Spec Brainstorm

Conversational design workshop for substantial work that produces a focused, reviewed spec.

One question at a time. Multiple approaches explored. Design approved in sections. Ruthless scope control. No implementation until design is approved.

Run this skill only after atelier-orchestrator selects a Spec-backed Plan or the human explicitly requests a spec. Bounded work should go directly to spec-plan for an Inline Plan. Do not reclassify the planning mode here.

Artifact

docs/specs/YYYY-MM-DD-<feature>/
└── design.md  ← This skill's output

Requirements are inline — no separate requirements.json needed.

Exclusive output contract

This skill may create or update only the design.md shown above. It may inspect the repository and discuss drafts in conversation, but it must not modify any other file, create plan.json, create tracker entries, invoke another workflow skill, or write implementation code. This boundary still applies when the human asks to brainstorm, plan, and implement in one request. Finish design.md, report it, and stop.

Lessons

These principles apply to every spec, every time.

Specs must earn their cost

Persisted specs are for work whose discovery, architecture, dependencies, or coordination needs justify a durable artifact. Do not pull bounded work into this workflow merely because it touches multiple files or takes time. That work belongs in an Inline Plan.

Design for isolation and clarity

Break the system into units with one clear purpose each. Well-defined interfaces between them. Each unit independently understandable and independently testable. If you can't explain a unit's job in one sentence, it's doing too much. Apply oracle-codebase-design: deep modules, seams placed where behaviour varies, and the interface as the test surface.

Working in existing codebases

Explore the current structure first. Follow existing patterns. Targeted improvements only. No unrelated refactoring. Understand why things are the way they are before proposing changes. Treat loaded skills as relevant guidance, not as a requirement to apply every pattern they contain.

Decomposition

If the request describes multiple independent subsystems, flag it immediately. Decompose into sub-projects before diving into details. Each substantial sub-project gets its own spec and plan; bounded sub-projects can use Inline Plans. A spec that tries to cover three subsystems helps no one.

Keep migrations separate from authorization, product behavior, infrastructure, schema, and test-platform projects. Do not use a migration as permission to redesign adjacent systems.

YAGNI ruthlessly

Remove unnecessary features from all designs. If a capability isn't needed for the first user story, it doesn't go in the spec. Every feature is a cost — to build, to test, to maintain, to understand later. Push back on scope creep during discovery.

Future consumers do not justify shared infrastructure. A "reusable foundation" may describe an architectural quality, but it is not a user story or a current requirement.

Step 1: Orient

Before diving in, understand where you are.

1. Read project context — AGENTS.md, README, existing architecture docs
2. Check existing specs — Scan docs/specs/ for previous work. What domain model

exists? What patterns are established? What has been built before?

1. Read recent specs — What was the last thing built? Is this feature building

on existing work, extending it, or something greenfield?

This is silent — don't narrate it. Let the context inform where you focus.

Step 2: Discovery

Ask questions to understand what to build. Skip this step if requirements are already clear from context (existing specs, human provided details, etc.).

Interview style

Ask one question at a time. Multiple choice preferred when possible — give 2-4 concrete options rather than open-ended prompts. Keep the conversation moving.

Good: "Should this be real-time or batch-processed? (a) Real-time via WebSocket, (b) Periodic polling every 30s, (c) On-demand when user requests it."

Bad: "How should the data synchronization work?"

When to skip discovery

• Human provides clear, detailed requirements
• Feature extends existing work with well-defined scope
• Human says "spec out X" and X is specific enough

Decomposition check

Before asking any detail questions, assess scope. If the request describes multiple independent subsystems (e.g., "build a notification system with email, SMS, push, and an admin dashboard"):

1. Flag it immediately: "This looks like multiple independent projects. Let me

propose a decomposition."

1. Break it down: Identify the subsystems and their dependencies.
2. Get agreement: "Which of these should we spec first?"

Do not try to spec everything in one document.

YAGNI check

During discovery, push back on scope:

• "Do you need this in the first version, or is it a nice-to-have?"
• "Can we ship without this and add it later if needed?"
• "This feature adds significant complexity — is the use case real or hypothetical?"

If the human insists, include it — but flag the trade-off in the spec.

Discovery questions

Adapt these to context. Not all are needed every time.

1. What problem are we solving? — Concrete problem statement, not solution description
2. Who has this problem? — User roles
3. How do they solve it today? — Current workflow and pain points
4. What does success look like? — Measurable outcomes
5. What does this integrate with? — Existing systems, APIs
6. What constraints exist? — Technical, business, regulatory

If you already know answers from orientation, confirm rather than ask.

Tell the human: "Based on my research, here's my understanding of what we're building. Does this look right?"

STOP. Wait for human confirmation.

Step 3: Research

Read the relevant codebase deeply. Not signatures — implementations, edge cases, error handling, data flows. Trace callers and callees. Read tests to understand expected behaviour.

When research spans more than a few files, delegate it to a read-only subagent per area (see Delegation in atelier-orchestrator) and assemble the research table from their cited findings.

Collect research findings for the spec as the foundation. Do not write design.md until the approved design sections are assembled in Step 4c.

The research section must map each relevant concern to the code that already handles it and the current requirement that drives the decision:

Use only reuse, modify, delete, or new in the Decision column. Every new concept must map to a present requirement; a future or hypothetical consumer does not qualify.

Tell the human: "I've written the research section of the spec. Ready for you to review before I continue with the design."

STOP. Wait for human review.

Step 4: Design

Design happens in three phases: explore approaches, present the design in sections, then write the spec file.

4a. Explore approaches

Before settling on a design, present 2-3 approaches with trade-offs.

