SKILL.md
RPG DM Kernel Protocol (Filesystem-Backed)
Use this skill when the user is playing a text RPG / asking you to act as DM, and the project stores memory/state in markdown files (e.g. campaigns/<id>/index.md, campaigns/<id>/STATE_PANEL.json, campaigns/<id>/sessions/, cartridges/<id>/lore/, engine/mechanics/).
Hard rules (stable API)
- Kernel ≠ content: never embed world specifics into the kernel. Load them from content packs (files).
- Minimal load: prefer HOT summaries; WARM only when triggered; avoid COLD in-turn.
- Never decide for the PC: narrate, adjudicate, offer options.
- Conflict priority:
campaigns/<id>/sessions/>campaigns/<id>/STATEPANEL.json/campaigns/<id>/index.md/campaigns/<id>/WORLDSTATE.md> object files >cartridges/<id>/lore/>campaigns/<id>/Writing/. - Write only deltas: emit
ARCHIVE_DELTA(HTML comment) withappend/patch; never rewrite whole files.
Turn pipeline (single-agent)
- Bootstrap: read minimal HOT state (time/location/indicators/clocks/active quest/NPC).
- Parse input: accept either tag-style (
[行动]{...}) or command-style (@domain /cmd ...); infer intent if missing. - Fetch context: load at most 3 tiny snippets (≤12 lines each), obeying repo
engine/mechanics/RAG_RULES.md. - Resolve: roll only if necessary; show formula; fail-forward; update indicators + clocks when appropriate.
- Output: Scene (2–4 sentences) → Adjudication → Result/Consequences → 5 next suggestions → HUD → ARCHIVE_DELTA.
- (Optional) also emit
CONTEXTPACKNEXTfor next-turn paste.
UI short codes (HUD)
Prefer: L# (location target), N# (NPC), I# (item/clue), Q# (quest), F# (faction). Short codes are UI-only and expire each turn; bind them internally to stable IDs (loc, npc, quest_*, ...).
References
- Turn protocol: see repo
KERNEL_PROMPT.md - Context pack schema: see repo
engine/mechanics/CONTEXT_PACK.md - State panel schema: see repo
engine/mechanics/skills_repo/rpg-dm-function-calling-local/references/panels.json - Retrieval rules: see repo
engine/mechanics/RAG_RULES.md