---
name: brainstorming
description: "Use when defining ambiguous or high-complexity new features, product behavior, UI/component design, architecture choices, contract changes, or when grilling/pressure-testing a plan or design. Routine small requests stay on the fast path."
---

<EXPLICIT-MODE-GATE>
If activation mode is explicit (`~/.config/aegis/config.toml` has
`activation_mode = "explicit"`, or `AEGIS_ACTIVATION_MODE=explicit` is visible
in the environment) and the current user request did not explicitly invoke
Aegis or this skill by name, exit back to the fast path: answer concisely
without this workflow's checklist, ceremony, or document requirements. If the
user explicitly named Aegis or this skill, proceed normally.
</EXPLICIT-MODE-GATE>

# Execute

→ Direct grilling or plan/design pressure-test? → Enter `Grilling Mode` below. Soft challenge intent? → Use its one-line mode confirmation. Do not start normal design artifacts, document writing, task planning, or implementation during the interview.
→ New feature, product behavior, UI/component design, architecture/contract change, or ambiguous medium/high-complexity work? → **Design first. No implementation until the needed design/spec is approved.**
  1. Explore project context → read authority docs, check for existing patterns
  2. Ask clarifying questions one at a time (prefer multiple choice)
  3. Propose 2-3 approaches with trade-offs and your recommendation
  4. Present design sections → get user approval after each
  5. Write spec → self-review → user review → transition to writing-plans
→ HARD GATE: For tasks that match this skill, do NOT write code, scaffold projects, or invoke implementation skills until design/spec approval is satisfied.

## Route Fixtures

These rows are calibration expectations for method behavior, not a runtime
regex router. The Agent selects the route from evidence; route selection is
not a user question.

| Scenario | Route |
| --- | --- |
| 想法还没想清楚，先梳理功能设计 | normal brainstorming, compact output first |
| 讨论公共 API 契约和兼容边界 | normal brainstorming, design sections before implementation |
| 盘问/拷问/审问这个方案，不要顺着我 | `Grilling Mode` |
| 修复登录按钮的空指针 | `systematic-debugging` |
| review 当前 PR / diff / 当前代码 | `requesting-code-review` |
| 给我一个有目标的方案 | `goal-framing` when goal intent is explicit; otherwise normal brainstorming |
| 把按钮文案从保存改成提交 | fast-path; no design ceremony |

# Brainstorming Ideas Into Designs

Help turn ideas into fully formed designs and specs through natural collaborative dialogue.

Start by understanding the current project context and authority boundary, then ask questions one at a time to refine the idea. Once you understand what you're building, present the smallest design artifact that stabilizes the work and get the required approval.

<HARD-GATE>
Do NOT invoke any implementation skill, write any code, scaffold any project, or take implementation action for work that matches this skill until you have presented the required design/spec and the user has approved it where this workflow requires approval.
</HARD-GATE>

## Grilling Mode

### Mode Precedence

While Grilling Mode is active, it overrides the normal brainstorming execution
flow. Suspend `Checklist`, `The Process`, the `Compact output contract`, and
all documentation or design-transition requirements until the user exits the
interview; retain the no-implementation hard gates.

### Grilling Entry Signals

- **Direct:** The user asks to grill or interrogate an idea, plan, or design, or explicitly requests a pressure-test. Direct phrases include `grill me`, `grill this plan`, `审问我`, `盘问我`, and `拷问我`. Enter the mode immediately.
- **Soft:** The user asks to challenge assumptions, find holes, red-team, or "别顺着我" about a draft idea, plan, or design. Ask only: `Grill or normal brainstorming?` Enter the mode only after confirmation.
- **Not grilling:** A bare/literal phrase reference, or a PR, diff, or current-code review. Explain literal phrases normally; route implementation review to `requesting-code-review`.

### Opening Card

After the user has entered the mode, emit this once in the user's language, then begin the interview:

```text
◆ Grilling Session
Target: <idea / plan / design>
Question path: value -> boundaries -> failure modes -> acceptance
Pace: deep (default) | fast (user-requested)
```

### Pace

- **Deep:** Ask exactly one decision question per turn when it is blocking or depends on the previous answer. State the recommended answer and the most relevant trade-off before waiting.
- **Fast:** Only when the user explicitly asks for a fast or batched interview (for example, `fast`, `batch`, `快问`, or `一次问几个`), ask at most three independent decision questions. Give each question its recommendation and trade-off, then wait for the user's responses. Return to deep pace for dependent follow-ups.

