v1.3.0 — make me a game2026-11-09
"Make me a game" is the main road: /prototype asks a few questions, writes a one-page brief, builds a browser game from it, checks it against the brief, and fixes what the check finds before you get it. After you play it, you change it by asking, and when you like it you grow it into a real web project and build on it one feature at a time. A minor release after 1.2.0.
Highlights
- Make me a game. One round of at most about eight questions drawn from the genre, with the usual answer shown so "you pick" works. A vague idea gets two or three one-line pitches first. The answers become design/game-brief.md: your words kept exactly, how it looks, its first minute, its challenge, your systems, and Keep, Avoid and Out of scope lines.
- The build and its check. The game is built in your session as a browser game in prototypes/<name>/ that opens with a double-click: no emoji art, no copy of another game, and not trivially easy. A short scripted check runs before you get it: a headless browser presses a few keys, a screenshot is compared with the brief, and every line of the brief is marked pass, missing, broken or not assessed. The handover card shows what is still broken and what was guessed, and ends "Not yet played by a person".
- Change it after play. Each change is checked against the brief's Keep, Avoid and Out of scope lines, gets a light check, and is one commit you can go back to. "Too easy" adds a new rule before any number moves.
- The tweak panel, added when you ask to tune: a slider for each value in config.js, opened with the backquote key or ?tune. Paste the saved block back and /prototype writes it into config.js.
- Share it, or bring your own game in. one-shot-share.sh makes a copy without the working files; /prototype --from <path> copies in a browser game you already have so it can be changed and grown.
- Grow it. On PROCEED, a one-shot grows into a Vite and TypeScript web project with data files, tests and a same-seed run, then takes one feature a sitting. Stories, epics and gates only when you ask for them.
- /cut, a new skill: cuts and a polish list to finish the game from its brief, and you pick each cut.
- Web is an engine. engine.name: Web, with three.js, Phaser and PixiJS starters and a pinned web reference. Every engine's folders and commands come from one table, .claude/engines/engines.tsv. Genre sheets for 69 genres live in .claude/genres/.
- Studio mode. The agents, the team-* pipelines and the design-document ladder are now called studio mode, and /start and /next-step say when it fits.
- Renamed skills: /help is /next-step, /settings is /project-settings, /changelog is /write-changelog. The old names still work.
- Fixes, among them: rotating the session state can no longer lose its notes (#158); git commit -m "message" -a is checked on a machine without Python (#155); the Godot reference says Vulkan is the Windows default and Compositor is a resource (#143); five skills that change a file in place can now edit it (#156). The full list is in the CHANGELOG.
What is shown, and what is not yet
In a blind comparison, the brief written from these questions was picked over a hand-written brief for all three game ideas tried. Growing was shown on one game, grown through six stories. The engine route and deepening one system in a design document are built but not yet shown in a test like these.
Upgrading from 1.2.0
Take .claude/scripts/, .claude/hooks/, .claude/genres/ and .claude/engines/ whole. If you edited settings.json, every hook command has a new form, and there are 28 new allow rules for the read-only scripts: UPGRADING.md → v1.2.0 → v1.3.0 lists every merge and the three renamed skill specs to delete. A skillOverrides entry or permission rule naming an old skill needs the new name.
Tested on
- Platforms: Windows, Linux (bash 5 and 3.2) and macOS (bash 3.2 with BSD tools), with and without jq.
Contributors
Thank you to everyone who reported, proposed or fixed something in this release:
- @rudycelekli: found and proposed fixes for three bugs. Rotating the session state could lose its notes (#158). The commit check missed -a after a quoted message on a machine without Python (#155). Five skills could not edit a file in place (#156).
- @mamonbleau: reported that the Godot reference had the Windows renderer default and Compositor wrong (#143).
Full changelog: https://github.com/Donchitos/Claude-Code-Game-Studios/compare/v1.2.0...v1.3.0
v1.2.0 — the main path, rebuilt and tested2026-10-19
The story skills are rebuilt around shared helper scripts, a read guard stops whole-file reads of large design docs, the engine references move to Godot 4.7 and Unreal Engine 5.8, and no skill pre-approves every shell command any more. A minor release after 1.1.3.
