Agent skills

Installing Knit’s agent skills and commands into an AI harness.

Install the agent skills into a harness

Equip an AI harness to drive the experiment with knit skills install.

APIs: skills:install

Knit ships agent skills and commands — the know-how an external AI harness (Claude Code is the reference) needs to write a machine profile, author an experiment, plan and run a sweep, and analyze results. Install them into the project with knit skills install:

$ ./exp.sh skills install

This downloads the latest agent/ folder from the knit GitHub repository (with curl + tar, the same way bootstrap provisions its tools) and copies the skills and commands into .agents/ — the canonical cross-harness location. The skills are not embedded in knit.sh, so they always track the published repo rather than the installed script.

The copy merges into any skills/commands already there, so skills from other sources are preserved and re-running the command is idempotent. Install also drops an AGENTS.md pointer at the project root (so an agent orients itself before any skill is loaded), but never overwrites an AGENTS.md the project already has.

Claude Code. Claude Code looks under .claude/ rather than .agents/. Add --claude to point it at the same install — there is one real copy on disk, not a per-harness fork:

$ ./exp.sh skills install --claude

This keeps .claude/skills and .claude/commands as real directories and symlinks each knit item into them one at a time — every skill and every command its own symlink. So the project’s own .claude skills/commands and its .claude/settings.json sit alongside the links untouched, and an item whose name already exists is skipped with a warning rather than overwritten. Re-running is idempotent (knit’s own links are refreshed quietly).

Pinning a version. --ref installs the skills from a specific branch, tag, or commit of knit-sh/knit instead of the default branch:

$ ./exp.sh skills install --ref v1.4.0

Because install writes a tree of files rather than a single stream, there is no --print mode.