Use before code when brainstorming a feature, refactor, or architecture change. Reads the substrate the codebase already has, captures current scope and future pressure separately, proposes 2–4 design options grounded in that substrate, then attacks each option through a 25-question pressure-test battery before recommending one. Triggers on "brainstorm a refactor of X", "design a feature for X", "what's the right way to add X", "how should we restructure X", "I'm thinking about Y, what do you think". Always pair with discover-substrate first; never produce code from this skill.
Use before code when brainstorming a feature, refactor, or architecture change. Reads the substrate the codebase already has, captures current scope and future pressure separately, proposes 2–4 design options grounded in that substrate, then attacks each option through a 25-question pressure-test battery before recommending one. Triggers on "brainstorm a refactor of X", "design a feature for X", "what's the right way to add X", "how should we restructure X", "I'm thinking about Y, what do you think". Always pair with discover-substrate first; never produce code from this skill.
Brainstorm design
What this skill produces
A substrate-grounded design recommendation that:
captures current scope, future pressure, and non-goals separately
explores the option space — in one shot when shallow, as a multi-turn axis dialog when the space forks along ≥2 structural dimensions
pressure-tests the chosen direction through the question battery
recommends one option (or named hybrid) with main risk and structural mitigation
lists the substrate that must be created or updated before implementation begins
The output is the input to either rewrite-specs (if a direction is approved) or another round of brainstorming (if no direction survives pressure-testing).
Voice
Read ${CLAUDE_PLUGIN_ROOT}/references/output-voice.md before rendering chat output. The voice guide is the load-bearing source for verdict-leads, header-depth cap, density budgets, and forbidden phrasings; the imperative above is what triggers the model to load it via a Read tool call. Do not reproduce the imperative or any citation to the voice guide inside the Output format render template — instructions placed inside render templates leak verbatim into user-facing output.
Hard constraints
Never produce code from this skill. Not a snippet, not a function signature. Brainstorming ends at "here's the recommended direction."
Substrate discovery is internal to this skill. Step 0 of the Process dispatches cohesive:discover-substrate via the Skill tool with the change surface extracted from the user's brainstorm topic. The user never sees discovery output as a separate render — it's a sub-step that produces a persisted report this skill consumes. The router routes design to this skill directly (not chained discover-substrate → brainstorm-design); the discovery dispatch happens here, where it is plumbing for the brainstorm rather than user-visible output. Optional override: if the dispatch prompt names a discovery report path that's already been produced (e.g., another consumer skill ran discovery in the same session, or the user explicitly invoked cohesive:discover-substrate first), this skill reuses that path instead of re-running discovery.
Always propose at least two credible options for non-trivial changes. Single-option "design" is just a proposal, not a decision.
Never recommend an option whose main risk is mitigated by "we'll be careful." Mitigation is structure: a test, a linter, a boundary, a constraint.
Conversational mode is multi-turn; each turn asks at most one forced-choice question. Each conversational turn presents a verdict-led pick on one decision and asks the user to ratify or redirect — never a vague "what do you want?" prompt. Forbidden phrasings from ${CLAUDE_PLUGIN_ROOT}/references/output-voice.md apply per-turn. Render the per-turn forced-choice prompt (axis pick, sub-decision Pick/Confirm) through AskUserQuestion per ${CLAUDE_PLUGIN_ROOT}/references/output-voice.md §"Forced-choice questions".
Process
Step 0: Dispatch substrate discovery (internal)
Before grounding the brainstorm, dispatch cohesive:discover-substrate via the Skill tool with the change surface extracted from the user's request (the topic of the brainstorm). discover-substrate runs as a sub-step, persists its full report to disk, and returns a path. This skill consumes the path as input to Phase 1; the user does not see discover-substrate's chat output.
Skip condition: if the dispatch prompt to this skill includes "Discovery already complete; report at ", do not re-dispatch — read the named report directly. This handles three cases: (a) the user explicitly invoked cohesive:discover-substrate before brainstorm-design and the report exists; (b) another consumer skill ran discovery earlier in the same session; (c) the router (in legacy invocation patterns) passed the prereq state explicitly.