The first approach must keep the existing architecture and make the smallest correct change. Present broader approaches only when a current requirement makes their extra cost relevant. When approaches differ mainly in interface shape, use the design-it-twice reference in oracle-codebase-design.

For each approach, address:

1. What it looks like — Brief architecture sketch
2. Pros — What makes this approach good
3. Cons — What's painful, expensive, or risky
4. Complexity estimate — Rough sense of implementation effort

Lead with your recommended option and explain why it wins.

Before asking the human to approve an approach or design batch, remove anything justified only by completeness, consistency, or hypothetical reuse.

Example:

Approach A: Single table with JSON columns - Simple schema, fast to implement - Querying inside JSON is limited, migration pain later - Complexity: Low

>

Approach B: Normalized relational tables - Clean queries, easy to evolve schema - More joins, more migration files, more code - Complexity: Medium

>

Recommendation: Approach B — the query flexibility matters more here than implementation speed.

Get explicit approval on the chosen approach before presenting the design.

Tell the human: "Which approach should we go with? Or should I explore a different direction?"

STOP. Wait for human to choose an approach.

4b. Present the design in sections

Present the design in batches. Get approval after each batch before continuing.

Sections already confirmed in earlier steps (Problem, Scope, Constraints, Context) are written into the spec from those confirmations — do not re-present them.

Batch A: User Stories — the contract you're designing against. Formal stories with acceptance criteria and priorities. If rejected: revise. If the rejection reveals a scope misunderstanding, loop back to Discovery (Step 2).

Batch B: Architecture — component design, domain modeling, and layer boundaries. Use oracle-codebase-design and any installed language-specific architecture or API-design skills relevant to your stack. Then present: component structure, domain model, where business logic lives, where IO lives, where untrusted input is parsed (the external seam), and the seam each module is tested through. If rejected: revise. If the rejection undermines the chosen approach, offer to return to approach exploration (4a). If it reveals a fundamental gap, loop back to Research (Step 3). If the detail reveals the work is far more complex than estimated, say so and offer to revisit the approach.

Batch C: API Design + Data Model — contracts derived from the approved architecture. Skip sections that don't apply, but say so explicitly ("No API changes — moving to Trade-offs"). Never skip silently. If rejected: revise. If the rejection implicates the architecture, go back to Batch B.

Batch D: Trade-offs + Open Questions — alternatives considered, why this approach wins, known limitations, anything unresolved. Usually revisable inline.

Each batch ends with:

Tell the human: "Does this look right?"

STOP. Wait for approval before continuing.

If you loop twice on the same batch, stop and ask:

"We've looped on [batch] twice. Should we reconsider the approach?"

Terminology discipline: while drafting batches, challenge terms against CONTEXT.md and record resolved terminology in design.md. Do not update CONTEXT.md from this skill. If domain confusion runs deep, suggest pausing for oracle-grill-me before continuing.

4c. Write the spec

Once all batches are approved, write the full spec document.

#### What the spec should contain

# Feature Name

## Problem
- What problem are we solving
- Who has this problem
- How they solve it today

## Scope
- **In scope:** [specific capabilities]
- **Out of scope:** [explicitly deferred]

## User Stories
- US-1: As a [role], I want [action], so that [benefit]
  - Given X, when Y, then Z
- Priority: must/should/could

## Constraints
- [Technical or business constraints]

## Context
- What exists today, how it works end-to-end
- Existing patterns and conventions
- Dependencies and integration points
- Gotchas, assumptions, technical debt

## Architecture
- Component structure (functional core / effectful edge)
- Domain model: entities, value objects, aggregates
- Where business logic lives, where IO lives

## API Design
- Endpoints, request/response contracts
- Error handling approach
- Event contracts (published/consumed)

## Data Model
- Schema design, access patterns
- Migrations needed

## Trade-offs
- Alternatives considered
- Why this approach wins
- Known limitations

## Open Questions
- Anything unresolved needing human input

Scale each section to complexity — a few sentences if straightforward, detailed if nuanced.

#### Reference implementations

If the human provides reference code — from open source, from elsewhere in the codebase — use it as a concrete guide. Working from a reference produces dramatically better designs.

#### Self-review

After writing the file, check it with fresh eyes:

1. Placeholder scan — Any TBD, TODO, FIXME, or incomplete sections? Every

section should have real content or be removed.

1. Internal consistency — Do sections contradict each other? If the

Architecture section says "stateless" but the Data Model includes session state, resolve the conflict.

1. Scope check — Is this focused enough for a single implementation plan?

If the spec covers more than one independent subsystem, it should have been decomposed in Step 2. If it's still too broad, flag it now.

1. Ambiguity check — Could any requirement be interpreted two ways? If so,

pick one interpretation, state it explicitly, and let the human correct you.

1. Necessity check — Delete concepts justified only by completeness, consistency, or

hypothetical reuse. Confirm every remaining new concept maps to a current requirement.

Substance rule: if a fix changes the substance of an approved section, re-present that section for approval. Wording and consistency fixes go inline — note them at handoff.

Step 5: Handoff

Tell the human:

"Brainstorm complete. Design written to docs/specs/<path>/design.md. A separate spec-plan invocation can create plan.json after you approve this document."

If the human requests changes — in conversation or by annotating the file — address every note, update the spec, and re-run the self-review. If a change alters the substance of an approved section, re-present that section for approval before continuing. Resolve questions that affect scope, architecture, contracts, data, security, or task ordering before handoff.

Stop after reporting the completed design.md. Do not invoke spec-plan, offer to continue automatically, create plan.json, or write code. The human must start the next phase with a separate request.

If planning reveals design flaws, loop back to research. See atelier-orchestrator for iteration patterns.