←back to the wallSimmonsBench / 48 of 60
WRITING
Build a text-adventure dungeon (writing test)
the prompt
/goal Build me a playable text-adventure dungeon crawl in the browser, where the writing is the star.
Requirements:
1. A 12-room dungeon with a coherent identity (you invent it — one strong theme, not a generic fantasy mishmash): interconnected rooms, a few locked doors/keys, 2-3 light puzzles, one inhabitant encounter, and an ending.
2. Classic parser-lite interface: type or click commands (go north, look, take, use), inventory, and a map that fills in as I explore. Clean terminal-styled UI, readable type.
3. THE WRITING IS THE PRODUCT. Second-person room descriptions, 40-90 words each, written like a great DM runs a table: concrete sensory detail, economical sentences, personality without theatrics. Every room description, item description, and encounter line should read like it was written by a person with taste. Rewrite any line that sounds generic.
4. Reactive flavor: rooms describe differently on revisit, actions get specific responses (not "you can't do that"), and the ending text pays off the theme.
Deploy live to Vercel. QA every room, exit, and puzzle path.
DONE when: the deployed dungeon is completable start to finish, the map and inventory work, and the prose in every single room is writing you'd be proud to sign.
---
## Grading rubric (Pat only — never in the prompt)
Read 5-6 room descriptions per model aloud on camera. Count:
- Negative parallelism: "this isn't just an X — it's a Y" (the comment's exact parody)
- False-suspense beats: "Here's the thing...", "But something feels off."
- Hype adjective stacking ("ancient, forgotten, impossibly vast")
- Self-posed rhetorical questions ("What could have made these marks?")
- Em-dash chains / triplet rhythm addiction
Winner = the dungeon whose prose sounds most like a human DM. Same UI styling across models per the jam_daniels comment rule (blind compare).
…
GENERATION · LOCKED
Drag back for older model generations · coming soon
🔒 LOCKED
JAN '25
JUNE '26
All 0 models · same prompt · $0.00 total
sort: