Codify Skill

/codify <skill-name> — Extract deterministic logic from a skill's SKILL.md into a Python script.

What This Does

Takes a skill's SKILL.md and splits it into:

• Script (<skill-dir>/script.py) — all deterministic operations (file ops, validation, parsing, template copying)
• Slim SKILL.md — only AI judgment parts (questions, assessment, continuation logic) + calls to the script

Argument

/codify <skill-name> — the skill directory name under ~/.claude/skills/. If omitted, list available skills and ask.

Workflow

Phase 1: Analyze

1. Read ~/.claude/skills/<skill-name>/SKILL.md
2. Classify every block of instructions into one of two buckets:

DETERMINISTIC (goes into script.py):

• File/directory creation, copying, moving, deleting
• Template validation and repair
• Reading files and extracting sections/data
• Writing structured content to files
• Git operations (status, branch, mv)
• Environment detection (checking if files/dirs exist)
• Any operation that always produces the same output given the same input
• Pattern matching, categorization by filename/extension
• JSON/YAML/TOML parsing or generation
• Archiving, rotating, or resetting files

JUDGMENT (stays in SKILL.md):

• Asking the user questions (AskUserQuestion)
• Deciding what to do based on context (if/else logic about user intent)
• Assessing completion status
• Choosing what to work on next
• Dispatching agents for tasks requiring reasoning
• Evaluating quality (e.g., "is this content worth promoting?")
• Any operation where different AI instances might reasonably make different choices

3. Present the analysis to the user:

SKILL: <name> (<line-count> lines)

DETERMINISTIC operations found:
- <description of each operation that will become a script subcommand>

JUDGMENT operations (staying in SKILL.md):
- <description of each AI reasoning block>

Proposed script subcommands:
- <subcommand-name>: <what it does>
- <subcommand-name>: <what it does>

Ask user to approve or adjust before proceeding.

Phase 2: Build Script

Create ~/.claude/skills/<skill-name>/script.py following these rules:

Script Architecture:

• argparse with subcommands (one per deterministic operation)
• ALL output is JSON via a shared _out(data) helper
• Handle markdown code blocks correctly when parsing sections (track ``` state)
• Use pathlib for all file operations
• Include --help for every subcommand
• Make executable (chmod +x)
• Use #!/usr/bin/env python3
• No external dependencies — stdlib only

Subcommand Design:

• Each subcommand does ONE thing well
• Required args via --flag (not positional)
• Return {"status": "ok", ...} on success
• Return {"status": "error", "message": "..."} and exit(1) on failure
• Include enough detail in output for AI to make decisions without re-reading files

Code Quality:

• Proper error handling (file not found, permission denied)
• Idempotent where possible (safe to run twice)
• Comments only where logic is non-obvious

Phase 3: Rewrite SKILL.md

Rewrite the SKILL.md keeping:

• Frontmatter (name, description, user-invocable) — unchanged
• UX contracts, invariants, rules — preserved but condensed
• User question definitions — preserved
• Continuation/decision logic — preserved
• Self-assessment criteria — preserved

Replace with script calls:

• Every deterministic operation becomes a code block showing the script call
• Agent Ref / Agent Reference sections → deleted entirely
• Inline bash snippets → replaced with script subcommand calls
• Multi-step file operations → single script subcommand

Preserve and update frontmatter:

• Keep name, description, user-invocable unchanged
• Add allowed-tools if missing — list tools the skill actually uses (Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion, etc.)
• Update description if the workflow changed significantly

Add a Script Reference table at the bottom:

## Script Reference

All commands output JSON. Run from skill directory.

| Command | Purpose |
|---------|---------|
| `subcommand-1 --flag` | What it does |
| `subcommand-2 --flag` | What it does |

Target: Reduce SKILL.md by 50-70% while preserving all behavior.

Phase 4: Test

Run each script subcommand with --help to verify it parses. Run preflight or equivalent read-only command against the current project to verify JSON output. Report results.

Phase 5: Backup

Before overwriting SKILL.md, save original as SKILL.md.pre-codify.

Edge Cases

• Skill has no deterministic operations (pure judgment): Report "Nothing to extract" and skip.
• Skill already has a script.py: Ask whether to merge, replace, or abort.
• Skill is very short (<50 lines): Report "Too small to benefit from codification" and skip.
• SKILL.md references other files: Read those too and include in analysis.

Example

/codify start

Result: The transformation we just did — 637-line SKILL.md with 7 Agent Ref sections became 204-line SKILL.md + 320-line session-init.py script with 9 subcommands.