docs: move the dev skills to .agents/skills/

.claude/skills/agent-eval/SKILL.md → .agents/skills/agent-eval/SKILL.md

@@ -7,7 +7,7 @@ license: Unlicense
 # ncli agent-capability eval
 
 `integration/agent-eval/` is a different integration test from
-`.claude/skills/local-verify`: instead of building `ncli` from local source
+`.agents/skills/local-verify`: instead of building `ncli` from local source
 to check a code change, it drives a **fresh** Claude Code agent -- in a
 container that has never seen this repo -- through the same path any
 real external user/agent would follow (`Fetch

.gitignore

@@ -45,6 +45,11 @@ vendor/
 # Claude Code agent worktrees (transient per-session checkouts)
 /.claude/worktrees/
 
+# Locally-installed Claude Code skills. The shared source of truth for this
+# repo's own dev skills is .agents/skills/, which reaches teammates via
+# clone; what lands here is each person's installed copy.
+/.claude/skills/
+
 # Local cache of ohstr/docs-kit, cloned on demand by `just docs-dev`
 /.docs-kit/
 

AGENTS.md

@@ -149,4 +149,4 @@ or invoking a command in that area:
 
 These skills assume only the `ncli` binary is available — no access to this
 source tree. (Building/contributing to `ncli` itself is a different task —
-see CONTRIBUTING.md and `.claude/skills/local-verify`.)
+see CONTRIBUTING.md and `.agents/skills/local-verify`.)

integration/agent-eval/README.md

@@ -4,7 +4,7 @@ Tests **ncli as consumed by an agent**, not this repo's source. A fresh
 Claude Code agent, in a container that's never seen this checkout, is
 pointed at `https://ohstr.github.io/ncli/PROMPT.md` and follows it --
 same as any real external user/agent picking `ncli` up for the first
-time. Unlike `.claude/skills/local-verify` (builds from local source to
+time. Unlike `.agents/skills/local-verify` (builds from local source to
 check a change), this installs the **published** image/docs, so it catches
 docs drifting from what's actually shipped.