Highlights
- The main path is rebuilt. /start, /create-stories, /story-readiness, /create-epics, /dev-story and /story-done follow Anthropic's skill-authoring guidance: shorter, with rarely used branches in references/ and fixed steps handed to helper scripts, so the same question gets the same answer each time. /start works out the rigor from what you describe and shows it on one card. /create-stories proposes an S, M or L estimate for each story.
- Stories list their files, and evidence knows when its code changed. The new story template has a ## Files section: /create-stories plans it, /dev-story fills it and /story-done checks scope against it. /dev-story stamps a story's screenshots with the files they prove, and /story-done names any that changed since (#135, proposed by @jiangshoujie).
- A read guard for large design docs. A whole-file read of a doc over 40 KB under design/, docs/architecture/ or production/session-state/ is refused with the doc's headings, so the retry reads one section (#136, #140, suggested by @mamonbleau). It also ends the failed whole-ADR reads reported in #64 by @Shtruk. features.read_guard: off turns it off.
- Session start says when a newer release is out. One line, from at most one GitHub request a day made in the background; nothing about your project is sent. updates.check: false turns it off, and it is always off under CI.
- Engine references for Godot 4.7 and Unreal Engine 5.8. Unity stays on 6.3 LTS and lists 6.4 to 6.6 and the 6.7 beta. /test-setup writes current CI actions. Unreal 5.8 support was first proposed in #121 by @spirit986. Staying on Godot 4.6 or Unreal 5.7? Keep your current engine folder.
- Helper scripts and registry checks. New scripts in .claude/scripts/ read story facts, ADR status, a Markdown section, the stage and the session checkpoint for the skills. More ship for you to run (bug counts, sprint arithmetic, tech-debt scores, flaky-test rates and others), each with --help. registry.sh validates the three registries and hands out IDs that are never reused.
- Warnings for writes the skills never make, such as replacing an existing EPIC.md or project.godot, crossing the prototype boundary or renumbering a registry entry. They warn and never block.
- /sprint-plan checks that the sprint goal can be shown and lists anything it needs that no story builds (based on an idea in #58 and #76 by @FreedomPortal).
- Fixes, among them: the commit check skips GDD copies in design/gdd/ subfolders such as archive/ (#133, reported by @mamonbleau); a repeated modes: key in project.yaml is merged and named instead of wiping the first; a multi-line value in project.yaml is read as text, never as settings; a stage never advances past a gate check that could not run; /architecture-review reuses a saved review only when nothing it read has changed, and single-gdd writes only that system; playtest records count as evidence from production/qa/playtests/; /sprint-plan never takes your answer to its QA-plan question as approval to write; in a Godot project, evidence screenshots under production/ are no longer imported or packed into an export. The full list is in the CHANGELOG.
Security
- Skills pre-approve only the commands they run. 34 skills had a bare Bash entry in allowed-tools, which let them run any shell command without asking in the default permission mode. Each now names its framework scripts and read-only commands. Anything else, including your commands.test, asks first: approve it once, or add Bash(<your command> *) to .claude/settings.local.json.
- Tagging, merging and deploying are flagged before they run. The push check also warns on git tag, git merge, gh release, butler push and steamcmd +runappbuild.
- Agents never file upstream. No agent files or comments on an issue or pull request in an engine, plugin or vendor repository, even when asked; it writes a draft for you to submit.
- The same .env protection with fewer rules, and the local settings template no longer suggests Bash(git *), which approves git push too.
Upgrading from 1.1.3
Take .claude/hooks/ and .claude/scripts/ whole, in the same step: the hooks call the new helpers, and a 1.1.3 yaml-helper.sh breaks /settings and the rebuilt skills. If you edited settings.json, add the four new hooks, remove the PostCompact entry together with post-compact.sh, and add worktree.baseRef: "head" (after a git merge, check these rather than adding them again); UPGRADING.md → v1.1.3 → v1.2.0 lists every merge. If you wrote skill specs, move them from CCGS Skill Testing Framework/ to studio-tests/. If a registry has an unquoted status: superseded_by: ADR-NNNN line, quote it; registry.sh names each one. In a Godot project, take production/.gdignore and delete the .import files Godot made under production/. Staying on Godot 4.6 or Unreal 5.7 and upgrading with git merge? The merge takes the new engine folder without a conflict; run git checkout HEAD -- docs/engine-reference/<engine>/ before you commit it. Run /architecture-review once before the Technical Setup → Pre-Production gate: reports from 1.1.3 count for nothing there. Expect a permission prompt the first time a skill runs a command it no longer pre-approves.
