Turn Claude Code into a full game dev studio — 49 AI agents, 72 workflow skills, and a complete coordination system mirroring real studio hierarchy.
Studio Tests — Claude Instructions
This folder is the quality assurance layer for the Claude Code Game Studios skill/agent framework. It is self-contained and separate from any game project.
Key files
| File | Purpose |
|---|---|
catalog.yaml |
Master registry for all 74 skills and 49 agents. Contains category, spec path, and last-test tracking fields. Always read this first when running any test command. |
quality-rubric.md |
Category-specific pass/fail metrics. Read the matching ### section for the skill's category when running /skill-test category. |
skills/[category]/[name].md |
Behavioral spec for a skill — 5 or more test cases + protocol compliance assertions. |
agents/[tier]/[name].md |
Behavioral spec for an agent — 5 or more test cases + protocol compliance assertions. |
templates/skill-test-spec.md |
Template for writing new skill spec files. |
templates/agent-test-spec.md |
Template for writing new agent spec files. |
results/ |
Written by /skill-test spec when results are saved. Not gitignored — add it to .gitignore to keep results out of commits. |
Path conventions
- Skill specs:
studio-tests/skills/[category]/[name].md - Agent specs:
studio-tests/agents/[tier]/[name].md - Catalog:
studio-tests/catalog.yaml - Rubric:
studio-tests/quality-rubric.md
The spec: field in catalog.yaml is the authoritative path for each skill/agent spec.
Always read it rather than guessing the path.
Skill categories
gate → gate-check
review → design-review, architecture-review, review-all-gdds
authoring → design-system, quick-design, architecture-decision, art-bible,
create-architecture, ux-design, ux-review
readiness → story-readiness, story-done
pipeline → create-epics, create-stories, dev-story, create-control-manifest,
propagate-design-change, map-systems, vertical-slice
analysis → consistency-check, balance-check, content-audit, code-review,
tech-debt, scope-check, estimate, perf-profile, asset-audit,
security-audit, test-evidence-review, test-flakiness
team → team-combat, team-narrative, team-audio, team-level, team-ui,
team-qa, team-release, team-polish, team-live-ops
sprint → sprint-plan, sprint-status, milestone-review, retrospective,
changelog, patch-notes
utility → all remaining skills
Agent tiers
directors → creative-director, technical-director, producer, art-director
leads → lead-programmer, narrative-director, audio-director, game-designer,
systems-designer, level-designer, qa-lead
specialists → gameplay-programmer, engine-programmer, network-programmer,
ai-programmer, tools-programmer, ui-programmer, ux-designer,
technical-artist, sound-designer, performance-analyst, prototyper,
writer, world-builder
engine/godot → godot-specialist, godot-gdscript-specialist, godot-csharp-specialist,
godot-shader-specialist, godot-gdextension-specialist
engine/unity → unity-specialist, unity-ui-specialist, unity-shader-specialist,
unity-dots-specialist, unity-addressables-specialist
engine/unreal → unreal-specialist, ue-gas-specialist, ue-replication-specialist,
ue-umg-specialist, ue-blueprint-specialist
operations → devops-engineer, release-manager, live-ops-designer,
community-manager, analytics-engineer, economy-designer,
localization-lead
qa → qa-tester, security-engineer, accessibility-specialist
A tier is the directory a spec lives in. The rubric category is the agent's
category: in catalog.yaml and can differ: security-engineer and
accessibility-specialist sit in qa/ but are rated as specialist.
Workflow for testing a skill or agent
- Read
catalog.yamlto get thespec:path andcategory:— from theskills:entry for a skill, theagents:entry for an agent - Read the skill at
.claude/skills/[name]/SKILL.mdand every file in itsreferences/folder, or the agent at.claude/agents/[name].md - Read the project
CLAUDE.mdand the files it imports with@: every subagent receives them, so a rule stated there (for example a redirect incoordination-rules.md) counts as the skill's or agent's own - Read the spec at the
spec:path - Evaluate the spec's Static Assertions, each case's assertions, and its Protocol Compliance list
- Offer to write results to
results/and updatecatalog.yaml
Fixture defaults
Every skill spec has a ## Fixture Defaults section above its cases. A case's
sandbox project starts from this project.yaml block:
modes:
rigor: minimal # → workflow minimal, docs.density terse, qa.level minimal,
# story_granularity coarse, review_mode solo, team.size individual
automation: collaborative
- A setting the case's Fixture names wins over the block, and an explicit
sub-knob (
modes.workflow,modes.review_mode,qa.level, …) wins over therigorpreset (.claude/docs/config-resolution.md§ Defaults). - A case that says
project.yamlsets nomodeskeys, or that there is noproject.yaml, runs without the block: it tests the unset defaults. - There is no
project.local.yamlunless the case names one. - A case whose expected behaviour changes with a setting pins that setting in
its own Fixture, with the reason — e.g.
modes.workflow: standard(pinned: unset resolves to minimal, which recommends the next build instead of a sprint).
The block is pinned so that a change to the framework's defaults cannot silently change what a case tests. A spec may set a different default in its own section when most of its cases need one.
Workflow for improving a skill
Use /skill-improve [name]. It handles the full loop:
test → diagnose → propose fix → rewrite → retest → keep or revert.
Spec validity note
Specs in this folder assert the correct behavior, not whatever a skill or agent does today. A failing assertion means the skill or agent is wrong until shown otherwise: fix it and re-run, and never weaken an assertion just so it passes. Specs can be wrong too — when one asserts behavior that is itself incorrect, correct the spec toward the right behavior, not toward what the skill currently does.
This folder is deletable
Nothing in .claude/ imports from here, and every other CCGS skill and agent works
without it — but /skill-test and /skill-improve read it by path. Without it,
/skill-test static still works, /skill-test spec and /skill-test category
stop working, /skill-test audit reports that no catalog exists, and
/skill-improve skips its category checks (the table in README.md has the detail).