thesethrose/dungeon-agent/build-world

Build a World (author a new campaign, grilling-driven)

A process skill. The DM authors a NEW, runnable campaign by interviewing the player through the design decisions (see grilling), then produces the two content artifacts: stats-schema.json (this game's character stats) and world.md (premise, locations, monsters, loot, win condition). The rules engine, save/load, combat, and anti-cheat are inherited from AGENTS.md — you only define content.

When to use

• Player asks for a brand-new world/game/dungeon not in the existing campaign skills.
• Player wants a genre/style there's no campaign for yet.

The three artifacts to produce

1. stats-schema.json — character sheet definition: stats (name/type/min/max/start),

derived formulas, item & ability whitelist, win condition.

1. world.md — premise, locations & exits, monsters, loot, XP curve, tone.
2. persona.md — the DM's narrator voice/tone for THIS campaign (adopted on

load per SOUL.md Mode 2). Optional but recommended — without it, the DM narrates in a neutral default.

Procedure

0. Research first for established settings (skip grilling of facts)

If the player names a known IP / era (e.g. "Star Trek", "Middle-earth", "1920s Chicago", "Warhammer 40k"), the lore is a FACT you supply — pull characters, backstories, factions, tech, and canon from your knowledge / past sessions / web research. Do NOT grill the player about established canon. Summarize the relevant lore you're anchoring to, then confirm before building. (See grilling skill: facts → research; only decisions → grill.)

1. Grill the player on DECISIONS only (one at a time, via clarify)

Use grilling. ONE question per clarify call, each with a recommended default. Resolve dependencies in order:

1. Genre/tone — high fantasy / grimdark / sci-fi / comedy / horror.
2. Setting — one-line premise ("a haunted space station", "a flooded kingdom").
3. Goal / win condition — what ends the game.
4. Player power level — starting power (affects starter stat values & early monsters).
5. Combat or puzzle focus — how central combat is (tunes monster XP & stat weights).
6. Confirm shared understanding before building.

2. Design the stats schema (stats-schema.json)

Fields the engine's validator understands:

• campaign (unique id → save folder), title, start_location, schema_version.
• stats[] — each: name, type (int|float), min, max, start.

Typical set: hp, max_hp, atk, def, level, xp, gold, mana (adapt to genre).

• derived[] — auto-recomputed stats: target + expr (e.g.

{"target":"max_hp","expr":"20+(level-1)*4"}).

• items[], abilities[] — the STRICT whitelist. Validator rejects anything not

listed. Backbone of the no-invented-powers anti-cheat.

• win_condition.

Design rules:

• Every stat used must be declared; unknown stats are treated as cheats.
• start values within min/max; derived targets recompute cleanly.
• Keep whitelists complete but tight — every item/ability in world.md must be listed.
• The campaign folder is SELF-CONTAINED: schema at references/stats-schema.json,

saves at <campaign>/saves/<slot>.json (INSIDE the folder). Keeping saves inside the folder is what makes the campaign portable — copying the folder carries progress too. Do not place saves in a shared engine-level dir.

3. Write the world (world.md)

Structure: premise → starting gear/abilities (subset of whitelist) → locations with exits (flow start → ... → final, final named start_location) → monsters (name, hp/atk/def, XP) → loot (effects) → XP & leveling → win condition.

4. Register it as a campaign skill

Copy campaign-template, rename to campaign-<name>, write schema + world + persona into its references/ (and an optional copy under campaigns/<name>/ for the engine's default path). Update the SKILL.md description + starting procedure.

5. Validate (proof it runs)

engine/save_utils.py new --schema campaigns/<name>/stats-schema.json --player test --slot test
engine/save_utils.py validate --schema campaigns/<name>/stats-schema.json --save campaigns/<name>/saves/test.json

Must print VALID. Fix and re-validate, then delete the test slot.

Anti-cheat relevance

Well-formed campaigns are cheat-proof by construction: stats, items, abilities are schema-whitelisted and validator-enforced, so the player physically cannot have a power or stat the world doesn't define.

Pitfalls

• Forgetting to list an item in items[] that world.md references → "unknown_item".

Keep schema and world in sync.

• Derived formula errors → "derived_mismatch". Test with a fresh save.
• Don't set start HP so low the player dies instantly, nor so high there's no

tension. For a level-1 crawl, 15-25 HP is a safe default.

• Reuse this skill every time you author a world for a consistent format.