Change-surface clarification: if the user's brainstorm topic is unclear at the change-surface level (e.g., "design something for auth" without naming a subsystem), ask one precise clarifying question naming the candidate change surfaces from the repo's directory structure: "Which subsystem is the brainstorm about: , , ?" This is the canonical clarifying question for brainstorm-design — distinct from the retired discovery-state question that lived here pre-2026-05-06.
Phase 0: Resolve the artifact directory
If the brainstorm is going to be persisted (the user has asked for it, or the router's design route is chaining toward rewrite-specs), resolve where it will be written before grounding begins:
If docs/cohesive/brainstorms/ exists, write there.
Else if the repo carries docs/adr/, docs/specs/, docs/design/, docs/decisions/, or docs/architecture/, write to a brainstorms/ subdir alongside it.
Else default to docs/cohesive/brainstorms/.
If docs/ does not exist, still default to docs/cohesive/brainstorms/.
Brainstorms are not always persisted — many design conversations end at the recommendation. When persistence is requested, announce the resolved path in chat before writing.
Phase 1: Ground the brainstorm
Use the discovery report from Step 0 (the persisted report path returned by the internal cohesive:discover-substrate dispatch) as the starting material. If the discovery report carries **Empty-substrate verdict: yes**, broaden option-generation to fundamentals rather than grounding in nothing — this is a fresh-substrate codebase, not a mature one.
Then, in chat, capture three things separately (plus a fourth optional category for re-decide cycles):
Current scope
What this change must accomplish. The behavior the user is asking for.
Future pressure (not current scope)
Things the user has hinted at or that the substrate suggests will matter later — but that should not become current-scope features just because they're foreseeable. Capturing future pressure separately is the mechanism that prevents speculative implementation.
Non-goals
What this change explicitly does not do. Helps the options stay focused.
What we already tried (optional — re-decide cycles only)
When this brainstorm is invoked from a cohesive:validate-rewrite Approved trailer's Re-decide option (the user read the architectural reflection and judged the locked design unsound), capture the discarded direction and what made it feel wrong. This input is the substrate residue of a failed lock — without it, the next round of options re-derives the same path that just got discarded.
The dispatching validate-rewrite invocation passes this input as part of the Skill-tool dispatch prompt. The shape:
Discarded direction: <option name + 1-2 sentence summary from the discarded brainstorm's ## Direction block>
Reflection's harder-downstream concerns: <bullets from the discarded validate-rewrite Approved trailer's Architectural reflection §"Harder downstream">
Reflection's load-bearing-on-memory concerns: <bullets from the discarded validate-rewrite Approved trailer's Architectural reflection §"Load-bearing on memory">
Re-decide pass count: N (incrementing across the cycle; cap at 2-3 per the **Re-decide acknowledgment** convention in ${CLAUDE_PLUGIN_ROOT}/skills/validate-rewrite/SKILL.md §"Output format")
The brainstorm uses these inputs to bias option-generation: any new option must either resolve the harder-downstream concerns of the discarded direction or explicitly accept them with a different structural mitigation. Options that re-derive the discarded path without addressing its concerns are out of scope. The pressure-test battery in Phase 4 attacks the new options against the discarded reflection's concerns, not just against the substrate at large.
If any of the four input categories is unclear, ask one precise clarifying question. Suggested forms: "Which future pressure should this design optimize for most: , , or ?" (when future pressure is the unclear input) or "Which of the prior reflection's concerns is the most load-bearing for this re-decide: or ?" (when re-decide-cycle inputs are unclear).
Question precedence. Step 0's change-surface clarification (Hard constraint #2) takes precedence over Phase 1's future-pressure / re-decide question. If the change surface is unclear, ask Step 0's question first; Phase 1's question only fires after the change surface is known and the discovery report is in hand.
Phase 2: Identify axes and select mode
Identify the axes of disagreement the substrate exposes — the orthogonal structural dimensions along which credible options would differ. Common axes:
Centralization — per-component vs unified
Timing — synchronous-with-state-change vs async/eventual
Locality — context-required-to-change-this-area increases or decreases
Enforcement substrate — added invariants/linters vs trust-and-document
Migration shape — in-place vs parallel-path vs flag-gated cutover
Axes are internal at this stage — they determine mode but are not yet user-rendered. Naming them explicitly inside the agent's working state is what makes the auto-detect gate well-defined.
