Imported from RandallLiuXin/GodotMaker (skills/core/gm-gdd/SKILL.md). Install upstream with npx skills add RandallLiuXin/GodotMaker --skill gm-gdd. Copyright stays with the author.

GodotMaker GDD

$ARGUMENTS

You are running the design phase for one tag at a time. The pipeline is tag-iterative: each /gm-gdd invocation either bootstraps the whole project plus its first tag (initial mode), or focuses the next tag in ROADMAP.md (subsequent mode).

Session Setup

FIRST ACTION — before anything else: Write gdd to .godotmaker/current_role.

Resume Check

Read .godotmaker/stage.jsonl (treat as empty if missing) — each line is {"role": X, "ts": Y}.

• If project.godot does not exist → STOP. Tell user to run /gm-scaffold first.
• If the last event has role == "gdd" → STOP. Tell the user:

  "GDD already completed for the current tag at {timestamp}. Recommended next: /gm-asset. If you need to redo this step or have other plans, just tell me."

• Otherwise → proceed (fresh project, OR new tag after the previous tag's /gm-finalize).

Mode Detection

Detect the mode by inspecting on-disk state — there is no flag:

• Initial mode: ROADMAP.md does NOT exist. (GDD.md may also be missing — if it is, this is a brand-new project.)
• Subsequent mode: ROADMAP.md EXISTS. Determine the current tag as follows:
  1. Read ROADMAP.md, list tag entries in declared order.
  2. Run git tag --list 'v*' (capture stdout).
  3. The current tag is the earliest tag in ROADMAP that is not in git tag --list.
  4. If every ROADMAP entry already has a git tag, STOP and inform the user the roadmap is exhausted — they must edit ROADMAP.md to add new entries before re-running /gm-gdd.

State the detected mode + (if subsequent) the current tag explicitly to the user as your first conversational message after Resume Check passes. They should never have to guess which tag they're working on.

Freeform Intake

Before invoking game-planner in initial mode, collect the user's rough idea in normal conversation, not AskUserQuestion.

• If $ARGUMENTS already contains a game idea, treat it as the freeform intake and do not ask again.
• If $ARGUMENTS is empty, ask the user for one open-ended paragraph: what they want to make, any references, mechanics, visual style, constraints, and anything they already decided. Make clear that rough notes are enough.
• Pass this freeform intake verbatim into the game-planner brief as "Initial User Concept".
• game-planner must skip questions that the freeform intake already answers and use those details to choose smarter defaults.
• Any "your call" / "you decide" language in the intake is scoped to the named topic or the current intake round unless the user explicitly grants broader delegation.

This intake is NOT a confirmation gate. Keep AskUserQuestion for explicit GDD and ROADMAP confirmations only.

Hard Rules

1. You CANNOT write game code (.gd/.tscn/.tres). Code lives in workers in /gm-build.
2. You CANNOT write to assets/. Assets are produced in /gm-asset.
3. Use AskUserQuestion for confirmation. GDD must be explicitly confirmed by the user before generating ROADMAP / per-tag artifacts. ROADMAP must be explicitly confirmed before any artifact is written.
4. MUST NOT skip the ROADMAP confirmation gate — see Sub-stages below. Initial mode WITHOUT a confirmed ROADMAP cannot proceed to artifact generation; subsequent mode with a roadmap edit WITHOUT re-confirmation cannot proceed either.
5. Subsequent mode does NOT append tag-N sections to STRUCTURE/SCENES. It overwrites those root files with the current tag's scope. Prior tags' versions live in docs/tags/<prev_tag>/. The cross-tag accumulating files are GDD.md, ROADMAP.md, STYLE.md, ASSETS.md, and MEMORY.md.
6. GDD design changes that contradict shipped tags MUST be reflected as PLAN refactor tasks. When subsequent-mode interview reveals that a prior tag's behaviour now needs to change, the GDD update marks the old behaviour as (superseded by ...) rather than deleting it, AND the new PLAN.md gains an explicit refactor / removal task in the Main Build section.

Sub-stages

1a — Interview & GDD update

Invoke the game-planner skill (.claude/skills/game-planner/SKILL.md).

