Add tracked AI skills for Blender assets and release QA

tools/README.md

@@ -129,34 +129,25 @@ All Python tooling tests live in [`tests/`](tests/). They use the repository's
 `requirements-tools.txt` (Pillow) and do not start Godot or Blender. See
 [`tests/README.md`](tests/README.md) for the exact discovery command.
 
-## Which tools could become AI skills?
+## Project AI skills
 
-A skill is a **procedure and decision guide**, not a script wrapper. The useful
-split is project-specific judgement that an AI repeatedly needs, not a second
-copy of every command.
+Two project-authored skills now live under
+[`../.agents/skills/`](../.agents/skills/) and are tracked in the repository:
 
-### Recommended: elevate as a project skill
+- [`gig-line-blender-asset-contract`](../.agents/skills/gig-line-blender-asset-contract/SKILL.md) — contract, command order, failure playbook and verification for Blender/NPC/gear/prop work.
+- [`gig-line-release-qa`](../.agents/skills/gig-line-release-qa/SKILL.md) — test/smoke matrix, visual gates, clean build, checksums, tag/publish and release records.
 
-1. **`gig-line-blender-asset-contract` (high value).**
-   Bundle the ordering and contract around `blender_assets.py`,
-   `npc_pipeline.py`, `bake_ao.py`, `npc_proportions.py` and
-   `npc_baseline.py`: the 13-bone `npc_base` contract, tint slots, 8,000/8,000
-   character cap, gear attach space, AO/UV rules, audit → fit → measure →
-   baseline sequence, and the “inspect the diff before committing” rule.
-   This is the workflow most likely to be repeated by an AI and most likely to
-   be damaged by improvising.
+The third-party `blender-skills` pack remains local/gitignored. These project
+skills are guidance only; they do not move scripts into `.agents`, add runtime
+dependencies or replace the tools above.
 
-2. **`gig-line-release-and-visual-qa` (high value).**
-   Capture the sequence `check_docs.py` → `check_screenshots.sh` →
-   `screen_designs.py compare/judge` → clean-worktree `build_releases.sh` →
-   `PUBLISH_DRY_RUN` → checksum/upload checks. Include the optional nature of
-   the vision judge, the clean-worktree trap and the release rollback rules.
+### Remaining skill candidate
 
-3. **`gig-line-reference-provenance` (medium value).**
-   Guide `refboard.py` vs `refcollect.py`: committed Wikimedia-sourced boards
-   with sidecars versus local-only `tmp/research/` downloads, and the rule that
-   generated concepts never override photo evidence. This is more of a policy
-   skill than a command skill.
+- **`gig-line-reference-provenance` (medium value).** Guide `refboard.py` vs
+  `refcollect.py`: committed Wikimedia-sourced boards with sidecars versus
+  local-only `tmp/research/` downloads, and the rule that generated concepts
+  never override photo evidence. This is more of a policy skill than a command
+  skill.
 
 ### Keep as scripts