switch; open-ended design with no single candidate stays here.
| Question | Answer |
|---|---|
| What does it do? | Collaborative dialogue to clarify scope, pressure-test premises, explore approaches, and write a requirements-only unified plan |
| When to use it | Vague feature ideas, "let's brainstorm", multiple plausible directions, unclear scope; non-software decisions and projects |
| What it produces | Software: requirements-only unified plan in docs/plans/ with artifact_readiness: requirements-only and R/A/F/AE IDs. Non-software: chat synthesis plus optional handoff to ce-plan, not a software unified artifact. |
| What's next | /ce-plan, /ce-work for trivial scope, doc review, or publish to Proof |
Going straight from a vague idea to implementation produces:
- Code (or work) that solves the wrong problem because nobody pressure-tested the premise
- Scope creep because boundaries were never made explicit
- Plans that re-litigate product decisions every time someone touches them
- Requirements docs that are either over-ceremonial PRDs nobody updates, or one-line briefs that planning has to fill in by guessing
A typical "let's brainstorm" with an AI also has shape problems: it asks five questions in one message; you answer two and the rest get lost. It picks one approach immediately instead of presenting alternatives. It bakes implementation details into product discussion. The output is conversation, not a handoff-able artifact.
ce-brainstorm runs a structured but conversational flow that ends in a durable artifact:
- One question per turn, even when sub-questions feel related
- Right-sized ceremony — Lightweight / Standard / Deep / Deep-product tiers
- Named gap lenses force rigor on premises before generating approaches
- An opt-in blindspot pass maps the decision surface of territory the user doesn't know before the interview proceeds on it
- A background grounding scout gathers verbatim repo evidence on a cheap model while you answer the opening questions
- 2-3 concrete approaches with tradeoffs, then a stated recommendation
- Opt-in visual probes for decisions that are faster to judge as rough sketches than prose
- Synthesis Summary as the last opportunity to correct scope before the doc lands
- Fresh-context claim verification checks the doc's repo claims before it lands
- Right-sized Product Contract inside a unified plan with stable identifiers (R/A/F/AE) that flow into planning
Stacking three questions in one message produces diluted answers. ce-brainstorm asks one question per turn, every turn — and defaults to the platform's blocking question tool with single-select options when natural choices exist. Well-chosen options scaffold the answer without confining it (free-text fallback always available).
Not every brainstorm is the same. Lightweight covers small, well-bounded ideas with low ambiguity. Standard handles normal features with some decisions. Deep adds systemic-move probes for cross-cutting work. Deep-product additionally requires establishing product shape — actors, core outcome, positioning, durability — rather than inheriting it. Ceremony scales with the work, not against it.
Before generating approaches, the skill scans the user's opening for rigor gaps. Each gap has a name and probes the kind of confusion it catches:
- Evidence — "users want X" with no observable behavior backing it
- Specificity — beneficiary described abstractly; design will silently invent who they are
- Counterfactual — no visibility into what users do today, or what changes if nothing ships
- Attachment — a specific solution shape is being treated as the thing being built
- Durability (Deep-product only) — value rests on a current state of the world that may shift
These probes fire as prose, not menus — a 4-option menu signals which kinds of evidence count and lets the user pick rather than produce. Prose forces real observation.
Phase 2 surfaces 2-3 concrete approaches with at least one non-obvious angle — inversion, constraint removal, or cross-domain analogy. Approaches are presented at mechanism / product-shape granularity, not architecture. (Architecture decisions made on intentionally-shallow research tend to pre-commit you to bad choices; those belong in ce-plan.) Approaches are shown before the recommendation so the user sees alternatives without being anchored.
When a decision is spatial, behavioral, or visual, ce-brainstorm can offer a choice: keep exploring in text, or create a rough local visual probe. Visual probes are disposable sketches for product feedback, not polished UI, implementation prototypes, or final specs. The artifact is display-only; the user responds in chat.
