| name | brewcode:spec |
| description | Creates SPEC.md task spec via research + interaction. Triggers: create SPEC.md, brewcode spec, write spec. |
| disable-model-invocation | true |
| argument-hint | [-n] <description> | <path-to-requirements> — -n/--noask: no questions to user |
| allowed-tools | Read, Write, Glob, Grep, Bash, Task, AskUserQuestion |
| model | opus |
Input Handling
| Input | Action |
|---|
$ARGUMENTS empty | Read .claude/TASK.md → first line = path → derive task dir |
$ARGUMENTS has text | Use as task description |
$ARGUMENTS has path | Read file as task description |
Flag Parsing
Parse $ARGUMENTS for flags BEFORE input detection:
| Flag | Effect |
|---|
-n, --noask | Skip all user questions, auto-approve defaults |
Strip flag from $ARGUMENTS. Remaining text = description or path.
Naming
- Timestamp:
YYYYMMDD_HHMMSS (e.g., 20260208_143052)
- Name slug: lowercase, underscores, from description (e.g.,
auth_feature)
- Task dir:
.claude/tasks/{TIMESTAMP}_{NAME}_task/
Workflow
-
Detect SPEC Template (OPTIONAL)
EXECUTE using Bash tool:
test -f .claude/tasks/templates/SPEC.md.template && echo "PROJECT TEMPLATE" || echo "BUILT-IN STRUCTURE"
If a project template exists, use it. Otherwise use the built-in SPEC structure from the Output section below.
-
Read & Analyze Input
- Parse
$ARGUMENTS per Input Handling table
- Determine scope: files affected, areas of codebase
- Identify what needs clarification
-
Clarifying Questions (AskUserQuestion)
If --noask: Skip. Record in SPEC User Q&A: "Skipped (--noask mode)". Infer scope from description and codebase analysis.
Otherwise: Use AskUserQuestion tool to ask 3-5 questions, grouped in batches of up to 4 per AskUserQuestion call. Focus on:
| # | Category | Example Questions |
|---|
| 1 | Scope | What's in/out? Which modules affected? |
| 2 | Constraints | Required libraries? Backward compatibility? API contracts? |
| 3 | Edge cases / ambiguities | Concurrent access? Empty/null inputs? Error recovery? |
Record all Q&A for the User Q&A section of SPEC.
2.5. Feature Splitting Check
After gathering requirements, evaluate scope:
IF requirements cover >3 independent areas OR estimated complexity >12 plan phases:
→ AskUserQuestion: "I suggest splitting into X tasks: [A], [B], [C]. Agree?"
→ If yes: create SPEC only for first task, record others in Notes section
→ If no: continue with full scope
Dynamic Agent Resolution
Before spawning agents, check for project team agents:
- If
.claude/teams/ exists — read team.md for agent roster with domains
- If
.claude/agents/ has project agents — list available
- Match agent domain to current task area
- Priority: team agent > project agent > plugin agent > system agent
- If agent refuses (Task Acceptance Protocol) — re-delegate to suggested colleague (max 2 retries)
Always fall back to plugin agents when no project agents match the task domain.
-
Partition Research Areas (5-10 areas)
Analyze project and split into logical parts for parallel research:
| Area | Pattern | Agent |
|------|---------|-------|
| Domain-specific | matching pattern | matching team agent |
| Controllers | **/controllers/ | developer |
| Services | **/services/ | developer |
| DB/Repos | **/repositories/ | developer |
| Tests | **/test/ | tester |
| Config | *.yml, docker-* | developer |
| Docs | *.md, docs/ | Explore |
See references/SPEC-creation.md for detailed parallel research instructions.
-
Parallel Research (ONE message, 5-10 agents)
ONE message with 5-10 Task calls in PARALLEL
Task(subagent_type="Plan", prompt="Analyze architecture...")
Task(subagent_type="developer", prompt="Analyze services...")
Task(subagent_type="tester", prompt="Analyze test patterns...")
Task(subagent_type="reviewer", prompt="Analyze quality...")
Task(subagent_type="Explore", prompt="Find library docs...")
Agent prompt template:
> **Context:** BC_PLUGIN_ROOT is available in your context (injected by pre-task.mjs hook).
Analyze {AREA} for task: "{TASK_DESCRIPTION}"
Focus: patterns, reusable code, risks, constraints
Context files: {FILES_IN_AREA}
Output: findings (bullets), assets (table), risks, recommendations
NO large code blocks - use file:line references
-
Consolidate into SPEC
- Create task directory:
.claude/tasks/{TIMESTAMP}_{NAME}_task/
- Use the project
SPEC.md.template if detected in Step 0, otherwise the built-in structure
- Merge agent findings (deduplicate)
- Fill SPEC sections per Consolidation Rules in
references/SPEC-creation.md
- Write
.claude/tasks/{TIMESTAMP}_{NAME}_task/SPEC.md
- Include Research table with per-agent findings
-
Present Key Findings (AskUserQuestion)
If --noask: Skip validation. Auto-approve all findings.
Otherwise: Use AskUserQuestion to validate with user:
- Key architectural decisions made
- Risk assessment and proposed mitigations
- Any assumptions that need confirmation
- Completeness check: "Does this cover everything?"
Incorporate user feedback into SPEC.
-
Review SPEC (reviewer agent + fix loop)
Task(subagent_type="reviewer", prompt="> **Context:** BC_PLUGIN_ROOT is available in your context (injected by pre-task.mjs hook).
Review SPEC at {SPEC_PATH}
Check: completeness, consistency, feasibility, risks
Output: list of remarks with severity (critical/major/minor), specific fixes")
Iteration loop:
WHILE remarks.critical > 0 OR remarks.major > 0:
1. Fix all critical/major remarks in SPEC.md
2. Re-run reviewer
MAX 3 iterations. After 3 rounds, present remaining remarks to user via AskUserQuestion.
Exit criteria: No critical/major remarks remaining OR 3 iterations exhausted
Template source: Prefer a project .claude/tasks/templates/SPEC.md.template when present; otherwise use the built-in SPEC structure below.
Output
# Spec Created
## Detection
| Field | Value |
|-------|-------|
| Arguments | `{received args}` |
| Input Type | `{text description or file path}` |
| Noask | `{yes or no}` |
## Files Created
- SPEC: .claude/tasks/{TIMESTAMP}_{NAME}_task/SPEC.md
- Task Dir: .claude/tasks/{TIMESTAMP}_{NAME}_task/
## Next Step
> Clear context, then hand the SPEC to the developer agent for implementation.
1. Clear context: type `/clear` and press Enter
2. Delegate implementation to the `developer` agent, pointing it at the SPEC:
\`\`\`
Implement the spec at .claude/tasks/{TIMESTAMP}_{NAME}_task/SPEC.md
\`\`\`