Tested on
- Platforms: Windows, Linux (bash 5 and 3.2) and macOS (bash 3.2 with BSD tools), with and without jq.
- Claude Code: live skill runs on 2.1.296, 2.1.302 and 2.1.303.
- Godot: gdUnit4 6.2.2 on Godot 4.7.2, locally and in the /test-setup CI workflow on GitHub Actions.
- Upgrading: both git routes rehearsed on a 1.1.3 project.
Full changelog: https://github.com/Donchitos/Claude-Code-Game-Studios/compare/v1.1.3...v1.2.0
v1.1.3 — long skills keep their later steps2026-10-08
Long skills keep their later steps through a long session, a question nobody answered is never taken as an answer, and the shipped permissions approve less. A fix release for 1.1.2.
Highlights
- The 14 longest skills keep their later phases after a compaction. Claude Code re-attaches only the first part of a skill after a compaction, so the later phases of a long skill could drop out mid-session. /adopt, /architecture-decision, /architecture-review, /create-architecture, /design-system, /dev-story, /gate-check, /prototype, /review-all-gdds, /setup-engine, /smoke-check, /story-done, /test-setup and /ux-design now keep each phase in its own file under references/. The steps themselves are unchanged, and /skill-test and /skill-improve read those files too.
- A skipped, dismissed or empty question is not an answer, in any mode: nothing is written, set or picked on it, and it is asked again. Agents return their options as a list for the skill to ask, and CLAUDE.md names the three writes that need no "May I write?".
- Settings read the same from a subfolder. Started in src/, the settings reader missed per-system overrides, the notes about invalid or locked values, and the older review-mode.txt and stage.txt. /settings and /perf-profile work from a subfolder too. A local platform.cert_tier can no longer override the lock.
- Hooks and checks. The commit check works for a project in a subfolder of a larger repository, checks a JSON data file stored with Git LFS through its working copy instead of blocking it, sees files a git add in the same command stages, and stops asking non-system files in design/gdd/ for GDD sections (#133). /smoke-check no longer waits forever on Unreal.
- Every skill opens with a title (#83, thanks @specterslient95-lgtm), and /start writes a single modes: block.
- Smaller fixes: registry lookups no longer match combat-ai when asked for combat; thirteen skills name the minimal route as a next step; team-polish and team-ui keep their core members at lean; /scope-check's verdict bands no longer overlap.
Security
- The shipped allow rules no longer approve deleting a branch or writing files without asking. Four rules matched any text after the command name, so git branch -D, git log/git diff --output=<file> and python -m json.tool <in> <out> ran without a prompt. Those rules are removed; read-only git commands still run without one.
- Start Claude Code at the project root. It loads .claude/settings.json only from the folder it starts in, so a session started in a subfolder runs without the project's hooks, deny rules and ask-first mode. Claude now warns when it was started somewhere else.
- .env files are denied to Edit.
Upgrading from 1.1.2
No setting changes meaning and nothing is migrated automatically. If you merged settings.json or copied its allow rules into settings.local.json, remove the rules listed in UPGRADING.md → v1.1.2 → v1.1.3; a copy keeps approving them. If you edited one of the 14 long skills, take the shipped skill and make your change again in the phase file that now holds that text. If your project.yaml has two modes: blocks, merge them into one.
Tested on
- Platforms: Windows, Linux (bash 5 and 3.2) and macOS (bash 3.2 with BSD tools), with and without jq.
- Claude Code: live skill checks on 2.1.292 and 2.1.293.
Full changelog: https://github.com/Donchitos/Claude-Code-Game-Studios/compare/v1.1.2...v1.1.3
v1.1.2 — the default path works end to end2026-09-29
The default path works end to end, Unity and Unreal get the same care as Godot, and the checks tell you what they found. A fix release for bugs in 1.1.1, each reproduced before it was fixed.
Highlights
- The minimal path works start to finish. /help, /story-done, /sprint-status and the status line follow it. Skills that expect a concept doc read design/game-brief.md instead, and there are no more dead ends into sprints.
- Engine commands that actually run. Godot tests run, Unity tests compile, and Unreal tests compile and are found. /smoke-check runs your tests first.
