Skip to content

Turn Claude Code into a full game dev studio — 49 AI agents, 72 workflow skills, and a complete coordination system mirroring real studio hierarchy.

15.9 KB
Raw Blame

Skill Test Spec: /prototype

Skill Summary

/prototype [concept] [--path html|engine|paper] [--review full|lean|solo] [--spike] runs a throwaway concept prototype before GDDs are written. It first confirms intent with AskUserQuestion (Prototype this concept / Skip — concept already proven / Mid-production spike), forms a falsifiable hypothesis ("If the player [does X], they will feel [Y] — we will know this is true if [Z]") and asks for the riskiest assumption. It picks an HTML, Engine or Paper path (genre table, recommendation pre-stated), presents a 3–5 bullet plan, asks to record it as a checkpoint in production/session-state/active.md, then asks "May I create the prototype directory at prototypes/[concept-name]-concept/ and begin implementation?". Every prototype file starts with a PROTOTYPE - NOT FOR PRODUCTION header in that file's own comment syntax (# in GDScript and Python, // in C#, C++ and JavaScript, <!-- --> in HTML and Markdown), and standards are intentionally relaxed.

After a one-question-at-a-time debrief, it fills .claude/docs/templates/prototype-report.md, asks to write prototypes/[concept-name]-concept/REPORT.md and add its row to prototypes/index.md (one ask naming both). Verdicts are PROCEED / PIVOT / KILL. PIVOT writes a PIVOT-NOTE.md; KILL runs a five-item soundness check and appends to prototypes/GRAVEYARD.md. In full review mode creative-director reviews the result via CD-PLAYTEST and answers APPROVE / CONCERNS / REJECT: APPROVE keeps the recommendation; CONCERNS go to the user, who decides (Revise / Accept / Discuss); REJECT means a PROCEED cannot stand — the user chooses PIVOT or KILL, and REPORT.md records the director's reason. A NOT ASSESSED answer (the director lacked an input) is never an approval. --spike replaces the whole flow with a ~4-hour spike that asks before creating its folder, asks before writing SPIKE-NOTE.md and clearing the spike from active.md, and has no CD gate and no PROCEED/PIVOT/KILL.


Static Assertions (Structural)

Checked against the SKILL.md by /skill-test spec — no fixture needed.

  • Has required frontmatter fields: name, description, argument-hint, user-invocable, allowed-tools
  • Has ≥2 phase headings
  • Contains verdict keywords: PROCEED, PIVOT, KILL, NOT ASSESSED (a prototype nobody has played)
  • Contains "May I create" / "May I write" language before creating the prototype directory and the report
  • Has a next-step handoff per verdict (PROCEED → /design-review, /gate-check, /map-systems, /design-system; PIVOT → /prototype or /brainstorm; KILL → /brainstorm)

Director Gate Checks

  • Full mode: CD-PLAYTEST — creative-director receives the REPORT.md content, the hypothesis and the pillars (from design/gdd/game-concept.md, or the pitch and "what they feel" line from design/game-brief.md) and returns APPROVE / CONCERNS / REJECT, or NOT ASSESSED when it lacked an input. APPROVE → the recommendation stands; CONCERNS → shown with the recommendation, and the user decides (Revise / Accept / Discuss); REJECT → a PROCEED cannot stand: the user chooses PIVOT or KILL, and REPORT.md records the director's reason; NOT ASSESSED → never an approval: the missing input is named, and the gate is re-run or the review recorded as NOT ASSESSED. If neither document exists: "CD-PLAYTEST skipped — game pillars not yet defined at concept prototype stage."
  • Lean mode: skipped — "CD-PLAYTEST skipped — Lean mode."
  • Solo mode: skipped — "CD-PLAYTEST skipped — Solo mode."
  • Spike mode: no CD gate and no phase gate, in any review mode.

Fixture Defaults

Every case's project.yaml sets modes.rigor: minimal and modes.automation: collaborative, with no project.local.yaml, unless its own Fixture says otherwise. Rules: studio-tests/CLAUDE.md § Fixture defaults.


Test Cases

Case 1: Happy Path — Engine Prototype, PROCEED

Fixture:

  • design/gdd/game-concept.md describes a 2D platformer; project.yaml has engine.name: Godot and engine.language: GDScript
  • prototypes/ has no grapple-hook-concept/
  • Review mode: solo; modes.workflow: standard (pinned: unset resolves to minimal, which routes PROCEED to /create-stories and /dev-story instead of the step 7 list — Case 6 covers that route)
  • In the debrief the user reports the hypothesis CONFIRMED and answers PROCEED

Input: /prototype grapple-hook traversal

Expected behavior:

  1. AskUserQuestion confirms intent; user picks "Prototype this concept"
  2. Skill states a falsifiable hypothesis in the If / feel / know-if form and asks for the riskiest assumption
  3. Path widget pre-states an Engine recommendation (traversal feel is the hypothesis)
  4. Plan is presented and confirmed, with the ask naming production/session-state/active.md; the checkpoint is written there
  5. Skill asks "May I create the prototype directory at prototypes/grapple-hook-concept/ and begin implementation?"; every GDScript file it writes starts with # PROTOTYPE - NOT FOR PRODUCTION (GDScript comments use #; a // header would not parse)
  6. Debrief questions are asked one at a time; report drafted from .claude/docs/templates/prototype-report.md; "May I write this report to prototypes/grapple-hook-concept/REPORT.md and add its row to prototypes/index.md?"; a row is appended to prototypes/index.md
  7. Note "CD-PLAYTEST skipped — Solo mode."; PROCEED routes to /design-review, /gate-check, /art-bible, /map-systems, /design-system, /review-all-gdds

Assertions:

  • The hypothesis is falsifiable, with a measurable signal
  • Engine is recommended in the path prompt before the user chooses
  • The directory ask names prototypes/grapple-hook-concept/ and comes before any prototype file is written
  • Every prototype script carries the PROTOTYPE - NOT FOR PRODUCTION header as a # comment, never //; no file is written under the code root
  • REPORT.md is written only after its ask, that ask also names prototypes/index.md, and the index gains a row with verdict PROCEED
  • PROCEED next steps include /map-systems and /design-system

Case 2: PIVOT — Carry-Forward Note Written

Fixture:

  • A paper prototype of trade-routes has just been debriefed
  • User answers PIVOT: the route planning was engaging, the upkeep maths was not
  • Review mode: solo

Input: /prototype trade-routes --path paper

Expected behavior:

  1. Paper path writes prototypes/trade-routes-concept/rules.md and play-log.md after the directory ask
  2. REPORT.md is written after its ask with verdict PIVOT
  3. Skill asks, one at a time: "What specifically worked in this prototype that we should preserve…?" and "What is the single most important thing to change?"
  4. Skill asks "May I write this to prototypes/trade-routes-concept/PIVOT-NOTE.md?"
  5. The note holds the original hypothesis, what to keep, what to change and the revised hypothesis; next steps are /prototype [revised-concept] or /brainstorm [hint]

Assertions:

  • --path paper is honoured and produces rules.md and play-log.md
  • The two carry-forward questions are asked separately
  • "May I write" names PIVOT-NOTE.md before it is written
  • The note contains a revised hypothesis
  • Next steps offer /prototype for the revised concept or /brainstorm

Case 3: KILL — Soundness Check and Graveyard Entry

Fixture:

  • An HTML prototype of procedural-dialogue has been debriefed
  • User answers KILL; the debrief shows testers never understood the core action after 2 playtests, no fun moment was observed, and the concept only worked when the developer explained it

Input: /prototype procedural-dialogue --path html

Expected behavior:

  1. REPORT.md is written after its ask with verdict KILL
  2. Skill runs the five-item KILL checklist; 3 items apply, so the verdict is sound
  3. Skill asks "May I append this concept to prototypes/GRAVEYARD.md?"
  4. The entry has Kill reason, What worked, What failed and Next time, with a specific kill reason (not "it was boring")
  5. Next step: /brainstorm open or /brainstorm [new-hint]

Assertions:

  • The KILL checklist is applied, with the 2+ rule deciding soundness
  • "May I append" names prototypes/GRAVEYARD.md before writing
  • The graveyard entry has all four fields and a specific kill reason
  • Next steps route to /brainstorm

Case 4: Spike Mode — No Debrief, No Gate, No PROCEED/PIVOT/KILL

Fixture:

  • Project is in Production
  • Review mode: full; modes.workflow: standard (pinned: unset resolves to minimal, whose spike note and active.md wording Case 7 covers)
  • production/sprints/sprint-004.md is the current sprint and active.md shows it

Input: /prototype rope-physics --spike

Expected behavior:

  1. Skill skips Phases 1–9 and runs Spike Mode
  2. Asks for the spike question in plain text: "Can we [do X] using [approach Y]?"
  3. Path widget as in Phase 3; scope limited to 2–3 bullets; ~4-hour cap
  4. Asks "May I create prototypes/rope-physics-spike-[date]/ and build the spike there?" before writing any spike file
  5. Asks "Did the spike answer the question? YES or NO, and why in one sentence."
  6. Asks once, naming both files, "May I write prototypes/rope-physics-spike-[date]/SPIKE-NOTE.md and clear the spike from production/session-state/active.md?"; then writes the note (question, result YES / NO / PARTIAL, what to do next) and updates active.md
  7. The note's "what to do next" offers: add to the current sprint / investigate further / abandon the idea. active.md is cleared back to the current sprint state

Assertions:

  • The intent widget and hypothesis/debrief phases are skipped
  • The spike folder is created only after its ask
  • SPIKE-NOTE.md and active.md are written only after an ask that names both
  • Output is SPIKE-NOTE.md in a -spike-[date] folder, not a REPORT.md
  • No CD-PLAYTEST spawn or skip note, even though review mode is full
  • No PROCEED / PIVOT / KILL verdict is issued
  • At standard, the note's next-step options say "add to current sprint", and active.md returns to the current sprint state

Case 5: Director Gate — CD-PLAYTEST by Review Mode and Pillars

Fixture:

  • A concept prototype has been built and REPORT.md written with a PROCEED recommendation
  • Run A: --review full, design/gdd/game-concept.md exists with pillars; creative-director returns REJECT — the core fantasy is not present; the user, asked PIVOT or KILL, picks PIVOT
  • Run B: --review full, neither game-concept.md nor design/game-brief.md exists
  • Run C: --review lean

Input: /prototype [concept] --review [mode]

Expected behavior:

  1. Run A: creative-director is spawned for CD-PLAYTEST with the REPORT.md content, hypothesis and pillars. Its REJECT means the PROCEED recommendation cannot stand: the skill asks the user to choose PIVOT or KILL (it does not pick for them), the user picks PIVOT, and after an ask naming REPORT.md and prototypes/index.md the report records the director's verdict and reason and the final PIVOT recommendation; Phase 9 then runs the PIVOT route
  2. Run B: no spawn; note "CD-PLAYTEST skipped — game pillars not yet defined at concept prototype stage."
  3. Run C: no spawn; note "CD-PLAYTEST skipped — Lean mode."

