Skills, tools & sources

The skills tab shows what the agent knows how to do in this project: the skills under .agent/skills/ and the tools under .agent/tools/ — the same files the agent reads, rendered for you.

One source of truth, many agents

.agent/skills/ is the source; agent-specific folders are symlinks into it:

.claude/skills → ../.agent/skills

Add or edit a skill under .agent/skills/ and every agent sees it immediately. The rule is the same for you and the agent: edit there, never inside a dot-folder.

Sources: skills you didn’t write

Skills and tools can come from a Speccify source — a skills repo added under Sources in the skills tab. Expanding a skill from a source runs the expand step: normal skills land in .agent/skills/, tool contracts in .agent/tools/, and .agent/speccify/expansions.yaml records where each came from, at which version, and whether each tool’s implementation is verified on this platform.

Two rules keep the bookkeeping sane:

Edit the skills and tools, not expansions.yaml — the record is maintained by Speccify; hand-edits would make it lie.
Tools are pre-approved deliberately. For Claude Code, .claude/settings.json allowlists running what’s under .agent/tools/ — implementations are contract-checked, so the agent can call them without a permission prompt each time.

When the agent uses an expanded skill while working a ticket, it logs an agent_run line in the ticket’s history naming the skill and tool — the inspector is the trail of what was used; there is no separate log file.