Before writing the doc, ce-brainstorm emits a scoping synthesis shaped like what two product collaborators would confirm before writing a PRD. It surfaces what's being built, the key trade-offs the dialogue produced, what's been deferred, and any genuine forks the user should weigh in on. Each section renders only when it has something to say — no empty buckets padded for ceremony. The synthesis compresses to a single forward-looking sentence with an end-of-turn interrupt window only for a Lightweight brainstorm that asked no blocking questions. Every other case — Standard or Deep tier, or any tier that asked a blocking question — gets the full scoping synthesis with an explicit confirmation gate, including a richly pre-loaded Phase 0.2 fast path that needed no dialogue: pre-loaded substance earns a real checkpoint rather than a one-sentence preview.
The Product Contract carries plan-feeding identifiers — R-IDs (Requirements), A-IDs (Actors), F-IDs (Key Flows), AE-IDs (Acceptance Examples). ce-plan consumes these and traces every implementation unit and test scenario back to them. Origin scope boundaries (especially "Outside this product's identity") flow through unchanged.
Building a software feature? Standard flow. Naming a product? Choosing a vacation? Deciding a career move? ce-brainstorm routes to a domain-agnostic facilitator that preserves the one-question-at-a-time discipline and right-sized output.
Requirements describe what behavior is expected from the user's perspective. They do not describe libraries, schemas, endpoints, file layouts, or code structure — unless the brainstorm is itself about a technical or architectural decision. This keeps planning's job clean: invent the how, not the what.
The interview machinery assumes you can evaluate what it asks — and that assumption fails exactly when you're scoping work in territory you don't know. When you flag unfamiliarity ("I know nothing about the auth modules", "I don't know what color grading is"), or consecutive answers show you can't weigh the options rather than merely haven't decided, ce-brainstorm offers a blindspot pass before questioning you further on that territory: a grounded map of 3-7 decisions and hazards you didn't know to ask about, each with why it matters for your topic, the realistic options, and a recommended default. You pick which to walk through; the rest take defaults recorded as explicit assumptions. It converts unknown unknowns into known unknowns, so the interview extracts choices instead of guesses. Works on both the software and non-software routes.
On Standard and Deep brainstorms, a cheap extraction-tier scout is dispatched in the background while you answer the first question. It writes a grounding dossier — verbatim quotes with file:line pointers — to scratch storage and hands back a short gist, so the dialogue stays lean while the evidence stays available on demand. Before the requirements-only unified plan is written, a fresh-context verifier (a mid-tier model that never saw the dialogue) checks the Product Contract's repo claims — absence claims, file references — against the codebase, running while you review the synthesis confirmation. Refuted claims are corrected before the plan lands; unverifiable ones become explicit assumptions. The dossier path is handed to ce-plan so planning starts from verified quotes instead of re-scanning. On platforms without per-agent model selection, both run on the inherited model with the same read budgets; with no subagent support at all, the skill falls back to inline scanning and verification.
You start with a vague feature idea — "I want to add a way for users to pause notifications." ce-brainstorm reads the project's constraint files, classifies the work as Standard scope, and sends a cheap background scout to gather repo evidence while you answer the first question.
The pressure test detects a specificity gap (who are these "users"?) and an attachment gap ("pause" is already a specific solution shape). It probes both as prose, one at a time. You name the actual pain — your support team gets pinged at 3 AM for non-urgent stuff — and describe the smallest version that would solve it.
Three approaches surface — per-notification-type mute with TTL, a global do-not-disturb schedule, mute on the rule rather than the channel — with tradeoffs and a recommendation. The Synthesis Summary reads back the shape that emerged ("per-channel mute on notification rules, 24h preset for the 3 AM support pings"), names the trade-offs made in dialogue (per-channel over per-user, mute lives on the rule), what's deferred (presence-based mute, quiet-hours schedules), and a single call-out about the rule-delete loss path. You confirm and add a 24h preset.
A requirements-only unified plan is written under docs/plans/ and the Phase 4 menu offers next steps — /ce-plan (recommended), Product Contract review, publish to Proof, or skip-to-build only for trivial scope with a sufficient Definition of Done.
Reach for ce-brainstorm when:
- A feature idea is partially formed but you can't yet sketch the implementation
- A request has multiple valid solutions and you need to choose
- The scope is unclear ("add notifications" — what kind? for whom? when?)
