Computed Skills — Development skill for Claude Code

Computed Skills

[](LICENSE) [](https://www.python.org/) [](https://claude.ai/claude-code) [](https://github.com/Joncik91/computed-skills/stargazers) []()

**Skills that adapt to what's actually happening.**

A static skill gives the same instructions every time. A computed skill runs a script first — analyzes the situation, pre-digests data, picks a strategy — then hands the agent tailored markdown. The agent never knows a script was involved.

Table of Contents

• Why Computed Skills
• Quick Start
• Three Patterns
• Security Review
• When to Use What
• How to Build One
• Examples
• Production Case Study
• Prior Art
• Requirements
• Security
• License

Static:    SKILL.md ──────────────────────────→ Agent reads markdown
Computed:  SKILL.md → runs script → markdown ─→ Agent reads markdown

Works with [Claude Code](https://claude.ai/claude-code) and any skill system that supports `!`command`` preprocessing. Any language that prints to stdout.

Why Computed Skills

Static skills have a fundamental problem: they can't see context. A static code review skill gives the same checklist whether you changed 2 files or 40, whether you touched auth code or CSS. A static deploy checklist shows 20 items every time — and agents learn to skim them all.

Computed skills fix this by letting a script decide what the agent sees:

2 aut