game-plan
Use when starting a Godot game or a game jam entry, or when asked about its phases and gates, scope, milestones, risks, budget or business model.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-13
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- game-plan
- description
- Use when starting a Godot game or a game jam entry, or when asked about its phases and gates, scope, milestones, risks, budget or business model.
03 SKILL.md 正文
A-to-Z game plan (solo developer)
plan/plan.json holds seven phases, each with gates. plan shows the current phase, what blocks it,
and warns on timebox overruns. Genre-independent.
pwsh ${CLAUDE_PLUGIN_ROOT}/ci/godot/hyperdot.ps1 plan -Do init -Game "My Game" # once; never overwrites
pwsh ${CLAUDE_PLUGIN_ROOT}/ci/godot/hyperdot.ps1 plan -Json # where am I, what is next
pwsh ${CLAUDE_PLUGIN_ROOT}/ci/godot/hyperdot.ps1 plan -Do done -Id concept.scope # mark a manual gate (you or the user decides, not the agent)
pwsh ${CLAUDE_PLUGIN_ROOT}/ci/godot/hyperdot.ps1 plan -Do reopen -Id concept.scope # undo it, never by editing plan.json
| Phase | Weeks | Exit gate in one line |
|---|---|---|
| Concept | 2 | Pitch, pillars, scope ladder, kill criteria written |
| Prototype | 4 | Trailer script written; a stranger played 15 min unprompted; verdict PROCEED/PIVOT/KILL |
| Vertical slice | 8 | Menu -> level -> save -> quit, tests in CI, export -Run passes, tech debt listed, frame budget met at full density |
| Production | 24 | Playable milestones every 2-6 weeks; feature freeze |
| Alpha and beta | 8 | Feature-complete, playtested by strangers, perf and accessibility checked |
| Release candidate and launch | 4 | Store page, demo, CI-built gold master, release |
| Post-launch | 8 | Hotfix path, feedback triaged, retrospective |
Timeboxes are defaults for ~10 hours a week; set weeklyHours and resize at the slice gate with real velocity.
How an agent uses it
- Start every session with
plan -Json. Work only onnextActionsofcurrentPhase. checkismanualor automatic (tests: a test framework runner,ci,preset,file:<path>: a file that is not empty,runner:<name>: that runner is set up,placeholders: none left). Automatic gates turn true when the project makes them true; do the work, not the checkbox.plan -Do donerefuses automatic gates. Anoptionalgate does not hold its phase back.- Manual gates (the playtest happened, the verdict) need the user's word. Ask; do not mark them yourself.
- When
warningsreports an overrun: propose cutting scope from the "could" tier; never cut testing or the gate. - Work items are
tasksin plan.json, edited with commands, not by hand:
Columns:pwsh ${CLAUDE_PLUGIN_ROOT}/ci/godot/hyperdot.ps1 plan -Do add -Title "Paddle feel pass" -Feature F1 -Column todo # -Feature: the GDD feature it delivers pwsh ${CLAUDE_PLUGIN_ROOT}/ci/godot/hyperdot.ps1 plan -Do move -Id t1 -Column doing # refuses past wipLimit (default 2) pwsh ${CLAUDE_PLUGIN_ROOT}/ci/godot/hyperdot.ps1 plan -Do remove -Id t1 # a task that is no longer wantedbacklog, design, todo, doing, verify, done: the solo version of a Kanban work loop, with Verify meaning "tests green" (verify-before-done), not "a person reviewed it".-Columnalso takes the project column's name ("To Do"). - The GitHub Project is optional and experimental (tested against a stand-in
gh); plan.json stays the source of truth.plan -Do sync -Board @memakes a private project for the user (titled "<game> plan", linked to the game's repository, with a Column field: Backlog, Design, To Do, Doing, Testing, Done);plan -Do syncafter that creates an item per new task, pulls back items the user dragged, adds a task for each item the user added, archives the items of removed tasks and moves the items of tasks moved withplan -Do move. It uses the GitHub CLI signed in with the project scope (gh auth refresh -s project); nothing is stored in a file. The Stop hook of the plugin repeats the sync after a turn that changed the plan. status -Jsonis the whole picture in one call: GDD features with their tagged tests, the plan, latest runs, git activity. The user asks for/gameto see it judged and written toplan/STATUS.md.