- You want a structured artifact you can hand to another contributor or to planning
- A vague problem statement needs to become a real product decision
- You have to scope work in territory you don't know — the blindspot pass maps the decision surface before questions begin
- You're working on something non-software (named products, roadmap choices, decisions)
Skip ce-brainstorm when:
- You don't yet know what to work on →
/ce-ideatefirst - Requirements are already specified (PRD exists, GitHub issue is detailed) →
/ce-plandirectly - You have a known root cause for a bug →
/ce-debug - The change is trivial and obvious → just do it
/ce-ideate (optional — discover candidate directions)
| picks one survivor + carries warrant + rationale
v
/ce-brainstorm
| produces requirements-only unified plan
| software mode: R-IDs, A-IDs, F-IDs, AE-IDs + scope boundaries
| universal mode: a domain-appropriate brief
v
/ce-plan
| enriches the same plan to implementation-ready
| R-IDs flow into Requirements; A/F/AE-IDs trace into units and tests
| origin scope boundaries are preserved verbatim
v
/ce-work
When ce-plan loads with a requirements-only unified plan as input, it does not re-litigate product behavior. The Product Contract is authoritative. Plan-time decisions are about execution guardrails — not what's being built.
ce-brainstorm is the most common standalone entry point. Many teams skip ce-ideate (they already know what to explore) and skip ce-plan (the brainstorm is their full thinking artifact).
- Feature briefs — turn a vague idea into a stable artifact for stakeholders or new contributors
- Onboarding existing work — when a feature is in flight but the rationale was never written down
- Pre-PR alignment — when multiple people need to agree on scope before code starts
- Strategic decisions — Deep-product tier surfaces durability and adjacent-product risks
- Non-software brainstorms — name a product, plan an event, decide a roadmap
The Phase 4 handoff offers planning, agent doc review, publish to Proof, direct-to-work for lightweight scope, more clarifying questions, or pause.
| Argument | Effect |
|---|---|
| (empty) | Asks for the feature description |
<feature idea> |
Open-ended brainstorm |
<problem> |
Routes via the product pressure test |
Existing requirements-only plan path, legacy *-requirements.md path, or topic |
Resume offer |
output:html |
Write the requirements-only unified plan as a single self-contained HTML file instead of markdown. Exclusive — the artifact is .md OR .html, never both. Default is markdown. Set brainstorm_output: html in .compound-engineering/config.local.yaml to make HTML the default. Pipeline mode (LFG, disable-model-invocation) always forces markdown so downstream automation gets a stable text shape. |
Why one question at a time? Isn't that slow? Stacking three questions per turn produces diluted answers — users pick the easy one and the rest get lost. One question per turn produces sharper answers and is empirically faster to convergence.
Why does it pressure-test my premise? I just want to brainstorm. The named gap lenses catch the most common ways feature briefs fail downstream. They fire only when the gap is actually present in your opening — a concrete, well-framed prompt may earn zero probes.
Can I skip the requirements-only plan? Yes. The Lightweight tier and the announce-mode fast path support that. If you only need brief alignment, no doc is written.
What if I already have a PRD or detailed GitHub issue?
Skip ce-brainstorm and go directly to /ce-plan. The plan skill consumes any kind of input.
What does "Inferred" mean in the synthesis? The agent composes an internal three-bucket draft (Stated / Inferred / Out of scope) as a thinking step before presenting the scoping synthesis. Inferred items are the agent's bets to fill dialogue gaps. Those that survive the keep test surface as call-outs in the scoping synthesis; the rest dissolve into the Product Contract when the user confirms.
Does it work for non-software topics? Yes — a domain-agnostic facilitator preserves the one-question-at-a-time discipline and right-sizing for naming, decisions, planning, etc.
When you're on a cheaper session model, ce-brainstorm can still generate approaches with a higher-reasoning model: it dispatches approach generation to Fable via a subagent, so you get sharper, less-generic options without switching your whole session. Opt in per run by saying "use fable" in your prompt, or set brainstorm_use_fable: true in .compound-engineering/config.local.yaml. A mechanical host gate makes this a silent no-op on every non-Claude-Code harness (Codex, Cursor). See references/reasoning-elevation.md.
ce-ideate— upstream "what's worth exploring" discoveryce-plan— enrich the requirements-only unified plan into an implementation-ready plance-doc-review— persona-based review of the Product Contract or full plance-work— execute lightweight changes directly from a brainstormce-strategy— anchor brainstorms to a documented product strategy