Then apply the auto-detect mode gate:
Signal
Mode
≥2 distinct axes of disagreement, OR option space touches ≥2 substrate kinds (specs + invariants + linters)
Conversational (Phase 3b)
Single-axis option space, naming, placement within an established pattern, single-pattern refactor
Autonomous (Phase 3a)
User override at any phase boundary: "give me the autonomous version" / "walk me through it." Auto-detect is a default, not a lock. The table form (rather than binary if/else) makes a future third mode a sibling row, not nested prose.
Phase 3a: Propose options (autonomous mode)
Generate 2–4 named design options. Each option must:
Have a short, memorable name (not "Option A" alone — name the idea, e.g., "connector-local classification" or "shared decision kernel")
Describe the core architectural choice in 1–2 sentences
Be substantially distinct from the other options (not just parameter variations)
For each option, immediately note:
Substrate changes required — which docs/matrices/invariants/tests/linters would need to be added or updated
Locality impact — what context a future change to this area would require
Future fit — which future pressure this option makes easy or hard
Risks — initial risks before pressure-testing
Then proceed to Phase 4 (Pressure-test) and Phase 5 (Recommend) — autonomous mode renders the trailer in a single turn.
Multi-turn axis dialog → leaf direction summary → sub-decision routing. Axes and sub-decisions are different layers: an axis is a top-level structural fork where any pick reshapes the leaf design; a sub-decision is an open call within the chosen leaf, classified by who decides.
The first user-visible turn in conversational mode is the axes map — this makes the auto-detect gate's "≥2 axes" determination auditable from the chat surface and the persisted file's ## Decision dialog §"Axes walked".
Axis dialog
Open with a single short turn rendering the axes map — a one-line structural overview:
"These split along <axis 1> (one family) × <axis 2> (another). I'll walk through them in order of where your taste matters most."
Then drill axes one at a time. Order: user-impact-if-redirected first (taste-driven, future-pressure-shaping, organizational/risk axes), substrate-evidence-strength as tiebreaker (strongest agent pick last, so prior decisions shape the leaf direction). Each axis turn renders in this order:
Name the axis in a short header.
Render the option space as parallel labeled blocks — one option per block, blank line between blocks, with Pro and Con sub-bullets so benefits and costs are visible side-by-side. Each block opens with a bold label + shape clause as its own paragraph; sub-bullets render Pro and Con. For options that are structurally blocked (not credibly viable), substitute Status: for Pro/Con and name the structural reason — forcing a fake Pro on a blocked option is theater. Format:
Axis: <axis name>
A: <short name> — <shape in one clause>.
Pro: <load-bearing benefit — what would make someone pick this>
Con: <load-bearing cost — the structural failure mode if you went this way>
B: <short name> — <shape>.
Pro: <…>
Con: <…>
C: <short name> — <shape>.
Status: <structurally blocked — name the reason>
Pro and Con must be load-bearing structural consequences (not nice-to-haves). If you can't name a real Pro for an option, it isn't credible enough to enumerate. Showing Pro upfront — alongside the costs — prevents biasing the user toward the chosen option through cost-only enumeration; rule 2b applies to mid-dialog turns just as to trailers. The blank lines between option blocks are load-bearing for visual scannability — markdown bulleted lists collapse vertically and lose the side-by-side structure the user needs to compare.
Render the pick + substrate evidence in verdict-led voice — "I'd pick X because Y" with path:line / gotcha citation / invariant name as the substrate evidence.
Render counter-pressure for the strongest alternative (the best argument for it, not a strawman) — and for any option you're rejecting outright (like a structurally blocked one), say so explicitly with the structural reason.
Close with a forced-choice prompt: ratify, push back with substantive counter-pressure, or redirect.
If the user pushes back on an axis, that axis opens into one additional turn where the agent surfaces deeper substrate evidence and re-decides — same render shape (option space + pick + counter-pressure + forced choice), just with the substrate-evidence section deepened. Don't loop further; if disagreement persists after that exchange, the user is the Decider and their pick stands. The agent records the divergence for Phase 4's cross-branch graft check.