- Unity and Unreal get what Godot gets. Rules, hooks, checklists and agents know Assets/, Source/ and Content/, and check their own engine reference.
- Warnings you can see. Hook warnings now reach you. The commit and push checks find git however it's written, ignore here-doc text, and run in PowerShell too. The push warning names the branches it protects: main, master and release/*.
- Gates you can pass, and verdicts that don't overclaim. /gate-check passes a project that is ready. Every verdict has a word for "could not check" (NOT ASSESSED). A release waits for the release gate, which now checks security and localisation. QA works at every qa.level.
- Skills ask before they write, including team skills before their agents write code.
- One evidence rule, one bug severity scale. Stories, ADRs and sprints agree, and a Proposed ADR blocks a story again.
- Agents. Directors answer each gate in its own words. devops-engineer follows trunk-based development. Director memory stays in your project.
- Skills read the whole argument. /scope-check inventory crafting system no longer checks just inventory.
- Agent test specs run and match their agents (#131), and agent models are documented as they actually run (#67).
- Two small additions ride along: the brief's Reference game line, and /start spotting a clone still wired to this repo. devops-engineer and community-manager now run on Sonnet instead of Haiku, so they cost a little more per call.
Upgrading from 1.1.1
Nearly every skill, hook and agent changed. No setting changes meaning and nothing is migrated automatically. If you ran /setup-engine on 1.1.0 or 1.1.1, your project.yaml still holds the old test and build commands. UPGRADING.md → v1.1.1 → v1.1.2 lists what to overwrite, what to merge by hand (settings.json, rules, .gitignore) and the exact command changes. Python 3 and Bash are now listed as requirements.
Tested on
- Platforms: Windows, Linux (bash 5 and 3.2) and macOS, on Claude Code 2.1.281–2.1.284.
- Engines: every Windows engine command was run on Godot 4.6.1 (gdUnit4 6.1.3), Unity 6000.3.23f1 and Unreal Engine 5.7 before it was written down.
- Checked against documentation only, not run: the Unreal Linux and macOS commands, the Unity CI template, and the Unity assembly-definition guidance.
Full changelog: https://github.com/Donchitos/Claude-Code-Game-Studios/compare/v1.1.1...v1.1.2
v1.1.1 — skills and agents work again outside auto mode2026-09-24
Skills and agents work again outside auto mode. A fix release for #128.
Fixed
- Skills no longer abort before they start. In 1.1.0, every skill that reads your config began with a shell command Claude Code refuses to run unless it is explicitly approved. In the permission mode CCGS ships with (default), 66 of the 74 skills stopped before doing anything — from any subdirectory of your project, and whenever Claude invoked a skill on its own, in auto mode too. Each skill now runs the config lookup as one plain command and pre-approves exactly that command in its own frontmatter. Nothing else is approved.
- game-designer and creative-director launch again. Both preload skills, so the aborting command stopped them at startup with Shell command permission check failed … Contains expansion. They keep their preloaded skills and still have no shell access.
- Config is found from any folder. Starting Claude Code inside src/ or design/ now reads the project's project.yaml.
- /changelog works in a repository with no commits yet.
Checked on Claude Code 2.1.277 and 2.1.281, in every permission mode, from the project root and from subdirectories.
Upgrading from 1.1.0
Overwrite .claude/skills/, .claude/hooks/yaml-helper.sh and .claude/docs/config-resolution.md. No settings or config changes. If you customised a skill or wrote your own that copied the 1.1.0 config line, update two lines in it — see UPGRADING.md → v1.1.0 → v1.1.1.
Thanks @mamonbleau for the report and the follow-up that narrowed it down.
Full changelog: https://github.com/Donchitos/Claude-Code-Game-Studios/compare/v1.1.0...v1.1.1
v1.1.0 — one config file, and a process weight you choose2026-09-23
One config file, and a process weight you choose — one you can see and adjust, not a question you answer once at setup and forget. A solo game jam and a studio production no longer have to carry the same overhead.
Added
- Ask-before-write now holds, whatever your personal settings say. CCGS sets Claude Code's permission mode to default in the project settings, so every approval gate in the framework actually fires. If your own Claude Code config runs in an auto-accept mode, that used to switch the whole collaboration protocol off silently and agents would write files you never approved. Override it in .claude/settings.local.json if you really want to, but read .claude/docs/setup-requirements.md first.