1. Explore the codebase and current authority docs for facts before asking. Do not ask the user for facts that can be found locally.
2. The user owns the decision. Do not treat a recommendation, a tentative answer, or a shared-understanding checkpoint as final approval.
3. Aside from the one-time opening card, keep the turn to the observation, recommendation, and the selected pace's questions. Do not emit a full design ceremony, write docs, create a plan, or implement while the interview is active.
4. End when the user says to stop, defer, or that the questions are sufficient. Reconfirm in a structured `Challenge Result`. That summary does not grant completion authority.

```text
Challenge Result
- Survived assumptions
- Rejected assumptions
- New evidence needed
- Design changes required
- Residual risks
- Return state: interview | design | approaches | writing-plans
```
5. If the user asks to proceed after the interview, return to the normal brainstorming design gate. A design/spec still needs the required approval before planning or implementation.

## Route Away / Doc Necessity Gate

Do not force this workflow onto low-complexity work. A tiny
wording edit, single-owner bug fix, simple config/status question, local
utility change, or mechanical multi-file change can proceed through concise
intent, baseline check, TDD/debugging, and verification without any new
document. Run the Doc Necessity Gate before writing any spec, plan, ADR, or
baseline artifact:

1. Does an existing spec/plan/ADR/baseline already cover this change surface?
   -> Update that owner document in place; never create a sibling document.
2. Is the surface durable/irreversible (schema, public API, owner, dependency
   direction, migration, compat-path retirement), cross-session/cross-person
   handoff, approval-gated, or authority-required?
   -> Yes: write the smallest artifact for that surface (see Documentation).
   -> No: write no document; keep compact drafts in-session.
3. Re-check at edit time and at closeout; escalate to the smallest stabilizing
   spec if uncertainty or impact grows.

### Route Precedence

1. Route-away cases leave this workflow first (`systematic-debugging`,
   `requesting-code-review`, `goal-framing` when goal intent is explicit,
   fast-path micro-tasks).
2. `Grilling Mode` requires explicit challenge intent. Ordinary discussion,
   evaluation, or the need to clarify understanding is not grilling.
3. Otherwise run the normal brainstorming flow and escalate depth on evidence:
   contracts, owners, persistence, migration, security, consumer count, or
   blast radius. Escalate after evidence, not merely because the first
   request sounds ambitious.
4. File count alone is not a design signal: a mechanical multi-file change
   can still be fast-path, and a one-file contract change can still be full
   design.

## Role And Authority Contract

### Agent-owned decisions

Resolve these directly without asking the user:

- repository investigation strategy and evidence gathering order
- file and function organization inside an accepted owner
- testing commands and proportional verification mechanics
- inline versus subagent execution when policy already allows it
- reversible implementation structure that does not change product behavior,
  contract, authority, or durable boundaries

### User-owned decisions

Ask the user only for:

- product behavior or preference
- irreversible, destructive, external, public, production, or sensitive impact
- explicit product/contract commitments only the user can make
- necessary information unavailable from repository, tools, and authority docs

Every user question must pass this test:

> If the user chooses another answer, which design boundary, behavior, owner,
> acceptance criterion, or risk decision changes?

If none changes, do not ask. When a question passes this test,
attach a recommended option and the reason for it, so the user decides
between framed choices instead of researching. This classification clarifies which decisions
are user-owned; it does not remove the approval points this workflow already
defines.

