AI-Authored Content Coherence
How to keep content an AI or procedural engine generates coherent at scale — one voice, no retcons, figures and events that cohere into arcs rather than dissolving into interchangeable mush. This is the crux discipline for a headless / AI-authored game, and it is the least solved — treat its rules as engineering hypotheses to de-risk early, not settled law.
When to use this
- Designing the generation pipeline for lore, history, NPCs, items, quests
- Writing the voice spec / style corpus and the machine-readable lore bible
- Building the self-review / coherence pass over generated batches
- Diagnosing generated content that feels samey, contradicts itself, or drifts in voice
Scope
This skill owns coherence of generated content (mostly non-visual). Adjacent siblings:
- The generator itself + the sameness problem →
procedural-generation - The review pass that gates a batch →
procgen-review(this skill's self-review pass is that gate) - Visual coherence / house style (the visual analogue) →
art-direction-and-readability - The lore the bible encodes →
worldbuilding-and-lore - Quests generated over world state →
narrative-and-quest-design
How the pieces fit
GUIDE.md— the why, with exemplars and citations. Heavily flagged: this topic's verification phase largely failed (session limit), so nearly all rules are sourced but unverified. The sources themselves are strong (Compton; Grinblat & Bucklew / Caves of Qud; LLM-consistency papers).CHECKLIST.md— the what to do: Do/Don't plus "Test for" criteria for a generator, a corpus, and a self-review pass.
The one idea to anchor on
Mathematical uniqueness is not perceived variety. A generator can emit endless mathematically-distinct artifacts that players still read as one undifferentiated mass (the "10,000 bowls of oatmeal"). Coherence-at-scale is won by a shared voice + a never-violate fact bible + recurring thematic domains + an explicit self-review pass — and even then, coherence at full open-world scale is unsolved. De-risk with the oatmeal/fanfic tests from day one.
⚠️ Verification note: Built from a deep-research pass whose verification phase was cut short by a session limit; the synthesis step also partly failed. Almost every rule here is sourced but unverified — strong, well-cited defaults that should be re-verified, not treated as confirmed. See the banner in GUIDE.md.
Start with GUIDE.md, then apply CHECKLIST.md.