Assertions:

  • Run A spawns creative-director for CD-PLAYTEST after REPORT.md is written
  • Run A: the director's verdict is read as one of APPROVE / CONCERNS / REJECT, not as a PROCEED / PIVOT / KILL of its own
  • Run A: after the REJECT, PROCEED is not kept; the user is asked to choose PIVOT or KILL
  • Run A's final recommendation is the user's PIVOT, and REPORT.md records it with the director's reason
  • Run B skips with the "game pillars not yet defined" note instead of spawning without pillars
  • Run C skips with the Lean-mode note
  • No other director gate is spawned

Case 6: NOT ASSESSED (Unplayed) and the workflow: minimal PROCEED Route

Fixture:

  • Run A: the prototype was built this session but nobody has played it yet
  • Run B: a PROCEED recommendation, project.yaml has modes.workflow: minimal

Input: /prototype [concept]

Expected behavior:

  1. Run A: with no fun evidence yet, PROCEED/PIVOT/KILL are all unreachable; the skill reports NOT ASSESSED and stops rather than guessing, naming "play it, then run /prototype [same concept] again" as the next step
  2. Run B: Phase 9's PROCEED route offers /create-stories (from the brief) then /dev-story on the first story, not the GDD//map-systems//art-bible path

Assertions:

  • Run A's verdict is NOT ASSESSED, not a guessed PROCEED/PIVOT/KILL
  • Run A's report and prototypes/index.md row both read NOT ASSESSED until a playtest happens
  • Run B's recommended path is the workflow: minimal route (/create-stories, /dev-story), not the standard/full GDD pipeline

Case 7: Spike Mode at workflow: minimal — the build order, not a sprint

Fixture:

  • project.yaml: modes.workflow: minimal (pinned: this case tests the minimal wording, the opposite of Case 4); review mode full
  • No production/sprints/ directory. design/game-brief.md lists a build order; the story in progress is production/epics/core/story-002.md (Status: In Progress), and production/session-state/active.md names it as the current task
  • The user answers YES to the spike question

Input: /prototype rope-physics --spike

Expected behavior:

  1. The spike runs as in Case 4, with the same asks: the spike folder, then one ask naming SPIKE-NOTE.md and active.md
  2. SPIKE-NOTE.md holds the question tested, the result (YES / NO / PARTIAL) and what to do next: add the mechanic to the brief's build order / investigate further / abandon the idea. It does not offer "add to current sprint"
  3. active.md is updated to clear the spike and return to the story in progress (story-002), not to a sprint state
  4. No CD gate and no PROCEED / PIVOT / KILL

Assertions:

  • The note's next-step options name the brief's build order instead of the current sprint
  • After the write, active.md names story-002 as the current task and no sprint
  • The spike folder and the note-and-active.md write each follow an ask that names the files
  • No sprint file is created or referenced
  • No CD-PLAYTEST spawn or skip note and no PROCEED / PIVOT / KILL verdict is issued

Protocol Compliance

  • Confirms intent with AskUserQuestion before building (unless --spike)
  • Forms a falsifiable hypothesis and stops if the concept is too vague to form one
  • Presents the plan and gets confirmation before building
  • Asks before creating prototypes/[concept-name]-concept/ and before writing REPORT.md (with its prototypes/index.md row), PIVOT-NOTE.md or GRAVEYARD.md
  • In spike mode, asks before creating the spike folder and before writing SPIKE-NOTE.md and active.md
  • Keeps prototype files under prototypes/, each with the PROTOTYPE - NOT FOR PRODUCTION header in its own comment syntax
  • Applies the review mode before CD-PLAYTEST and prints the skip note when skipped
  • Ends with the per-verdict next steps

Coverage Notes

  • "Skip — concept already proven" (records the evidence, suggests /map-systems or /design-system, and stops) is not tested.
  • A concept too vague to form a hypothesis (skill stops and asks the user to narrow it) is covered only by protocol compliance.
  • Reading a prior PIVOT-NOTE.md on the next run to seed the revised hypothesis is not tested.
  • The engine-path multi-turn fix loop and the 2-hour sunk-cost rule need a live run to verify.
  • CD-PLAYTEST APPROVE (the recommendation stands), CONCERNS (the user decides: Revise / Accept / Discuss), REJECT on a PIVOT or KILL recommendation (it stands) and NOT ASSESSED (never an approval) are not separately tested.