- project.yaml — one place for your project's configuration. Engine, naming conventions, specialists and process settings all live here. It replaces three separate files that used to hold this between them.
- project.local.yaml — your personal settings, not your team's. Git ignores it, so you can run leaner reviews or fewer confirmations on your own machine without changing anything for anyone else.
- modes.rigor — one question instead of six. Set minimal, standard or full and it configures six underlying settings at once. You can still override any single one; the rest stay where rigor put them.
- minimal — no design documents required, terse writing, light test evidence. The fastest route from setup to running code.
- standard — 5 required design sections, balanced depth.
- full — all 8 design sections, thorough docs, full test evidence on every story.
- Per-system exceptions. One system can be held to a higher standard than the rest of the project — your combat system can require the full treatment while everything else stays light. Exceptions can only make requirements stricter, never looser.
- A real jam path, not the full pipeline with switches turned off. minimal now rests on a one-page game brief with six fields, replacing the 30-section concept document, the systems breakdown and the per-system design docs. Four steps to running code: pick your engine → write the brief → /create-stories → /dev-story. On Godot, /setup-engine now also creates the project.godot the engine needs to open what those steps produce — until it did, the path delivered real source files with no project to load them, and the claim above was not true. Unity and Unreal projects must still be created in their own editor first; CCGS will not fabricate one, because it cannot source what those files should contain.
- modes.reviewmode — how many director agents review your work. On the team- orchestrators, full and lean currently behave the same. Those nine skills only ever call phase-gate reviewers, and lean is defined as "skip the director gates that aren't phase gates" — so it has nothing to skip. solo does differ: it turns director gates off entirely. Elsewhere (design and architecture review) all three levels differ as described.
- modes.automation — how often you're asked to confirm before something happens. In autonomous mode, decisions are written to a log instead of interrupting you. You can always list categories that must ask regardless. Four skills ignore this setting on purpose and always ask: /setup-engine, /gate-check, /hotfix and /day-one-patch — engine choice is a one-time irreversible decision, a gate verdict nobody reads defeats the gate, and the two release skills need explicit sign-off. See .claude/docs/automation-modes.md for the full list and reasons.
- qa.level — how much test evidence is required before a story is done.
- docs.density — how deeply written sections go.
- team.size — which agents take part by default. Affects review depth only; it never changes what a story has to deliver.
- testing.strict per test type — a logic story can block on a failing test while a visual story stays advisory.
- /settings — see your effective configuration, check one value, or change it. /settings --local writes to your personal file. It warns you when a change would be overridden by something else.
- /start rework — asks everything it needs up front and writes a complete config in one pass, instead of scattering questions across later steps.
- Migration tooling for existing projects. A converter moves a v1.0 project onto project.yaml. It previews by default, only deletes old files once it has proved every value survived, and deletes nothing if anything is missing. /adopt spots an older project and runs it for you, and you'll get a reminder at session start if you have old config files and no project.yaml.
- Recommendations for how much process your project needs. Describe the game you're making and CCGS suggests a level, instead of asking you to guess what "standard" means for a weekend jam versus a two-year production.
Changed
- Claude no longer starts the four risky commands by itself. /hotfix, /day-one-patch, /dev-story and /story-done now run only when you type them. They bypass the sprint process, patch shipped builds, write game code and close out stories — none of that should begin because a conversation drifted near the topic. Everything else still gets suggested as before.
- Engine specialists can only hand work to their own sub-specialists. The Godot, Unity and Unreal specialists are now limited to the four sub-specialists under each of them. Previously any of them could call any agent in the studio. This was already the documented rule; now it is enforced rather than trusted.
- Fifteen agents that never delegated no longer can. They were carrying the ability to spawn other agents and never using it.
- Light is now the default, not just the recommendation. modes.rigor defaults to minimal instead of standard, so a project that never opens /settings gets the fast path: a one-page brief instead of a full design pipeline, terse writing, light test evidence, no director review panels. We changed this because we measured the old default. Built both ways, the heavier tier cost several times more to reach working code, did not produce a better result, and gave nothing back when someone else picked the project up. It was not buying more; it was buying nothing we could detect. If your project outgrows the light path, /help and /gate-check will say so, and /settings modes.rigor=standard is one command. Projects that already set rigor explicitly are completely unaffected.
