Refactor any Claude Code skill by extracting deterministic operations into a Python script.

Every time a skill runs, the AI re-reads hundreds of lines of prose to re-derive mechanical operations: "does this file exist?", "run this regex", "copy with timestamp". This is wasteful. codify splits skills into two parts: code for mechanics, prompts for judgment.

curl -fsSL https://raw.githubusercontent.com/anombyte93/codify/main/install.sh | bash

Installs to ~/.claude/skills/codify/. Run the same command to upgrade.

# Check for updates
bash ~/.claude/skills/codify/install.sh --check-update

/codify <skill-name>

Where <skill-name> is the directory name under ~/.claude/skills/.

Reads the skill's SKILL.md and classifies every block:

Creates a Python script with argparse subcommands. Every subcommand:

• Does one thing well
• Outputs JSON to stdout
• Responds to --help
• Uses stdlib only (no dependencies)

Replaces deterministic prose with script calls:

# Before (in SKILL.md)
Check if .taskmaster/ exists. If it does, look for PRD files
in .taskmaster/docs/. Read tasks.json to count tasks...
(~40 lines of detection logic)

# After (in SKILL.md)
python3 script.py preflight
# Returns JSON: has_taskmaster, prd_path, task_count, ...

• Runs --help on every subcommand
• Runs read-only commands to verify JSON output
• Saves original as SKILL.md.pre-codify

script.py = AI's calculator and file manager (deterministic tools)
SKILL.md  = AI's playbook for user interaction and decisions (judgment)

The AI calls script.py like any other tool. Script returns structured JSON. AI interprets the results and makes decisions. Neither replaces the other.

• Claude Code CLI
• Python 3.8+

MIT - see LICENSE

Built by Atlas AI