## Checklist

You MUST create a task for each of these items and complete them in order:

1. **Explore project context** — check files, docs, recent commits, authority
   docs, and passively consume relevant active `CONTEXT.md` language without
   loading active modeling
2. **Choose the path and scope** — real design? diagnosis? route accordingly or decompose first
3. **Ask clarifying questions** — one at a time, understand purpose/constraints/success criteria
4. **Draft working artifacts** — `TaskIntentDraft`, `BaselineReadSetHint`, `BaselineUsageDraft`, `ImpactStatementDraft`
5. **Run existence check when adding new surfaces** — only if an approach adds a new owner, skill, artifact, adapter, fallback, workflow step, or benchmark metric
6. **Propose 2-3 approaches** — with trade-offs and your recommendation
7. **Present design** — in sections scaled to complexity, get user approval where required
8. **Write spec artifact** — only after the Doc Necessity Gate passes and no existing owner spec/plan covers the surface; if covered, update that document instead of creating a sibling
9. **Spec self-review** — check for placeholders, contradictions, ambiguity, scope, boundary
10. **User reviews written spec** — ask user to review before proceeding
11. **Transition to implementation** — invoke writing-plans skill (terminal state)

**The terminal state is invoking writing-plans.** Do NOT invoke any other implementation skill.

## The Process

**Understanding the idea:**
- Check current project state first (files, docs, recent commits)
- Read relevant authority docs before asking deep questions
- Use existing canonical terms in questions, options, scenarios, and the spec.
  If terminology crystallizes or conflicts, compose
  `establishing-project-context`; do not leave the resolution only in the spec.
- If the request is diagnosis/root-cause/follow-up to an approved plan → route to correct workflow
- If the request spans multiple independent subsystems → flag and decompose first
- Ask clarifying questions one at a time, prefer multiple choice
- Separate facts, assumptions, unknowns while exploring

**Working artifacts:** Keep four drafts: `TaskIntentDraft` (outcome, goal,
success evidence, stop condition, non-goals, scope, risks),
`BaselineReadSetHint` (candidate docs, authority gaps),
`BaselineUsageDraft` (required refs, optionally delivered context refs,
acknowledged-before-plan refs, cited refs, missing refs, advisory decision),
and `ImpactStatementDraft` (affected layers, owners, invariants, compat,
non-goals). Refresh when scope changes.

**Compact output contract:** `Aegis Visibility`, `TaskIntentDraft`, `BaselineReadSetHint`,
`BaselineUsageDraft`, `Requirement Ready Check`, `ImpactStatementDraft`,
`Existence Check`, `Product Risk Lens`, `Architecture Integrity Lens`,
`Prior-Art & Reuse Lens`, `Baseline Role Alignment`, `Plan-Time Complexity
Check`, `Options`, and `Decision Needed`. Use this compact shape before expanding into a full design
structure.

`Aegis Visibility` for this workflow names why design/spec clarification comes
before implementation and what drift, overbuild, wrong-owner, or missing
acceptance risk that restraint reduces. Keep it natural and task-specific; do
not turn it into a fixed skill trace.

Use a compact `BaselineUsageDraft` whenever the design direction depends on
specific baseline docs or current-authority refs:

```text
BaselineUsageDraft:
- Required baseline refs:
- Delivered context refs:
- Acknowledged before plan refs:
- Cited in design refs:
- Missing refs:
- Decision: continue | needs-baseline-readback | needs-verification | pause-for-user | blocked
```

`Delivered context refs` is optional host-projected bookkeeping only. It is not
authoritative proof that a host injected or the model internally consumed a
context payload. The artifact exists to make baseline/context attention drift
visible before the design is recommended or approved.

Use a compact `Requirement Ready Check` before recommending a design when the
requirement is not already confirmed and complete:

```text
Requirement Ready Check:
- Requirement source refs:
- Goals and scope refs:
- User / scenario refs:
- Requirement item refs:
- Acceptance / verification criteria refs:
- Open blocker questions:
- Decision: ready | needs-source | needs-goal-alignment | needs-scenario | needs-acceptance-criteria | needs-clarification | needs-user-decision | blocked
```

Treat task intent, conversation, source documents, and agent inference as
candidate requirement sources until project authority confirms them. If the
decision is not `ready`, keep the design at proposal/spec clarification level;
do not turn the gap into implementation tasks.