What 20 postmortems say went wrong most (${CLAUDE_PLUGIN_ROOT}/docs/studio-lessons.md): testing too thin or too late (9), scope
creep (8), effort underestimated (8), no vertical slice (7), platform surprises (7). What went right most: outside playtests early
and often (9), pillars with cuts to protect them (9), custom tools (6). Hence the gates concept.decisions (multiplayer, saves,
localisation, accessibility and the camera decided up front, not bolted on), prototype.variants, production.playthrough (a bot reaches
the ending on every CI run) and launch.archive (ship exactly the build that was tested, with its source tagged). Keep the
first estimate of each phase, and report how far the actual time has slipped from it.
Croteam shipped earlier games as "the alpha plus bug fixes"; The Talos Principle changed a lot after its alpha, through
team, closed-beta and public-demo playtests. Its biggest failure was a localisation nobody had played. Hence beta.iterate
(redesign what playtesters struggled with, then play it again) and beta.localization (a native speaker who has not played
finishes each language). In a story game, production's draft stage ends when every quest is playable start to end with
placeholder dialogue (CD Projekt Red); any later stage may send a quest back.
Volition (Saints Row, Red Faction) learned that a slice must measure, not only impress. Saints Row's slice streamed two
city chunks and never tried the places where three met; with its tech debt never counted, the game ran at 10 to 15 fps three
weeks before launch. Saints Row: The Third's art slice polished single assets instead of a full scene, so performance was
fixed in production. Hence slice.debt (automatic: docs/TECH-DEBT.md, from ${CLAUDE_PLUGIN_ROOT}/ci/godot/templates/TECH-DEBT.md,
lists each hacked, stubbed or unstarted system with its cost at full size, and the total the user accepts), slice.budget
(manual: hyperdot.ps1 bench -Scene <busiest scene> -BudgetMs <n> passes with that scene filled as the finished game will be), and a
slice.verdict that rejects a feature list or a tech demo and takes more slice time while core gameplay is still moving.
A plan made before these gates existed lacks them (plan.json is the game's after init): offer to add them.
Valve restarted Half-Life when it was not fun weeks before its date ("we couldn’t see how a month or two would make any
significant difference"): a verdict is proceed, pivot or kill, never "extend". It then worked in 15-minute rough slices
with a playtest each week setting next week's work (game-playtest); designed only around technology that already
worked, with risky bets first; banned one-off engineering (a feature is done when two scenes use it); froze scenes that
had passed playtests; and ran an alpha pass over the whole game that added nothing new. Its Cabal put every discipline
in the design room: with agents, one review pass per discipline over each design, each listing the work it creates
(${CLAUDE_PLUGIN_ROOT}/docs/valve.md).
Derek Lieu, a trailer editor, writes the trailer before the game: if nobody can picture a trailer worth watching, the
prototype is probably not worth building on (Lucas Pope's test). Hence prototype.trailer (automatic: docs/TRAILER.md,
from ${CLAUDE_PLUGIN_ROOT}/ci/godot/templates/TRAILER.md): the genre in the first 5 seconds, then the hook, content last;
three standout moments; each core mechanic with the on-screen feedback that shows a viewer the player caused it. An empty
feedback cell is design work for the prototype (game-juice); a scene the trailer needs and the game lacks becomes a task.
Write it with the user, from their answers: it is a draft to argue over, and the launch trailer starts from it.
Hellgate: London's lead said its biggest challenge was "realistically limiting ourselves": a new idea moves something
out of the milestone or goes to the GDD's "Out of scope" section, parked; an optional mode gets a quick prototype
and stays only if it keeps being played. A mod team called its own slipped dates "hasty and unplanned": promise a playable demo,
not a date, until the slice gate passes. Story, lore and voice planning (voice recorded after content lock) are
narrative; rewards and update cadence are progression.
