codify（codify 命令行）

Codify

Identified pattern → component mapping → implementation.

<when_to_use>

• Spotting repeated behavior worth codifying
• User explicitly wants to capture a workflow
• Recognizing orchestration sequences in conversation
• Identifying decision heuristics being applied

NOT for: one-off tasks, simple questions, well-documented existing patterns

</when_to_use>

<pattern_types>

Workflows: Step-by-step processes with defined stages and transitions. Orchestration: Tool combinations that work together for a goal. Heuristics: Conditional logic and decision trees for common situations.

</pattern_types>

<component_mapping>

Match pattern type to implementation:

Is it a multi-step process with stages?
├─ Yes → Does it need tool restrictions?
│        ├─ Yes → Skill (with allowed_tools)
│        └─ No → Skill
└─ No → Is it a simple entry point?
         ├─ Yes → Command (thin wrapper → Skill)
         └─ No → Is it autonomous/long-running?
                  ├─ Yes → Agent
                  └─ No → Is it reactive to events?
                           ├─ Yes → Hook
                           └─ No → Probably doesn't need codifying

Composites:

• Skill + Command: Skill holds logic, command provides entry point
• Skill + Hook: Skill holds logic, hook triggers automatically
• Agent + Skill: Agent orchestrates, skill provides methodology

</component_mapping>

Pattern spec format (YAML):

name: pattern-name
type: workflow | orchestration | heuristic
trigger: when to apply
stages:  # workflow
  - name: stage-name
    actions: [...]
    exit_criteria: condition
tools:   # orchestration
  - tool: name
    role: purpose
    sequence: order
rules:   # heuristic
  - condition: when
    action: what
    rationale: why
quality:
  specific: true | false
  repeatable: true | false
  valuable: true | false
  documented: true | false
  scoped: true | false

All five quality checks must pass before codifying.

1. Identify: Spot repeatable behavior in conversation

- If hint/argument provided, focus analysis on that specific pattern - Otherwise scan for: workflows, orchestrations, and heuristics worth capturing - For deep analysis, load outfitter:codebase-recon skill and use outfitter:patterns techniques - Extract success, frustration, workflow, and request signals - Look for 3+ occurrences of similar behavior

1. Classify: Workflow, Orchestration, or Heuristic?
2. Map: Which component(s) should implement it?
3. Specify: Document with pattern spec format
4. Quality: Validate against SRVDS criteria
5. Implement: Create the component(s)

Task stages:

- Identify { pattern description }
- Classify { pattern type }
- Map { component decision }
- Specify { pattern name }
- Implement { component type }

SRVDS criteria — all must pass:

Skip if: < 3 occurrences, context-dependent, simpler inline

<anti_patterns>

• Premature abstraction: Codifying after first occurrence
• Over-specification: 50-line spec for 5-line pattern
• Wrong component: Hook when Skill needed, Agent when Command suffices
• Missing trigger: Pattern exists but no clear activation
• Scope creep: Pattern grows to handle edge cases

</anti_patterns>

ALWAYS:

• Identify pattern type before choosing component
• Validate all SRVDS criteria
• Start with minimal implementation
• Document trigger conditions clearly
• Test pattern in at least 2 contexts

NEVER:

• Codify after single occurrence
• Create Agent when Skill suffices
• Skip quality validation
• Implement without clear trigger
• Add "might need later" features

• pattern-types.md — extended examples by type
• component-mapping.md — decision tree details
• examples/ — captured pattern examples

Identification vs Implementation:

• patterns skill identifies and documents patterns
• This skill (codify) implements patterns as Claude Code components

Use patterns first to identify what's worth capturing. Use codify to turn identified patterns into skills, commands, hooks, or agents.

Component skills (loaded during implementation):

• claude-skills — skill authoring
• claude-commands — command authoring
• claude-hooks — hook authoring
• claude-agents — agent authoring