**Existence Check:** Before recommending an approach that adds a new owner,
skill, artifact, host adapter, fallback, compatibility path, workflow step, or
benchmark metric, check whether it needs to exist. Use
`docs/current/AEGIS_MINIMALITY_REFERENCE.md` as the reference. Do not force this
onto ordinary feature design that reuses existing owners and artifacts.

```text
Existence Check:
- Proposed new surface:
- Existing owner / reuse candidate:
- Why existing surface is insufficient:
- Creation proof:
- Entropy / retirement impact:
- Decision: reuse-existing | add-with-proof | defer | reject | needs-first-principles-review
```

If the decision is `reuse-existing`, recommend the reuse path instead of a new
surface. If the decision is `add-with-proof`, carry the proof, verification
signal, and any retirement trigger into the design/spec.

**Product Risk Lens:** For ambiguous product, feature, UI, workflow, or
architecture choices, add a compact review lens, not persona roleplay:

```text
Product Risk Lens:
- Value:
- Non-goals:
- Trade-offs:
- Decision needed:
```

This is a review lens, not persona output. It does not override baseline evidence,
approved requirements, or current authority docs; it only makes the product risk
and decision point visible before implementation.

**Plan-Time Complexity Check:** Before choosing an implementation direction for
medium/high work, inspect the likely owner files and current shape. This is an
advisory design pressure check, not a gate and not completion authority. Do not
force it onto tiny low-risk edits.

Use `using-aegis/references/complexity-governance.md` for the shared artifact
classes, pressure-signal interpretation, and over-budget handling.

```text
Complexity Budget:
- Artifact class:
- Target files / artifacts:
- Current pressure:
- Projected post-change pressure:
- Budget result: within-budget | at-risk | over-budget
- Planned governance:

Plan-Time Complexity Check:
- Better file boundary:
- Recommendation: edit-in-place | extract helper | add owner file | split task | defer refactor
```

**Exploring approaches:** Propose 2-3 approaches with trade-offs and
recommendation. Make scope boundary explicit: what's in, what's deferred, what
belongs elsewhere.

Before approach selection, use `Existence Check` for any proposed new surface.
Escalate to `first-principles-review` and its `Decision Hygiene Review` when
the candidate direction still introduces a new owner, duplicate owner,
fallback, adapter, compat-only carrier, delete-first question, unverified
assumption, or "long-term stable" claim after the existence check. Do not make
either check a universal design ceremony; return to this workflow once the
decision surface is clean.

When the central decision is internal retirement vs compat retention vs
persistent-state confirmation, compose `anti-entropy-governance`. It classifies
the deletion target, chooses `delete-first | compat-exception |
confirmation-first`, and keeps destructive authority outside the design skill.

Use the narrower `Architecture Integrity Lens` when the main risk is not broad
strategy but architecture coherence: unclear canonical owner, responsibility
overlap, caller-side fallback, stale path carrying real logic, or a possible
higher-level owner / contract / source-of-truth simplification. The lens should
answer invariant, canonical owner / contract, responsibility overlap,
higher-level simplification, retirement / falsifier, and verdict before the
approach is recommended.

**Prior-Art & Reuse Lens:** When a candidate approach would introduce a new
mechanism, protocol, artifact shape, or nontrivial interaction pattern, check
proven external practice before inventing one. This lens is behavior-triggered:
research precedents when the direction is novel for the project, plausible
approaches remain after internal reuse checks, or the domain sits outside
current repository evidence. Do not run research ceremony for routine work that
already maps to well-known framework patterns, and do not let an unavailable
web/search tool stall approach selection.

```text
Prior-Art & Reuse Lens:
- Searched precedents: <bounded sources; index-first summary; cite anchor per claim>
- Adopt verbatim: <proven pattern + source>
- Adapt with reason: <tailored part -> project constraint / non-negotiable it maps to>
- Reject with reason: <project fact that makes the pattern inapplicable>
- Degraded: <no web/search tooling -> external basis unknown; internal-only evidence stated>
```

Search results are evidence candidates, not prompt payload: summarize
index-first and cite anchors instead of pasting raw pages. An "industry
standard" claim without a citable anchor stays `unknown`. Every adapt/reject
decision binds to a named project constraint or fact, not taste. The lens feeds
only the approach recommendation; it stays advisory and grants no completion
authority.