Initial-mode brief MUST include the Freeform Intake as Initial User Concept; game-planner skips already-answered topics and uses the intake to choose smarter defaults.

• Initial mode: game-planner runs the full Socratic interview → produces fresh GDD.md.

• Subsequent mode: brief game-planner with:

  • Current tag id (e.g. v0.2.0)
  • The current ROADMAP.md entry for that tag (its bullet list)
  • The full existing GDD.md content
  • The previous tag's docs/tags/<prev>/PLAN.md Tag Mechanics list (so the conversation knows what already shipped)

  Game-planner asks the user: "We're about to plan {Tag}. ROADMAP currently says {bullets}. Do you want to keep that scope, adjust it, or change the underlying GDD design?" — and runs a focused interview. If the user changes design intent, game-planner updates GDD.md in place: new sections appended, replaced sections marked (superseded by ...). Old GDD content is never silently deleted.

Gate 1a:

• GDD.md exists and (if subsequent mode) reflects the user's latest intent
• User has explicitly confirmed the GDD update via AskUserQuestion (or said "no changes needed")

1b — ROADMAP generation / adjustment

This sub-stage exists in BOTH modes but does different work.

Initial mode:

1. Read GDD.md (now confirmed).
2. Derive a tag list following the SemVer convention from templates/ROADMAP.md:
   • First tag is always v0.1.0 and MUST deliver the first playable unit.
   • Every tag is a minimal playable unit: the player can experience a complete slice of gameplay with a completion, fail, or exit state.
   • Every tag includes the player-facing information needed to understand and play that slice.
   • Subsequent tags add one playable unit at a time.
3. Write a draft ROADMAP.md populated with this tag list.
4. MANDATORY gate: Use AskUserQuestion to ask the user:

   "Here is the proposed roadmap. Is it OK to proceed with v0.1.0 as defined? You can also reorder, split, merge, or rewrite tags before we move on."

5. If user requests changes, edit ROADMAP.md accordingly and re-confirm. Do NOT proceed to sub-stage 1c until the user explicitly confirms the ROADMAP.

Subsequent mode:

1. Read existing ROADMAP.md.
2. If sub-stage 1a's interview revealed any roadmap-affecting decisions (user wants to reorder remaining tags, drop one, add one, or move scope around), edit ROADMAP.md accordingly. Tags that already have git tag <tag> are immutable — never modify their entries.
3. If you modified ROADMAP.md in step 2: use AskUserQuestion to re-confirm the updated roadmap before continuing. If you did not modify it, no extra confirmation needed.

Gate 1b:

• ROADMAP.md exists, with at least the v0.1.0 entry (initial) or the current tag's entry intact (subsequent)
• User has confirmed the roadmap (either fresh confirmation or "no changes" acknowledgement)
• Current tag id is established (initial: always v0.1.0; subsequent: per Mode Detection)

1c — Per-tag decomposition

After GDD + ROADMAP are confirmed, decompose the tag in two phases. PLAN.md is the canonical source of task IDs, current-tag mechanic IDs, affected files, assets needed, and verify expectations. Do not generate STRUCTURE.md, SCENES.md, STYLE.md, ASSETS.md, or TOC.md in parallel with PLAN.md; those artifacts must read the finalized PLAN.md instead of guessing task mappings.

Phase A — PLAN first. Launch one decomposer for plan-package only. It owns only PLAN.md.

Agent({
  subagent_type: "decomposer",
  description: "Decompose current tag PLAN.md",
  model: "{decomposer_model from .godotmaker/config.yaml, default: sonnet}",
  prompt: "{shared brief below + Work Package: plan-package; Owned Files: PLAN.md}"
})

After it returns, the lead performs Gate 1c-A from disk and edits PLAN.md directly if needed. PLAN.md must be stable before Phase B starts:

• PLAN.md exists with **Tag:** header matching the current tag
• Tag Mechanics section is populated with stable [<Tag>-M<N>] ids
• Inherited Mechanics section is populated for subsequent mode, or omitted for v0.1.0
• Playable Unit section is populated and references existing mechanic ids
• Playable Unit describes player experience, unit outcome, scenes involved, and per-mechanic player operation / effect / visible evidence
• PLAN covers the player-facing state, feedback, and presentation needed to play the current tag normally
• PLAN Runtime Asset Assignments binds required visible content to ASSETS.md rows, procedural output, UI text, or not required this tag with a deferral reason
• Risk/Main tasks have stable task IDs and all Task Status rows start as pending
• Tasks list affected systems/scenes/assets clearly enough for downstream artifacts

Phase B — remaining artifacts in parallel. After Gate 1c-A passes, launch the remaining decomposer packages in the same message. Both packages MUST read the finalized PLAN.md and MUST NOT invent task IDs, mechanic IDs, affected files, or asset mappings that are absent from PLAN.md.

File ownership remains disjoint:

1. architecture-package: owns only STRUCTURE.md and project.godot.
2. scene-asset-package: owns only SCENES.md, STYLE.md, ASSETS.md, and TOC.md.

All Phase B packages may read every input path, prior archive, and finalized PLAN.md, but each package may write only its owned files. After all reports return, the lead performs Gate 1c-B from disk and edits any mismatches directly.

Agent({
  subagent_type: "decomposer",
  description: "Decompose current tag STRUCTURE.md and project settings",
  model: "{decomposer_model from .godotmaker/config.yaml, default: sonnet}",
  prompt: "{shared brief below + Work Package: architecture-package; Owned Files: STRUCTURE.md, project.godot}"
})

Agent({
  subagent_type: "decomposer",
  description: "Decompose current tag SCENES.md, STYLE.md, ASSETS.md, and TOC.md",
  model: "{decomposer_model from .godotmaker/config.yaml, default: sonnet}",
  prompt: "{shared brief below + Work Package: scene-asset-package; Owned Files: SCENES.md, STYLE.md, ASSETS.md, TOC.md}"
})

Single-agent fallback. If dispatching decomposer subagents is unavailable, launch one decomposer with no Work Package; it owns the full artifact set and runs all steps serially. If Phase A succeeds but Phase B parallel dispatch is unavailable, launch one decomposer for the remaining artifact set after PLAN.md is finalized.

Shared brief:

## Task: Decompose current tag into per-tag artifacts

### Mode
{initial | subsequent}

### Current Tag
{vX.Y.Z}

### Project Root
{absolute path to project root}

### GDD Path
{absolute path to GDD.md}

### Roadmap Path
{absolute path to ROADMAP.md}

### Templates Dir
{absolute path to .claude/templates/}

### Project.godot Path
{absolute path to project.godot}

### Manifest Path
{absolute path to assets/manifest.json — include only if file exists}

### Prior Tag Archives (subsequent mode only — empty list if no prior tags)
- v0.1.0: {absolute path to docs/tags/v0.1.0/}
- ...

### Inherited Mechanics (subsequent mode only)
{copy the union of Tag Mechanics from every prior tag's docs/tags/<prev>/PLAN.md;
each line must keep its `[<prev>-MN] description` format so decomposer can paste
them into the new PLAN.md's Inherited Mechanics section verbatim}

### Cross-Tag Refactor Hints (subsequent mode only — empty if none)
{any GDD changes confirmed in 1a that supersede prior-tag behaviour. For each:
- "<prior tag>'s <feature>" superseded by "<new design>"
- which files / systems likely need refactoring (best-effort guess; decomposer
  decides the exact PLAN tasks)}

### Work Package (two-phase mode only)
{plan-package | architecture-package | scene-asset-package}

### Owned Files (two-phase mode only)
{exact list of files this decomposer may write}

The decomposer overwrites root PLAN.md, STRUCTURE.md, SCENES.md with the current tag's scope. For STYLE.md it writes the initial visual seed and updates it only for confirmed visual-direction changes. For ASSETS.md it operates differently: in initial mode it writes the skeleton; in subsequent mode it APPENDS new rows for assets this tag introduces (with Tag = <current tag>) and never modifies prior-tag rows. It does NOT touch GDD.md, ROADMAP.md, MEMORY.md, or any docs/tags/ archive. It returns a short report; do NOT relay raw decomposer output to the user — run the gate first.

