AI-generated skills get reviewed too: skill-cutter removes duplicates and unnecessary rules

A skill is a recipe for an agent to handle a specific task. skill-cutter calls for removing generic engineering advice, duplicates and rules the agent already received in its system instructions. It must preserve task-specific decisions and the required sequence of steps.

Extra rules get in the way. On Aug 8, 2026, SmolForge described removing JFDI 3 days after adding it: 86 lines in SKILL.md and 4 lines of description for the interface. The same write-up mentioned an agent turn lasting 7 hours 23 minutes, rather than a controlled comparison of speed before and after removal.

SmolForge also removed the 63-line Codebase Maintainability Guardrails: its broad trigger attached architecture and testing requirements to routine tasks. The activation conditions for the adjacent Test Strategy Hardening and Autoreview skills were narrowed. skill-cutter checks exactly these parts of description: it removes triggers such as “all requests”, adds exceptions and aligns the description with agents/openai.yaml.

How to set up the review. According to the swyxio/skills README as of Oct 3, 2026, the repository is installed with `git clone https://github.com/swyxio/skills.git`. Then place a symlink to the selected skill folder in your tool's directory:

- Codex: `~/.agents/skills/`. - Claude Code: `~/.claude/skills/`. - Cursor: `~/.cursor/skills/`.

In Audit mode, skill-cutter reviews the skill and suggests cuts without editing it. In Cut mode, it applies local changes when explicitly asked to shorten or simplify the skill.

Primary sources: [swyxio/skills](https://github.com/swyxio/skills), Oct 3, 2026. SmolForge write-up, Aug 8, 2026.

After making cuts, skill-cutter requires before-and-after counts of lines, words and files, reasons for removals, trigger changes and validation results, rather than reducing size for its own sake.