Own the end-to-end game outcome. Build the playable loop, route through the right phases, verify evidence, and do not call prototype-quality work premium. "Less basic" from the user means the current visual level is rejected: treat it as the premium bar.

Before planning, note what this runner can do and adapt:

Try in order, expanding to the user's home directory when the read tool requires absolute paths:

Reference files resolve the same way: . Sibling skills point back to this ladder instead of restating it.

For broad work (complete, premium, AAA, polished, high-fidelity, showcase, from-scratch, upgrade, release-ready), load all five phase skills before implementation: , , , , . For narrow director-invoked work, load the directly relevant sibling plus . Do not skip sibling loading because this director bundles a phase playbook.

Load generator skills before deciding generated assets are unnecessary, whenever their trigger surfaces exist in premium/AAA/showcase/complete/release-ready/"less basic" work:

References are phase-entry gates, not optional enrichment. The canonical per-phase Required References list lives in ; load that file at planning time for broad work and at phase entry otherwise.

If Task/subagent/workflow tools are available, delegate each major phase to a focused worker with the phase plus its required references explicitly loaded. If unavailable, execute serially after loading the same files.

Keep four ledgers: skill-loading, reference, external asset sourcing, and phase execution. Templates live in .

Compaction rule: report every row that has meaningful state (yes/no/blocked/done/skipped plus path or evidence), and collapse consecutive rows into a single line naming them. Never omit or compress rows that carry real state.

When a sibling skill file is loaded, follow its workflow for that phase. Phase entry/exit evidence, ledger templates, and the fallback procedure for unloadable siblings all live in .

New projects use the gameplay skill's scaffold creator; canvas verification uses the generated game's or the QA skill's packaged inspector:

python3 <threejs-gameplay-systems-skill-dir>/scripts/create_threejs_game.py ./my-game
node <threejs-qa-release-skill-dir>/scripts/inspect-threejs-canvas.mjs --url http://127.0.0.1:5188

Premium, AAA, polished, complete, release-ready, and showcase requests require visible quality across gameplay, hero/player, obstacles/enemies, rewards/interactables, world kit, HUD/menu states, render/lighting/materials, feel, performance/mobile, and QA. If screenshots are dominated by primitives, flat roads/arenas, generic stat cards, sparse worlds, or glow-only detail, the task is not done. The full completion gate is in .

When shell tools are available, draft the final evidence report to a markdown file and audit it before finalizing broad or premium work:

python3 <director-skill-dir>/scripts/audit_reference_report.py --premium /path/to/final-report.md

Use for premium/AAA/showcase/high-fidelity/polished/complete/release-ready/"less basic" claims; add for physics-heavy games; add when generated or integrated audio is in scope; add only for debug/perf/QA-only reports with no gameplay claims. If the audit fails, fix the missing sections or state the exact blocker instead of claiming completion. If the script is unavailable, manually enforce the same sections listed in Required Verification.

Report the ledgers (compacted per the rule above), game design brief, core loop contract, level/encounter plan, files changed, run URL, controls, verification commands, screenshots/artifacts, renderer/performance notes, technical art budget, visual test harness decision, quality gates passed, skipped phases, and remaining risks. For premium/AAA/showcase claims, include the filled visual scorecard with measured evidence and automatic failures remaining. Be precise: "invoked" means a slash/tool skill invocation; "loaded" means the file was read into context; "executed phase" means the work was performed under loaded skill guidance or the phase playbook.