Grow the commitment in steps, each earned by reactions to the last: Google's 8-bit Maps went from a demo made in hours to
a fifth of its maker's time, then full time. A decision made in a call or a chat exists for agents only once it is in
docs/decisions/ or plan.json, and a dropped idea is parked in the GDD's "Out of scope", not on one person's machine:
ideas come back years later (Untold History, ${CLAUDE_PLUGIN_ROOT}/docs/studio-lessons.md).
Concept: risks, money and the business model
- Risk log (
docs/RISKS.md, gateconcept.risks): the top five risks, each with likelihood, impact, how it is retired and by when. Every system the developer has never built counts as one; the riskiest is prototyped first. - Comparable games (
concept.market): a teardown of each of the three, as ingame-designsection 4. - Business model: name it (premium, demo plus full game, free with ads, free to play) and what is sold. Never sell
an advantage over another player, or a shortcut past what a learning game teaches; no energy timers or roulettes.
Paid random items, timed offers and daily rewards raise the age rating (
legalsection 1). - Budget and runway (
docs/BUDGET.md, the optional gateconcept.budget; skip it for a hobby game): the categories, filled in by the user, never guessed. Costs: people, outsourcing (art, music, trailer), workspace (rent, power, internet, insurance), monthly software and services, per-project licences, events (date, cost, go or no-go), marketing (one-off and monthly), age ratings per region, taxes. Income: cash on hand, store payouts and when they arrive, funding (savings, a publisher, platform funds, public cultural grants, crowdfunding), tax credits. Runway is cash divided by monthly cost; it should cover the plan's remaining weeks. Net per sale: price, less the average discount, sales tax or VAT, the store's fee, a publisher's share and refunds, each taken from the current terms. A grant application can reuse the GDD, plan.json and this budget.
Alpha and beta: exit criteria
- Placeholders are listed in
docs/TECH-DEBT.md(its Placeholders table): every stand-in asset or text, labelled in the game, with what replaces it and by which milestone; a stand-in asset is also"placeholder": trueinassets.json.beta.placeholdersis met when both are empty. - Bug classes: A blocks finishing the game, crashes or loses data; B is major and seen by most players. Alpha and
beta both play to the end with no undeclared A or B bug (
game-playtesttriage uses the same classes). beta.alphacomes with a statement of completion indocs/MILESTONES.md: each missing feature and the build it lands in. A long list means it is not alpha yet.- Beta: placeholders are a small part of the content (one studio template allows at most a fifth); any analytics
tested end to end (
game-telemetry); a release plan with dates for builds, marketing beats, store review and the first patches. The marketing cadence and the release ladder (a free portal, a soft launch, then the main store) are instore-page.
Jam mode (2 to 9 days)
Start with plan -Do init -Profile jam: phases sized in days, for a 7-day jam (scale timeboxDays to yours), with
gates from the day table of ${CLAUDE_PLUGIN_ROOT}/ci/godot/templates/GDD-jam.md, which becomes docs/GDD.md (jam.gdd).
Ask "browser build?" before anything else: Godot 4.7 stops a C# Web export ("Exporting to Web is currently not supported
in Godot 4 when using C#/.NET", mobile-web), so a browser jam is made in GDScript. Day 1 proves the riskiest feel
with plain shapes; day 2 has the loop with win and lose; the day before the end exports to the jam's platform and
launches it (export -Preset Web -Run for the browser); the last day is fixes only, and the game goes in with an hour or
more to spare. A feature you are unsure of is cut. Small finished games (jams, contests) before the big one build the
network the big one needs: Enix found Yuji Horii through its game contests.
Rules of thumb that drive the gates: scope is cut, never quality or the date; a prototype that nobody asks to replay is
killed or pivoted, not polished; the store page goes up months before launch because wishlists compound; every
milestone is a build someone can play. Sources and reasoning: docs/plan.md.
04 同仓库更多 Skills
over-budget
Use when a fixture exceeds the size budget.
bad-yaml
暂无简介
bad-workflow-description
Use when testing; runs the kit and then reads the JSON.
bad-version
Use when a fixture quotes a stale package version.
bad-value
Use when a fixture passes a value outside a parameter's set.
bad-trigger
Helps with Godot tests.