Create ArchitectureSAFE

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

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: create-architecture
description: "Author the architecture blueprint before code is written. Validates decisions against the pinned engine, flags knowledge gaps."
argument-hint: "[focus-area: full | layers | data-flow | api-boundaries | adr-audit] [--review full|lean|solo]"
user-invocable: true
allowed-tools: Read, Glob, Grep, Write, Edit, Bash, AskUserQuestion, Agent, Bash(bash "*/.claude/skills/create-architecture/../../hooks/yaml-helper.sh" resolve_config *)
model: sonnet
---

!`bash "${CLAUDE_SKILL_DIR}/../../hooks/yaml-helper.sh" resolve_config --keys review_mode,automation,workflow,docs.density`

Resolved above — use as-is; `--review` overrides `review_mode`. No block →
defaults in `.claude/docs/config-resolution.md`.


# Create Architecture

This skill produces `docs/architecture/architecture.md` — the master architecture
document that translates all approved GDDs into a concrete technical blueprint.
It sits between design and implementation, and must exist before sprint planning begins.

**Distinct from `/architecture-decision`**: ADRs record individual point decisions.
This skill creates the whole-system blueprint that gives ADRs their context.


See `.claude/docs/director-gates.md` for the full check pattern. Individual gate definitions live in `.claude/docs/director-gates/[gate-id].md` — the spawned agent reads its own gate file; do not read it in the parent session.


Every `AskUserQuestion` call follows `.claude/docs/automation-modes.md`
(collaborative asks always · guided major-only · autonomous logs and proceeds;
`automation_always_ask` categories always prompt).

**`docs.density`** — it controls per-section *depth*, where `workflow`
controls which sections exist. `modes.rigor` sets both together; set
`docs.density` explicitly to vary depth alone: `terse` (the default, via `rigor: minimal`) = layer diagrams + decision bullets,
no essays; `balanced` = diagrams + paragraph explanations of layer choices
(`rigor: standard`); `thorough` = full prose with rationale, trade-offs, and alternatives
considered per layer. Apply it to every section you author.

**`workflow`** (see `.claude/docs/workflow-modes.md`):
- `full` — full architecture: all layers, module ownership, data flow, API
  boundaries, full ADR audit.
- `standard` — simplified: system layer map + critical ADR list only.
- `minimal` — not required. Can still be run voluntarily.

**Argument modes:**
- **No argument / `full`**: Full guided walkthrough — all sections, start to finish
- **`layers`**: Focus on the system layer diagram only
- **`data-flow`**: Focus on data flow between modules only
- **`api-boundaries`**: Focus on API boundary definitions only
- **`adr-audit`**: Audit existing ADRs for engine compatibility gaps only

---

## Phase 0: Load All Context

Before anything else, load the full project context in this order:

### 0a. Engine Context (Critical)

Read the four project-wide engine documents in full — they are small, and every
part of each is used:

1. `docs/engine-reference/[engine]/VERSION.md`
   → Extract: engine name, version, LLM cutoff, post-cutoff risk levels
2. `docs/engine-reference/[engine]/breaking-changes.md`
   → Extract: all HIGH and MEDIUM risk changes
3. `docs/engine-reference/[engine]/deprecated-apis.md`
   → Extract: APIs to avoid
4. `docs/engine-reference/[engine]/current-best-practices.md`
   → Extract: post-cutoff best practices that differ from training data

Then read **only the module docs whose domain this game actually uses** —
not the whole `modules/` directory:

5. `docs/engine-reference/[engine]/modules/` — glob it to establish what exists,
   then match against the domains present in `design/gdd/systems-index.md`
   (the same domain vocabulary the ADR template uses: Physics, Rendering, UI,
   Audio, Navigation, Animation, Networking, Core, Input). Read the matching
   modules; skip the rest.
   → Extract: current API patterns per domain

   A game with no multiplayer system does not need the networking module loaded
   to write its architecture, and loading it costs the same as one that does.
   **If the domain match is ambiguous, read the module** — a missed engine
   constraint is far more expensive here than a redundant read, because this
   phase is where those constraints get baked into the architecture.

If no engine is configured, stop and prompt:
> "No engine is configured. Run `/setup-engine` first. Architecture cannot be
> written without knowing which engine and version you are targeting."

### 0b. Design Context + Technical Requirements Extraction

Load the approved design documents and extract technical requirements from each:

1. `design/gdd/game-concept.md` — game pillars, genre, core loop
2. `design/gdd/systems-index.md` — all systems, dependencies, priority tiers

**Check both exist before reading either. Neither is optional here, and both
need an absence branch** — §0a stops for an unconfigured engine, and these two
matter just as much:

- **`systems-index.md` absent** — stop:
  > "No systems index found. Run `/map-systems` first. An architecture written
  > without it invents layers for systems nobody mapped, and every ADR, epic and
  > story downstream inherits that invention."
  At `minimal` the index is not required (§ tier note above) — say so and proceed
  from the brief instead.
- **`game-concept.md` absent** — at `standard`/`full`, stop and point at
  `/brainstorm`. At `minimal`, read `design/game-brief.md` in its place; if that
  is absent too, stop — there is no design record to architect against.
- **Either present but empty or still template placeholders** — treat as absent.
  Present-but-empty is the case that most looks like present.

Do not proceed on a partial read and note it later. This phase is where design
assumptions get baked into ADRs, and an assumption made here is re-derived by
everything downstream rather than re-checked.
3. Project config — `naming.*` and `performance.*` from `project.yaml` (for any
   key absent or empty,

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.