gigio1023

Review real Minecraft LLM agent runs from transcript files, canonical runtime artifacts, social-cycle reports, actor workspaces, logs, Langfuse traces, human-visible behavior notes, and code diffs, then infer concrete implementation flaws and next fixes. Use for prompts like "review this agent run", "extract implementation insights from the run", "analyze the NPC behavior loop", "review the 100-cycle result", "why did the NPC stall?", "the matrix broke at 12/12", "collectLogs stopped at 3/4", "the screenshots look wrong", "review visual evidence", or "use Langfuse traces to propose runtime fixes" after Minecraft runtime, action-skill, Actor Turn, generated Mineflayer, social-cycle runs, or embodied co-actor legibility session runs.

Judge and design meaningful research directions for this Minecraft LLM-agent project before implementation. Use for prompts about "research direction", "research gap", "paper-worthy", "experiment design", "legibility", "co-actor", "observer prediction", "논문감", "연구 가치", "실험 설계", "유의미한 연구 주제", or historical WAM/F-native/F-loop/F-society plans that need re-evaluation under the current active plan. Trigger when a plan hides behind verified/structured/validated instead of proving a substantive claim.

Write, review, validate, or prepare Minecraft runtime run reports, HTML/static reports, model-comparison summaries, Qwen/OpenAI/Gemini writeups, visual evidence reports, public artifact summaries, or Korean/English "보고서" requests from this repo's social-cycle, action-skill, provider, visual, or embodied co-actor legibility artifacts including transition-row, public-history, experiment-declaration, prediction, and score-report bundles. Use before drafting report prose or publishing experiment conclusions.

Generate or review bounded Mineflayer TypeScript action skill candidates for this repo's Actor Turn `author_mineflayer_action` path. Use when writing generated source, helper allowlists, input schemas, verifiers, timeouts, or Mineflayer behavior notes for actor-owned action skills.

Use before any live provider-backed Minecraft run, LLM/provider benchmark, model availability decision, quota reset check, model-comparison lane, report rerun, long social-cycle run, OpenAI/Gemini/ModelScope smoke, or provider budget/429/auth issue. Checks requested provider/model candidates across OpenAI, Gemini, and ModelScope using reset windows, local usage ledger, built-in policy matrix, local emergency brakes, dashboard approval state, and whole-run request/token estimates. Trigger for Korean/English prompts like "quota 확인", "초기화됐는지", "run gpt/qwen/gemini benchmark", "사용 가능한 모델인지 체크", and "provider preflight".

Use when writing, editing, or reviewing dream-of-one player-facing content — Korean NPC dialogue and tone, storylets, scenario canon, suspicion why-lines, HUD strings, or six-locale localization. Triggers include "storylet", "대사", "말투", "why-line", "이유 문장", "Korean voice", and "localization keys". NOT for engineering docs, code comments, runtime logic (dream-npc-runtime), or client presentation (dream-godot-client).

Use when changing, reviewing, or diagnosing dream-of-one's Godot client under godot/ — the first-person 3D town, actors, doors and props, HUD, subtitles, modal conversation, input, assets, or headless smokes. Triggers include "first person", "HUD", "conversation surface", "navmesh", "subtitle", "interaction ray", and "scene smoke". NOT for suspicion/record/run logic (dream-npc-runtime), Korean wording and tone (dream-content-authoring), or engine-generic Godot questions.

Use when Dream of One work would require a long or high-volume Godot AI MCP session, repeated runtime inspection or captures, or any hands-on game playtest. Keeps the lead session in place, spawns one native in-session high-capability subagent with exclusive run ownership, and returns compact evidence. NOT for code-only Godot edits, a bounded session preflight, backend-only work, or generic parallelization.

Use when operating, playtesting, or live-debugging Dream of One through its configured Godot AI integration, including runtime captures, scene-tree inspection, bounded player input, or reproducing a player-facing Godot issue. NOT for code-only Godot changes without a live run (dream-godot-client), backend-only checks (dream-npc-runtime), or OS-window testing outside Godot AI.

Use when changing, reviewing, or diagnosing backend/npc-runtime in dream-of-one — run/session lifecycle, NPC scheduler and memory, stance/hearing judgment, records and civic ledger, provider ports/adapters/profiles, locale-aware envelopes, fail-visible provider interruption, or runtime tests. Triggers include "RunService", "provider profile", "adapter", "envelope", "provider failure", "stance", "hearing", "ledger", "tool validation", and "run/session API". NOT for Godot scene/HUD work (dream-godot-client) or authoring player-facing content (dream-content-authoring).

Use when implementing, debugging, reviewing, or upgrading a Godot project, including project.godot, GDScript, Godot C#, .gd, .tscn, .tres, scenes, resources, signals, nodes, UI, physics, shaders, imports, exports, addons, or engine-version changes. Also trigger on Godot resource terms such as ExtResource, SubResource, NodePath, and res:// even when "Godot" is omitted. NOT for engine-neutral game design, unrelated C#/GLSL work, asset creation alone, or contributing to the Godot engine source itself.

Use when creating, restructuring, rewriting, humanizing, or reviewing prose documents in English or Korean: README, guides, specs, API references, ADRs, memos, wiki pages, blog drafts. Covers de-AI-ifying tone ("humanize", "deslop", "AI 티 빼줘"), turning notes into standalone docs ("이 메모를 문서로"), surgical Korean polishing ("윤문해줘"), terminology review ("용어 이상한지 봐줘", "term check"), and repository-grounded doc authoring ("README 작성", "문서 정리해줘", "스펙 작성"). NOT for Python docstrings, prompt coaching, diagrams, PR/commit/issue copy, translation, or implementation changes.