Leaf direction summary
After all axes converge, render a one-turn summary:
"OK — that lands us at . Open sub-decisions: [1] [Pick], [2] [Confirm — I'd pick X because Y], [3] [Default — Z]. Want to start with the Pick, push back on the Confirm, or challenge a Default?"
Sub-decision routing
Each open sub-decision inside the chosen direction is tagged in chat. Tags are Pick / Confirm / Default — not Decide/Lock/Build, which are the gate-vocabulary tokens reserved for the methodology's three-gate user-facing model:
Pick — multiple substrate-coherent paths exist; the call shapes future-pressure outcomes; or there's an organizational/risk-tolerance dimension. The user picks; the agent surfaces options and tradeoffs but does not pick.
Confirm — substrate evidence points clearly to one answer with low organizational impact. The agent recommends; the user ratifies or pushes back.
Default — mechanical follow-on (naming, placement within an established pattern); one option is overwhelmingly conventional and reversible. The agent applies the default, surfaces it, and the user can promote any Default to discussion.
Default classification posture: aggressive Default. A sub-decision is material (Pick or Confirm) only if different choices would land on different specs / invariants / tests / linters / matrices — not different identifiers, file names, or placements within an established pattern. Tighter dialog beats thoroughness theater. The Default list is visible in the leaf-direction-summary turn so the user can promote any item, but Defaults don't block convergence.
Sub-decisions surface together (in the leaf-direction-summary turn) so the user sees the open-seam set in one render. Within that set, drilling proceeds one decision at a time as the user picks/ratifies/challenges. Backtrack to any earlier axis or sub-decision is always available — "back up to , I want to revisit" — and re-enters the dialog at that point.
Once all sub-decisions converge — Picks made, Confirms ratified, Defaults left in place — proceed to Phase 4.
Phase 4: Pressure-test
Apply the question battery from ${CLAUDE_PLUGIN_ROOT}/references/design-pressure-testing.md to the surface this mode produced:
Autonomous mode — apply to every option from Phase 3a. Don't skip the one you already prefer; attack it as hard as the others.
Conversational mode — apply to the assembled leaf from Phase 3b. Open with the cross-branch graft check before running the rest of the battery: aim Q15 (premature centralization), Q19 (future idea this makes hard), and Q20 (appears-easy-but-isn't) at the leaf with explicit attention to whether the leaf's main risk would be reduced by grafting structure from a non-chosen axis branch. If yes, surface the hybrid candidate with substrate evidence, offer backtrack to the divergence axis, and re-enter Phase 3b's axis dialog there if the user accepts. Then run the remaining battery categories on the leaf as a whole.
The battery covers eight categories:
Spec impact — what docs change/become obsolete/get renamed
Invariant impact — what gets stronger/weaker/newly required, what enforcement story
Test impact — what becomes misleading/inadequate, what new test is needed
Locality and abstraction — required context, premature centralization, seams
Future fit — easy/hard/appears-easy-but-isn't
Gotcha and scar surface — rediscovered/retired/new failure modes
Enforcement edges — invalid changes still easy, semantic linters needed
Refuse to recommend until every applicable question has a concrete answer (or "N/A" with justification). "I don't know" must be turned into "we should find out by [doing X]" before the recommendation.
Phase 5: Recommend
Recommend exactly one option, or a named hybrid. The recommendation surfaces:
The main risk of the chosen option in one sentence
The structural mitigation for that risk (not "we'll be careful")
A list of substrate that must exist before implementation (specs, matrices, invariants, tests, linters) — in the persisted brainstorm file, agent-facing; not in the chat trailer
Whether the recommendation is ready for rewrite-specs or needs another brainstorm round
The agent-facing substrate list is what rewrite-specs reads as input — it's load-bearing for the next chain step. It belongs in the persisted brainstorm file (substrate-shape vocabulary the agent uses to do the rewrite). The chat trailer renders the user-facing ## Direction block — Direction + Main risk + Structural mitigation — which is what the user reads to decide whether to approve. The two surfaces carry the same recommendation in different shapes.