- Setup steers light by default. /start now recommends the minimal, guided path for new projects, so you land somewhere light and step up if you need to.
- Setup asks about your game, not about the framework. The question is "what best describes what you're building?" — a game jam, a focused project, something systems-heavy — rather than "how much process do you want?"
- Your process level is always visible. The status line shows it next to your project stage, e.g. Concept · minimal.
- Advice goes both ways. /help can suggest tightening up as a project grows, not just lightening up when you sound overwhelmed.
- Moving into Production offers a check-in on whether the level you chose at the start still fits. It's a suggestion — it never changes the verdict.
- team.size: solo is now team.size: individual, so it isn't confused with review_mode: solo, which means something different.
- Less of your session is spent on the framework itself. Roughly 40% of the per-turn overhead was removed, leaving more room for your actual game.
- Picking up after a /clear or /compact is more reliable. Your session checkpoint is now a defined part of your session notes rather than an ever-growing file, so what comes back is the same every time and leaves far more room for actual work. Older notes can be archived without losing anything.
v1.0.0 — Claude Code Game Studios2026-05-13
First stable release. Upgraders from v1.0.0-beta: see UPGRADING.md.
Headline changes
/prototype and /vertical-slice are now two distinct gates
- /prototype = concept prototype, run after /brainstorm and before GDDs. New HTML / Engine / Paper paths. New --spike mode for ad-hoc mid-project technical questions.
- /vertical-slice = brand-new pre-production gate. Production-quality end-to-end build that validates the full game loop before Production commits.
- prototyper agent rewritten to operate in both modes.
- Two new templates (prototype-report.md, vertical-slice-report.md) and two new coordination patterns (8, 8b).
Pre-production workflow restructured (workflow-catalog.yaml)
- New entity-inventory step before asset/UX authoring.
- UX specs minimum raised to 3 (main menu + HUD + pause).
- prototype step now optional in pre-production (it runs earlier).
- vertical-slice step rewired to the new skill.
- Solo-dev skip language throughout.
Polish
- AskUserQuestion widgets added to 7 skills that were missing them.
- --review full|lean|solo flag now visible in argument-hints for all team-* skills.
- Skill consistency audit pass — prompts, flow, naming.
- ~17 skill frontmatters got their missing model: sonnet field.
- README platform-support copy made honest about Windows-first testing.
- Windows PowerShell fallback in /smoke-check.
Bug fixes
- #21 log-agent hooks logged "unknown" agent_type
- #36 missing allowed-tools in /architecture-decision and /story-done
- #42 rg --type gdscript is invalid — now uses --glob *.gd
- #43 session-start preview showed oldest state instead of newest
- #45 duplicate ## 0. heading and broken step numbering in /architecture-decision
Project docs
- New CONTRIBUTING.md — contributor guidelines for the framework itself
- New SECURITY.md — coordinated disclosure policy
- UPGRADING.md — new v1.0.0-beta → v1.0 section
---
Full diff: https://github.com/Donchitos/Claude-Code-Game-Studios/pull/50
v1.0.0-beta — Claude Code Game Studios预发布2026-04-07
> This is a beta release. The system is functional and has been through a full QA pass, but some rough edges remain. We welcome bug reports, feedback, and contributions. Expect things to evolve rapidly before v1.0 stable.
Claude Code Game Studios is a game development studio framework for Claude Code — 48 coordinated subagents and 70+ workflow skills covering every phase of indie game development, from concept through live ops.
---
What's New Since v0.3.0
This release encompasses 25 commits of development organized across new skills, agents, infrastructure, and a full QA pass.
