brainstorm: hunker

Capture of the chat that produced the hunker tracer. Kept as the reasoning trail behind the spec.

How we got here

• Started from the seed epic list (persuade-the-guard, secret-rule, interrogation, …). Persona toys were the cheapest first tracers but felt slight.
• Pivot: make the first tracer substantial and de-risking — the deliberate embryo of the MUD-LLM-roguelike graduation target (agent-mud), and have building it stand up the shared web launcher every later prototype needs.
• Reference frame: Hoplite — tile-based, one-screen, turn-based, "minimal but complete environment." Not a large world; boxed.
• Landed on a one-screen survival homestead: a valley + a cozy home you live in, where the LLM runs weather, events, growth/decay, and sieges.

Key decisions (and why)

• Adjudicator is the soul, not the Director. Both ship, but the fun-test measures improvised survival because it's falsifiable in a 10-minute session and more distinctly "only an LLM could do this." Director intelligence is hard to tell from good RNG in a short session, so it's the delivery layer, not the thing under test.
• Finite goal ("survive until the thaw"). Completable → clean fun-test. Endless-score roguelike is a later option.
• Bigger than 9×9. ~20×14 exterior so crossing costs real time (route tension), ~8×7 interior so the home feels cozy and you move around. Still boxed — regions/teleport are the production game.
• Homestead stations as adjudication surfaces. Hearth, forge, bed, storage now; research/magic wards + deep farming as slice-2 fast-follows on the same pattern.
• Interior/exterior split doubles as the smallest honest "rooms" — direct MUD lineage.
• Anti-"yes-and" is structural, not vibes. Engine owns integers + all dice; LLM proposes deltas + difficulty; engine validates, rolls, applies.

Meta: build for churn

• The force multiplier isn't a saved process — it's a clean engine/content seam so game #2 is clone-and-swap (see the engine-content-seam ADR). A pivot like "maybe secret-rule would've worked better" = clone hunker/, swap content into the same manifest, new soul on the same harness.
• Skillify the process only after hunker actually works, so we encode a proven process, not a guessed one — in the spirit of "prove it before you build it." Running notes now make writing that skill cheap later. The skill would scaffold a new game folder from the template, fill the manifest, drop in starter prompts, and register it in the launcher.

The process we're encoding (draft, to skillify later)

1. Name the fun-test as one emotional sentence.
2. Find where the LLM is load-bearing — or admit it's a lookup table and stop.
3. Draw the honest split: engine owns truth + all dice; LLM proposes + narrates.
4. Concrete minimal-but-complete environment.
5. Cut slice 1 to just prove the feeling; defer the rest onto the same harness.
6. Play it, judge the fun-test, log the lesson (fun-proven / not-fun-shelved).