Spec

---
name: spec
description: Create specifications through parallel analysis
disable-model-invocation: true
argument-hint: "[REQUIREMENTS]"
---

# Spec Command

Creates specifications with intelligent clarification.

## Usage
```
/sf:spec [REQUIREMENTS]
```

---

## Project Context
- Branch: !`git branch --show-current 2>/dev/null`
- Recent commits: !`git log --oneline -5 2>/dev/null`
- Working tree: !`git status --short 2>/dev/null | head -20`

## Clarification Check

If requirements are vague (< 15 words or unclear), use the **AskUserQuestion** tool to gather missing context. Do NOT output questions as plain text — always use the tool so execution pauses for the user's answer.

**Questions to consider asking:**
- What specific problem are you solving?
- Who are the users?
- What's the desired outcome?
- Any technical constraints?
- What's the minimal viable version?

## Directory Management

Claude Code will use `.claude/.sf/` as the working directory.

**Command-level logic:**

```
If .claude/.sf/spec.md exists:
    Use AskUserQuestion tool: "Existing spec found. What would you like to do?"
      Options: "Update existing" / "Create new"
    If user chooses update: Write "update" to .claude/.sf/mode
    If user chooses new: Write "new" to .claude/.sf/mode
Else:
    Write "first" to .claude/.sf/mode
```

## Execution

After directory setup and clarification (if needed), run agents:

**Pre-execution:**
- Task: manage-spec-directory (maxTurns: 3) (reads mode from $SF_DIR/mode file)
- **If manage-spec-directory fails (non-zero exit), halt immediately — do not run downstream agents.**

**Batch 1 (Parallel):**
- Task: define-scope (maxTurns: 6) with requirements: $ARGUMENTS
- Task: create-criteria (maxTurns: 6) with requirements: $ARGUMENTS
- Task: identify-risks (maxTurns: 6) with requirements: $ARGUMENTS

**Batch 2:**
- Task: synthesize-spec (maxTurns: 12) to combine all research, following the structure in `${CLAUDE_SKILL_DIR}/spec-template.md`

Output: `$SF_DIR/spec.md` (direct file or symlink to timestamped spec)

## Error Recovery

If any agent fails:
1. Claude Code shows the specific error
2. Fix the issue (usually unclear requirements)
3. Re-run /sf:spec with clearer input
4. No partial state - each run starts fresh