---
New Skills (33 added, 1 removed)
Implementation pipeline:
- /dev-story — reads a story, loads full context (TR registry, ADR, control manifest, engine prefs), routes to the correct programmer agent via Task, drives implementation through to test
- /create-epics — architect-level epic scoping, one epic per module
- /create-stories — decomposes a single epic into story files with TR-ID references and test evidence paths
- /quick-design — lightweight spec path for small changes; bypasses the full GDD pipeline
Story lifecycle:
- /story-readiness — validates a story is fully implementation-ready before a developer picks it up (ADR status, TR-ID validity, asset references, manifest version)
- /story-done — end-of-story completion review: verifies acceptance criteria, checks test-criterion traceability (BLOCKING if >50% of ACs have no test), detects deviations
Design & architecture:
- /ux-design — guided section-by-section UX spec authoring covering screens, flows, HUDs, and interaction patterns
- /ux-review — UX spec validation with APPROVED / NEEDS REVISION / MAJOR REVISION verdict
- /architecture-review — holistic architecture doc validation (runs on Opus)
- /create-architecture — guided section-by-section architecture authoring
- /create-control-manifest — generates the system control manifest with version stamp
- /propagate-design-change — when a GDD is revised, propagates impact analysis across all dependent systems
- /review-all-gdds — cross-GDD holistic consistency + design theory review (runs on Opus); includes Phase 4 Cross-System Scenario Walkthrough
- /content-audit — GDD-specified content vs implemented content gap report (COMPLETE / IN PROGRESS / EARLY / NOT STARTED per system)
- /art-bible — guided art bible authoring
- /asset-spec — per-asset visual specification generation
QA & testing (10 new):
- /test-setup — scaffold test framework + CI configuration per engine (Godot/Unity/Unreal)
- /qa-plan — generate a sprint test plan classifying every story by type (Logic / Integration / Visual / UI / Config)
- /smoke-check — critical path gate (PASS / PASS WITH WARNINGS / FAIL) before QA hand-off
- /team-qa — orchestrate qa-lead + qa-tester through the full QA cycle
- /bug-triage — severity/priority classification, sprint assignment, trend analysis
- /regression-suite — GDD critical path coverage map + fixed-bug regression audit
- /test-evidence-review — quality review of automated tests and manual evidence
- /test-helpers — generate engine-specific helper library boilerplate
- /soak-test — extended play session protocol with timed checkpoints and memory monitoring
- /test-flakiness — CI log analysis, flaky test detection, quarantine recommendations
Sprint & status:
- /sprint-status — fast 30-line sprint snapshot, read-only (runs on Haiku)
Production & release:
- /day-one-patch — day-one patch preparation workflow
- /security-audit — security audit for shipped game builds
- /team-live-ops — post-launch content planning orchestration
Tooling:
- /help — context-aware what's-next skill driven by workflow-catalog.yaml
- /adopt — brownfield onboarding for existing projects not started with CCGS
- /skill-test — three-mode skill QA tool: static linter / spec behavioral assertions / audit coverage report
- /skill-improve — test-fix-retest loop covering static + behavioral checks
Deprecated and removed:
- /create-epics-stories split into /create-epics + /create-stories and fully removed; all references updated
---
New Agent: godot-csharp-specialist
Covers the full Godot C# surface: partial class requirement, [Export]/[Signal] attributes, SignalName emit pattern, ToSignal async, Godot.Collections vs .NET collections, [GlobalClass] resources, .csproj configuration, and C#-specific anti-patterns.
/setup-engine now asks GDScript vs C# vs Both when Godot is chosen — the choice determines the CLAUDE.md template, naming conventions, and specialist routing table.
---
Director Gates System
New .claude/docs/director-gates.md — a central registry of 18 named gate prompts (CD-, TD-, PR-, LP-, QL-, ND-, AD-*) covering all 7 production stages. Skills reference gate IDs instead of embedding inline prompts, eliminating drift when prompts need updating. 15 skills updated to use gate IDs.