**Baseline Role Alignment:** When a question may involve both "what should be
built" and "where it should live", keep requirement truth separate from
architecture truth:

```text
Baseline Role Alignment:
- Product / Requirement Baseline:
- Architecture / Runtime Boundary Baseline:
- Result: aligned | Design Defect | Implementation Drift | missing-authority | needs-clarification
- scope: requirements | architecture | both
- Next action:
```

Use `Design Defect` when the relevant requirement, design, or baseline is wrong.
Use `Implementation Drift` when the work deviates from a correct unchanged
baseline. `Architecture Defect` and `Architecture Drift` remain compatibility
aliases for architecture-scoped `Design Defect` and architecture-scoped
`Implementation Drift`. This is a review lens, not a runtime gate or completion
authority.

**Presenting the design:** Scale sections to complexity. Cover only the surfaces that matter: architecture, components, data flow, error handling, testing, compatibility boundary. Get approval for the design before implementation when behavior, contract, architecture, or user-facing flow is being decided.

**ADR signals:** When the design/spec touches durable architecture surfaces
(owner, public contract, artifact shape, dependency direction,
source-of-truth, host compatibility, runtime-ready boundary, fallback,
adapter, or retirement schedule), mark the ADR signal, source refs, real
alternatives, and expected baseline-sync question for later completion. Do not
create accepted architecture memory from unexecuted ideas.

**Design for isolation:** Each unit = one clear purpose, well-defined interface, testable independently. Can someone understand it without reading internals? Can you change internals without breaking consumers?

**Existing codebases:** Follow existing patterns. Include targeted improvements only when they serve the current goal. If the design touches contracts, compat, fallbacks, or duplicated owners → call it out directly.

## Conditional Detailed Guidance

Read `expanded-design-guidance.md` completely only when current evidence needs
one of its conditional surfaces: a design probe, a scenario profile, or
workspace / spec documentation after the Doc Necessity Gate passes. Do not
load it for route selection, first-turn clarification, route-away cases, or an
active Grilling Mode interview. The reference supplies detail; this main skill
continues to own routing, design approval, and handoff.

## Design Ready And Design Complete

Approach selection is ready when:

- the desired outcome and primary scenario are known;
- scope and non-goals are explicit;
- current behavior and the target delta are grounded;
- key invariants and the likely canonical owner are identified;
- at least one observable acceptance criterion exists;
- no open unknown can still change the approach category.

Not every unknown must be eliminated; only decision-changing unknowns block
convergence.

The design can hand off when all applicable conditions hold:

- the selected approach and canonical owner are explicit;
- fixed behavior/contract and implementation-owned choices are separated;
- alternatives were materially compared or excluded by evidence;
- critical assumptions have evidence or explicit acceptance;
- failure/recovery and consumer impact are covered where applicable;
- acceptance criteria are observable and usable by verification;
- new owners, fallbacks, adapters, compatibility, or persistence have creation
  proof and retirement/rollback treatment;
- every user-owned decision has real user approval.

Design Complete is method readiness, not completion authority. Transition to
writing-plans only after these conditions hold; do not carry unresolved
decision-changing unknowns into the plan.

## After the Design

After approval, **Write the validated spec artifact when needed**. If the Doc
Necessity Gate selects workspace or spec documentation, read the applicable
sections of `expanded-design-guidance.md`; otherwise keep the accepted design
in-session. When workspace support is selected, its helper remains outside the
target project and is invoked as `<aegis-workspace-helper>`.

**Aegis Project Workspace:** workspace creation and spec persistence remain
conditional on the Doc Necessity Gate and project authority.

For a Design Spec, complete its self-review and obtain explicit user review
before planning. Apply requested changes and review again. A small Spec Brief
that only pins medium-task acceptance may use concise review unless project
authority requires a formal gate.

**Implementation:**

- Invoke the writing-plans skill to create a detailed implementation plan
- Do NOT invoke any other skill. writing-plans is the next step.

## Key Principles

- **One question at a time** - Don't overwhelm with multiple questions
- **Multiple choice preferred** - Easier to answer than open-ended when possible
- **YAGNI ruthlessly** - Remove unnecessary features from all designs
- **Explore alternatives** - Always propose 2-3 approaches before settling
- **Incremental validation** - Present design, get approval before moving on
- **Be flexible** - Go back and clarify when something doesn't make sense