Output format
The skill renders the centralized chat trailer per ${CLAUDE_PLUGIN_ROOT}/references/templates/chat-trailer.md — without the verdict slot (brainstorm is not verdict-led; per the §"Variants" brainstorm-design row, the body block carries ## Direction instead). When the brainstorm persists, the persisted file at docs/cohesive/brainstorms/YYYY-MM-DD-<slug>.md carries the full substrate-shape body — agent-facing.
Mode-aware rendering. The trailer renders in one turn at Phase 5 — same shape in both modes (Direction + Main risk + Structural mitigation, optionally preceded by a Pressure test summary table when ≥3 options were considered). In autonomous mode it follows Phase 4's option pressure-testing directly. In conversational mode it follows Phase 3b's dialog (axes → leaf summary → sub-decisions) and Phase 4's leaf battery (cross-branch graft check + remaining categories); the Phase 3b dialog turns are short verdict-led exchanges (one axis pick + counter-pressure + forced choice), not mini-trailers.
Body block specification (chat trailer). Per the §"Variants" brainstorm-design row of the centralized template: ## Direction block — **Direction:** + **Main risk:** + **Structural mitigation:** <test/type/constraint/linter — not "we'll be careful">. Optionally a ## Pressure test summary table renders above when ≥3 options were considered.
### Next block (chat trailer). One entry, decision-shape leading, skill citation parenthetical, payload following:
Ready for spec rewrite: Rewrite the docs to make this direction true. (cohesive:rewrite-specs.)Files to edit: <enumerate the docs/matrices/invariants the chosen direction touches, with the specific change in each>. Slug: <derived-from-topic>. Classification: <Pure implementation / Design / Mixed>.
Needs another round: Refine the design before writing it down. (cohesive:brainstorm-design.)Design question: <name the unresolved question — e.g., "should X be one concept or two given the future pressure of Y">.
Persisted file shape (substrate-shape, agent-facing). When the recommendation is accepted, the persisted brainstorm file uses the full substrate-shape body below — what rewrite-specs consumes:
# Brainstorm — <topic>### Current scope- ...
### Future pressure (not current scope)- ...
### Non-goals- ...
## Design options### Option A: <namedidea>**Summary:** <1–2 sentences>
**Substrate changes required:**<docs / matrices / invariants / tests / linters>**Locality impact:**<requiredcontextdelta>**Future fit:**<whichfuturepressurethisserves; whichitdoesn't>**Initial risks:**<beforepressure-testing>### Option B: <namedidea>
... (same structure)
(Option C, D as needed)
## Pressure test summary
| Option | Cohesion | Substrate delta | Future fit | Locality | Main risk |
|---|---:|---|---|---|---|
| A | High/Med/Low | Small/Med/Large | ... | ... | <onesentence> |
| B | ... | ... | ... | ... | ... |
## Breakage analysis
For each option (or just the recommended one if the others are clearly out):
### Option <X>-**Docs that would change:** ...
-**Existing assumptions that break:** ...
-**Behavior matrix impact:** ...
-**Invariant impact:** ...
-**Test guarantee impact:** ...
-**Gotchas triggered:** ...
-**Locality / centralization concerns:** ...
-**Easy invalid change still possible:** ...
## Recommendation**Direction:**<Optionnameornamedhybrid>**Main risk:**<onesentence>**Structural mitigation:**<test / type / constraint / linter / runtimewrapper / CIcheck — not "we'llbecareful">**Required substrate before implementation:**- Specs: ...
- Matrices: ...
- Named invariants: ...
- Tests / checks: ...
- Gotchas: ...
- Semantic linters (proposed): ...
### Next<decision-shapedsentence>. *(`cohesive:rewrite-specs`.)***Files to edit:**<enumerated>. Slug: `<derived-from-topic>`.
The persisted file's ## Recommendation block carries the substrate-shape "Required substrate before implementation" list — agent-facing, consumed by rewrite-specs as input. The chat trailer renders only the ## Direction decision-shape block; the substrate list does not appear in chat.
Persistence
When the user accepts a recommendation (or after Phase 4 if the chain proceeds to rewrite-specs), persist the brainstorm output to:
docs/cohesive/brainstorms/YYYY-MM-DD-<slug>.md
The persisted file is agent-facing and uses the substrate-shape body in §"Output format". rewrite-specs consumes it as input, closing the soft-prereqs hand-off gap that previously made brainstorm-design → rewrite-specs rely on human memory.