Gate intensity modes (--review): Three-level review control added to all 13 gate-using skills and globally configurable via production/review-mode.txt:
- full — all director gates fire (default)
- lean — per-skill gates skipped, phase gates retained
- solo — all director gates skipped (solo developers)
---
Traceability Backbone
- New docs/architecture/tr-registry.yaml — persistent stable TR-IDs per GDD requirement; never renumbered
- Stories embed TR-ID references (not quoted text) — survives GDD edits without breaking links
- ADR status gate: stories referencing Proposed ADRs are automatically set Status: Blocked
- /story-done and /story-readiness both perform TR-ID registry lookups and manifest staleness checks at runtime
---
Agent Upgrades
- memory: project added to 14 specialist agents for cross-session learning: art-director, audio-director, economy-designer, game-designer, lead-programmer, level-designer, localization-lead, narrative-director, performance-analyst, qa-lead, systems-designer, ux-designer, world-builder, writer
- context: fork added to 6 analysis skills to preserve the main context window: review-all-gdds, architecture-review, code-review, design-review, asset-audit, reverse-document
- isolation: worktree on the prototyper agent — throwaway prototypes never touch the main working tree
- 12 agents fixed in the QA pass: all 4 directors (gate verdict token format), engine-programmer, ui-programmer, tools-programmer, technical-artist, gameplay-programmer, godot-gdextension-specialist, systems-designer, accessibility-specialist, live-ops-designer, qa-tester
---
Model Tier Routing
Formalized across all skills:
| Tier | Skills |
|---|---|
| Haiku (read-only/format) | /help, /sprint-status, /story-readiness, /scope-check, /project-stage-detect, /changelog, /patch-notes, /onboard |
| Opus (multi-doc synthesis) | /review-all-gdds, /architecture-review, /gate-check |
| Sonnet | Everything else |
---
Hooks & Infrastructure
- log-agent-stop.sh — SubagentStop event completes the agent audit trail (start + stop)
- post-compact.sh — fires after every compaction, reminding Claude to re-read production/session-state/active.md
- notify.sh — Windows toast notifications (PowerShell balloon tip) on Notification events
- validate-skill-change.sh — advisory reminder to lint after any skill file is edited
- Directory CLAUDE.md scaffolding in docs/, design/, and src/
- L0 summary scan phase on heavy analysis skills (reduces token cost on large projects)
- Session Extract auto-appended to active.md after review-all-gdds, architecture-review, and story-done
---
CCGS Skill Testing Framework
Self-contained QA layer for the framework itself:
- 72 skill specs + 49 agent specs
- catalog.yaml machine-readable coverage index
- quality-rubric.md evaluation criteria
- Shared fixtures and templates
v0.3.0 — /design-system, /map-systems, Status Line, UPGRADING Guide2026-03-09
What's New
New Skills
- /design-system — Guided, section-by-section GDD authoring. Walks through all 8 required sections collaboratively, writes each to file immediately on approval, and survives session crashes via file-backed state.
- /map-systems — Renamed from /design-systems for clarity. Decomposes a game concept into a full systems index with dependency mapping, priority tiers, and design order.
Custom Status Line
- statusline.sh displays a 7-stage production pipeline breadcrumb in the terminal footer
- Auto-detects project stage from artifacts, or reads from production/stage.txt
- In Production/Polish/Release stages, shows Epic > Feature > Task breadcrumb from active.md
UPGRADING.md
- New migration guide covering both v0.1→v0.2 and v0.2→v0.3
- Documents breaking rename (/design-systems → /map-systems), new files, and safe-to-overwrite lists
Bug Fixes
- log-agent.sh and validate-commit.sh hook execution fixed
- Cross-reference gaps repaired across skills
- Stale workflow chain links updated
Breaking Changes
/design-systems has been renamed to /map-systems. Update any saved notes or scripts that invoke the old name.
Upgrading
See UPGRADING.md for step-by-step migration instructions.
v0.2.0 — Context Resilience & Systems Decomposition2026-02-20
What's New
Context Resilience (Phase 8)
- File-backed session state at production/session-state/active.md — work survives context compaction and session crashes
- Incremental document writing — agents write sections to disk as they're approved, not all at once at the end
- Auto-recovery hooks — session-start.sh detects and previews previous session state; pre-compact.sh dumps context before compaction
AskUserQuestion Integration
- Explicit AskUserQuestion tool usage added to 26 files (14 agents, 8 skills, 3 protocol templates, 1 principle doc)
- Enforces the Explain → Capture pattern: full analysis in conversation, then structured UI picker for decisions
- Design/creative agents use it at every decision point; team skills at phase transitions
/design-systems Skill
- Bridges game concept → per-system GDDs with professional studio methodology
- New systems-index.md template: enumeration, categories, priority tiers, dependency map, design order
- 3 modes: map (full decomposition), [system-name] (design specific), next (highest-priority)
- Each completed GDD passes /design-review before the next system starts
- Pre-production → Production gate now requires systems-index.md
Cross-Platform & Permission Fixes
- Fixed Windows compatibility bugs (no grep -P usage)
- Added missing tool permissions to agent definitions
- Updated docs for v0.2.0
Full Changelog
https://github.com/Donchitos/Claude-Code-Game-Studios/compare/v0.1.0...v0.2.0