skills
Install the AI orchestration skills that ship with the CLI. The skills sequence the versioning workflows — run the commands in order, stop at the human review gates — and never generate report or changelog content themselves; the version commands produce all content.
Usage
specs skills install [--dir <path>]| Flag | Description |
|---|---|
--dir <path> | Skill directory to emit into (default: .claude/skills) |
Two skills are emitted:
| Skill | Orchestrates |
|---|---|
specs-cli.premerge | specs version figmapremerge fetches, generates, and reports → human reviews the report → merge proceeds or not |
specs-cli.release | specs version report → human review → specs version cut → specs version tag → publish |
specs-cli.roles.propose | specs analyze → read the live components over the bridge → write analysis/roles.proposed.md → human marks up the Decision column |
specs-cli.roles.apply | Parse the approved decisions → annotate the Figma file one page at a time → human confirms in Dev Mode → fetch, generate, analyze roles to prove the round trip |
roles.propose is the one skill here that authors its report rather than only sequencing commands. Which role an element takes is a judgment — the discriminator between checkbox, switch and togglebutton is what the control does when operated, which no prop records. A rule reading checked and indeterminate would be right about checkbox and wrong about toggle, switch and togglegroup. So the report carries a confidence column, the evidence behind each row, and a human gate before anything reaches a Figma file.
The emitted files are canonical and overwritten on every install — re-running after a CLI upgrade refreshes them to match the commands they orchestrate. Keep local customization in your own separate skills rather than editing the emitted ones.
See Also
- version — the commands these skills orchestrate
- Reports & Changelogs — the outputs the human gates review