Files
code_of_conquest_dnd/.claude/skills/world-building/SKILL.md
Phillip Tarrant b15bd2bb1e feat(skills): add project-local world-building skill
Author gritty, disposition-gated Margreave content (lore bible + hand-authored
quest/item/location/origin JSON) that reconciles against existing canon before
writing, holds the world's hard-by-default / warm-when-earned NPC tone, and keeps
the content_build --check gate green.

- SKILL.md: read-before-write workflow, read-only on existing canon (propose
  diffs, require approval), build+check+pytest as the done bar.
- references/schema.md: full bible + JSON contract, secrecy scale, gate model.
- references/tone.md: tone contract + NPC disposition-warmth model.
- scripts/canon_index.py: dumps existing ids for reconciliation.
- evals/: 3 test cases (all green in isolated-worktree runs).

Un-ignore /.claude/skills/ so project skills are versioned like code; all other
.claude/ dirs (root + nested api/client) stay ignored.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HuHRPE7VfppUJEaoGBEUqZ
2026-07-12 11:15:34 -05:00

7.1 KiB

name, description
name description
world-building Author gritty, dark, disposition-gated world content for Code of Conquest (the Margreave). Use this WHENEVER the user wants to write, expand, or add lore, storylines, plot threads, secrets, factions, regions, towns, dungeons, NPCs and their knowledge/dialogue, quests, items, locations, or starting origins — even if they don't say "world-building" or name the content/ folder. Triggers on "write an NPC", "add a town/region/faction", "new quest", "a secret the party can uncover", "expand the Tallow Reach", "flesh out Duncarrow", "someone the player meets on the road", "cursed item", "give me a storyline", and the like. The skill reconciles new content against existing canon (never contradicting or silently changing it), writes it in the exact source format the content build tool consumes, keeps NPCs true to the world's hard, uncaring tone (warm only when disposition is earned), and verifies the build stays green. Do NOT use it for engine/UI code, prompts (/api/prompts), or combat rules.

World-building — the Margreave

Author new content for Code of Conquest, a gritty single-player CRPG. Read CLAUDE.md (repo root) once per session for the charter — this skill assumes it.

The one rule that governs everything you write

Code owns state. AI owns text. (charter §2)

You are authoring static content, not runtime state. Facts, personas, places, and knowledge chains are text the engine references by stable id. You never write initial disposition numbers into world content, never gameplay logic. A starting value (where the player begins, a seeded disposition) belongs to an origin, never to a town or NPC.

The two things that make this hard

  1. New content must not contradict old content. The build tool checks structure (ids resolve, secrecy routes correctly) but it CANNOT tell that you just described the north road as untraveled when canon says elves walk it. A contradiction ships silently and the DM narrates a broken world. So: read the existing canon before you write a word, and treat every existing file as read-only unless the user explicitly approves an edit (see below).

  2. The tone is specific and easy to get wrong. Gritty, not grim. The world is hard and does not care about the player — but an NPC who has come to trust them becomes genuinely warm. Getting either pole wrong (grimdark misery, or friendly-NPC-from-hello) breaks the world. See references/tone.md.

Workflow — follow in order

Make a todo per step.

1. Understand what's being asked, and which layer it lives in

Content kind Format Lives in
People, places, regions, factions, world rules, knowledge (rumor/fact/secret), NPC dialogue layers Markdown bible```yaml blocks content/lore/*.md (source)
Quests / story skeletons Hand-authored JSON content/world/quests/
Items (incl. cursed/blessed) Hand-authored JSON content/world/items/
Locations (map nodes) Hand-authored JSON content/world/locations/
Origins (starting seeds) Hand-authored JSON content/origins/
Degraded-DM fallback prose Markdown/text content/fallback/

The lore bible is the heart — it produces the disposition-gated dialogue that makes NPCs feel alive. content/world/ and content/server/ under canon/, topics/, and npcs/ are BUILT from the bible — never hand-edit them. The JSON under quests/, items/, locations/, and origins/ is hand-authored and loosely schema'd (POC seeds); edit those directly.

If a storyline you're writing implies a quest, item, or location that doesn't exist yet, say so and author the stub, or flag it — don't leave a dangling id.

2. Load and reconcile against existing canon — before writing

Run the canon index to see every id, type, secrecy, and one-line gist already authored:

.venv/bin/python .claude/skills/world-building/scripts/canon_index.py

Read the entries your new content touches (people it references, the region it sits in, secrets it brushes against). New content must slot into this without contradiction. Reuse existing ids in related: rather than inventing parallel ones. If your idea only works by changing something already canon — a fact's body, a disposition, who knows what — STOP. Do not edit. Present the exact before/after change and the reason, and get explicit approval first. New files and new entries need no approval; edits to existing canon always do.

3. Draft in the correct format

  • Bible entries: see references/schema.md for the full field-by-field contract, the id/namespace/type rules, the secrecy scale, and the knowledge-gate model. Match the voice and structure of content/lore/duncarrow.md — it is the reference specimen.
  • Hand-authored JSON: mirror the shape of the existing file in that folder (references/schema.md documents each). Keep ids stable and lowercase-snake.

Write NPCs to the tone and disposition-warmth model in references/tone.md. A hard NPC starts guarded (cold/neutral, sometimes reachable-hostile), withholds more than a stranger would expect, and both reveals more and softens in manner as the player climbs their ladder. That warmth is earned, never default — it's what the five-rung gate system exists to deliver.

4. Build and verify — not done until green

The build regenerates content/world/ + content/server/ from the bible and the --check gate is what CI enforces (.github/workflows/content-build.yml). Run, from repo root, with the repo venv active (source .venv/bin/activate) — the content_build package lives under tools/, so PYTHONPATH=tools is required:

PYTHONPATH=tools python -m content_build          # regenerate world/ + server/ from lore/
PYTHONPATH=tools python -m content_build --check   # must exit 0: fresh, valid, no orphans, no secrecy leaks
PYTHONPATH=tools python -m pytest tools/content_build/tests -q   # schema + secrecy + cli coverage

(If python isn't found, the venv isn't active — use .venv/bin/python.) If any step fails, read the BUILD FAILED / STALE: message — it names the source file and line — fix the bible, rebuild, recheck. Never hand-edit the generated JSON to make --check pass; fix the source. (Hand-authored JSON under quests/items/locations/origins isn't built — the build tool ignores it. If you touched those, just confirm they're valid JSON.)

5. Report

Summarize: new ids added (by layer), any existing id you reused in related, anything you flagged for approval, and confirmation the build + check + tests are green. Update content/lore/canon-roadmap.md's Authored/Pending lists if you added a new bible file or closed out a pending thread.

Reference files

  • references/schema.md — the complete bible + JSON schema, all validation rules, the secrecy scale, the knowledge-gate model, worked examples. Read before authoring bible entries.
  • references/tone.md — the Margreave's tone contract and the NPC disposition-warmth model. Read before writing any NPC or narrative body.