Skill

Thariq Shihipar, on the «Claude Code» team at Anthropic, spends the opening of his 3 June 2026 post correcting a common reduction of skills to a single markdown file. A skill is a folder: instructions, scripts, resources, configuration and «hooks conditionnels», which the agent explores and manipulates. The directory tree itself carries progressive disclosure, routing the model toward the right reference file for the situation.

The advice that follows cuts against the usual documentation instinct. A skill should record what contradicts the model's default behaviour, since Claude already knows how to code and can read the codebase. The highest-signal content is the gotchas section, which collects failure points actually encountered. Description fields are addressed to the model and state when to trigger the skill. Configuration arrives through a setup flow that stores parameters in config.json or asks the user via «AskUserQuestion».

Hugo Lassiège describes the same object from a solo practice. In his 28 July 2026 account of his usine logicielle, skills answer one of three questions about the agent: what it can do deterministically, sitting between what it knows and what stops it when it errs. He keeps around thirty, creates one at the third repetition of a procedure, and finds those covering multi-file procedures the most profitable.

Anthropic tracks which skills get used with a PreToolUse hook that logs invocations; skills emerge in a shared sandbox folder, circulate on Slack, then reach an internal marketplace by pull request. Lassiège reports the other end of that pipeline: for lack of packaging, his skills get copy-pasted.