Conversational-mode addition: Decision dialog record. When Phase 2 selected conversational mode, the persisted file adds a ## Decision dialog section between## Design options and ## Pressure test summary. The section records what was discussed — not as transcript, but as a structured residue:
This section is agent-facing (consumed by rewrite-specs to understand which sub-decisions are user-load-bearing vs Default-applied). It is omitted in autonomous-mode files.
If the user declines persistence (one-shot brainstorm, no rewrite intended), the skill is conversation-only.
Acceptance criteria
At least two credible options for non-trivial changes.
Future ideas are captured under "Future pressure," not promoted to current scope.
Every option has a substrate-changes line (not just a code-changes line).
The recommendation states a main risk and a structural mitigation.
The recommendation lists required substrate by category.
Exactly one next-skill recommendation.
When the recommendation is accepted, the brainstorm output is persisted to docs/cohesive/brainstorms/YYYY-MM-DD-<slug>.md.
Mode is selected explicitly per Phase 2's auto-detect gate (≥2 axes / ≥2 substrate kinds → conversational). User overrides ("walk me through it" / "give me the autonomous version") are honored at any phase boundary.
In conversational mode: the axes map is the first user-visible chat turn (auditable from the persisted file's ## Decision dialog §"Axes walked"); each turn asks at most one forced-choice question; sub-decisions carry explicit Pick / Confirm / Default tags; the cross-branch graft check opens Phase 4 before the trailer renders.
In conversational mode: the persisted file carries a ## Decision dialog section recording axes walked, sub-decision outcomes, and the cross-branch graft check result.
Red flags
Producing code (snippets, function names, file paths to create). This skill is design-only.
"Option A is clearly best" without showing the comparison table or pressure-testing alternatives.
Recommending an option whose main risk has no structural mitigation.
Treating future pressure as current scope ("since we'll need X eventually, let's add it now").
Claiming "no behavior matrix needed" without checking whether the change introduces branchy behavior.
Skipping pressure-test questions because "they don't apply" without saying why they don't apply.
Conversational mode drifting into "what do you want?" / "tell me more" / open-ended hedging — voice stays verdict-led per ${CLAUDE_PLUGIN_ROOT}/references/output-voice.md. The user's role is forced-choice ratification or substantive redirect, not authoring.
Tagging sub-decisions as Confirm by default to look thorough. Default posture is aggressive Default; a sub-decision is material only if different choices land on different specs / invariants / tests / linters / matrices.
Reusing gate-vocabulary tokens (Decide / Lock / Build) as sub-decision tags. Sub-decision tags are Pick / Confirm / Default.
Skipping Phase 4's cross-branch graft check in conversational mode because the dialog "felt right." Q15/Q19/Q20 against the leaf is what catches premature commitment at the top-level axis.
Treating axis-pick disagreement as something to silently resolve. If a user redirected on an axis, that divergence is recorded for Phase 4's graft check and persisted in ## Decision dialog.
What this skill is not
Not an implementation planner. After validate-rewrite Approved, cohesive:implement-cohesively drives implementation against the spec diff (single-pass composition with superpowers:writing-plans and superpowers:executing-plans plus end-of-run reviewer dispatch); for non-substrate-shaped implementation, superpowers:writing-plans is also available directly.
Not an architecture review. That's cohesive:review-codebase.
Not a substrate audit. That's cohesive:audit-substrate.
Composition
Internally dispatches:cohesive:discover-substrate as Step 0 per Hard constraint #2; optional override skips re-running discovery if a report path is supplied in the dispatch prompt.
Modes: autonomous (one-turn trailer) or conversational (multi-turn axis dialog) — selected by the auto-detect gate in Phase 2; user override available either direction at any phase boundary
Often followed by:rewrite-specs (if direction is approved and large enough to justify a spec rewrite). When rewrite-specs runs, the chain continues validate-rewrite → implement-cohesively. If the change is small and substrate is already in good shape, the user may go directly to superpowers:writing-plans without the rewrite chain.