Gate 1c-B:

• Phase B packages used the finalized PLAN.md as the source of task IDs, mechanic IDs, affected files, asset mappings, and verify expectations
• STRUCTURE.md exists with **Tag:** header, scoped to this tag's additions / refactors
• SCENES.md exists with **Tag:** header, scoped to this tag
• STYLE.md exists
• ASSETS.md exists and any new rows are tagged correctly (per templates/ASSETS.md and gm-asset/SKILL.md)
• ASSETS.md Visual Asset Contract covers current-tag gameplay-visible objects with runtime size, scene/mechanic use, readability requirement, and source relationship
• ASSETS.md applies the Gameplay Actor Asset Rows contract to gameplay actors introduced or changed in this tag: only final runtime outputs are rows; canonical and action-source material remains production evidence
• ASSETS.md Visual Asset Contract binds large character UI display slots to portrait/display rows instead of gameplay runtime frame rows
• ASSETS.md Visual Asset Contract gives every current-tag gameplay actor concrete runtime size as both % viewport height and target pixels
• TOC.md updated (if decomposer touched it)
• Parallel-only consistency check: PLAN task IDs referenced by STRUCTURE/SCENES/ASSETS exist in PLAN.md; every PLAN task's affected scene/system/asset appears in the corresponding artifact or is intentionally marked deferred.
• Parallel-only mechanic ID check: every current-tag mechanic ID referenced by STRUCTURE/SCENES/ASSETS exists in final PLAN.md Tag Mechanics; no artifact references a guessed, renumbered, or stale current-tag mechanic ID.
• Playable Unit scene check: every scene named in PLAN.md Playable Unit appears in SCENES.md, or PLAN.md marks it deferred.
• Scene asset binding check: every SCENES.md gameplay object and non-text UI element binds to ASSETS.md, procedural output, UI text, or not required this tag with a deferral reason.

Fallback when subagent doesn't finish. If any gate item is unmet (whether the decomposer reported failure or just produced incomplete artifacts), do NOT respawn the subagent — instead, take over directly. Read whichever artifacts exist, identify the missing pieces, and write them using the same templates the decomposer would have used. The templates document their structure conventions.

User-facing 1c summary (after gate + any fallback complete). Build the announcement from the final on-disk state, not the raw decomposer report. Include: (a) decomposer's Risk Tasks Identified and Key Architecture Decisions (still useful as design rationale), (b) a fresh "files now on disk" line you observed yourself after fallback. If you took over for any file, say which ones — the user reads the actual files themselves.

Cross-session backstop — what each hook actually does:

• stage_reminder.py fires the moment you write {"role": "gdd", ...} to .godotmaker/stage.jsonl. It validates that the files declared in the gdd schema (config/stage_schemas.json → gdd.files) exist; if any are missing, it surfaces that to the lead before the gdd phase is considered done.
• check_stage_prerequisites.py fires when /gm-build (or /gm-fixgap) dispatches a worker — it re-validates the same schema. It does NOT gate /gm-asset (asset is not in WORKER_DISPATCH_ROLES); the asset phase relies on its own SKILL.md Resume Check instead.
• check_completion.py does NOT validate gdd at all — it only enforces worker-dispatch roles. Don't rely on it here.

So a partial 1c can still slip past /gm-asset if you skip the stage_reminder warning. The fallback above (lead takes over to fill missing files) is the primary safety net; the hooks are catch-up checks at later stages.

Available Skills & Subagents

When Done

After all three gates (1a, 1b, 1c) pass:

1. From the project root run python tools/append_stage_event.py gdd --tag=<current tag> to append a {"role": "gdd", "ts": "<server-generated UTC>", "tag": "<current tag>"} line to .godotmaker/stage.jsonl. Do NOT hand-write the JSON or the timestamp — the helper exists so the timestamp comes from the system clock, not your own output.
2. git add -A && git commit -m "chore(gdd): <Tag>"
3. Inform the user: GDD complete for <Tag>. Recommended next: /gm-asset (or skip straight to /gm-build if no new assets are needed for this tag — /gm-asset is manual and will simply STOP if there's nothing MISSING).