Create, edit, and pressure-test Claude Code skills and agents against model behavior.
npx skills add https://github.com/blauwtje/exo --skill skills-tool-blauwtje
Please help me install this Agent Skill. Skill: skills-tool Source: https://github.com/blauwtje/exo/tree/main/skills/skills-tool Command: npx skills add https://github.com/blauwtje/exo --skill skills-tool-blauwtje
scripts/pressure.mjs across model:effort cells in with-skill and without-skill arms, writing full answers to files for comparison. - Structured authoring guidance: References cover description writing, skill shape templates, wording registers, fix placement, hole plugging, and blind evaluation. - Size and structure enforcement: Defines token and character budgets for bodies and descriptions, verified with node verify.mjs. - Use Case: When editing a skill that the model keeps ignoring, run a pressure scenario on the pre-edit and edited clones, then have a blind Sonnet judge score the answers to confirm the wording change actually moved behavior. ## Quick Start Ask the AI to use the skills-tool skill to create a new skill for a recurring mistake, running pressure tests before calling it done.High-intent search queries and answers about installing and using this skill.
Run scripts/pressure.mjs with a prompt file, model:effort cells, and the plugin directory. It executes without-skill and with-skill arms in parallel, writing each full answer to a file so you can compare actions and justifications across arms.
A skill description carries only the trigger: the moments it fires, the words the model would have in mind, and "not for" cases naming the owner. Aim at 300 characters, stay within 375, and never summarize the workflow, which the body owns.
Do not create a skill for a one-time fix, a habit of a single project, or a limit a regex can check. Those belong in a commit, CLAUDE.md, or the verifier, because a skill exists only for calls that need judgment.
An untested edit is a guess about model behavior. Run the pressure scenario without the skill first; if that arm passes, the case proves nothing. Close each excuse the with-skill run still produces using the plugging-holes reference, then rerun all cases.
Aim the body at 2,000 tokens, computed as bytes after the frontmatter divided by 4. The verifier fails skills over 2,500 tokens, and a skill that grew too long drops its weakest rule rather than a reason clause.
jellyrockAudit Claude Code session transcripts for friction, performance, permission, and accuracy issues.
vercel-labsRun live Claude Code eval sessions for the vercel-plugin and generate coverage reports.
kamelmhOrchestrates user tasks via selected execution patterns with plan review and progress caching.
DingJun1028Fixes React code patterns that fail ESLint and React Compiler rules.