This is a submission for the Sanity Challenge, Path Two: Vibe-Code Something Strange Cryptid Field Station - a field guide where hikers report cryptid sightings, an AI + human workflow verifies them, and rangers triage from a live command center.
submitted ? in-review ? verified ? published); obvious fabrications auto-reject; everything in between waits for a human verification document next to the sighting - one audit trail, two authors (agent + human).
The control sample is my favorite part: the seed data includes a jackalope report that openly admits to being taxidermy. If the pipeline ever publishes it, the pipeline is broken. It gets rejected.
studio/ today (npx sanity dev); the hosted Studio URL lands here right after sanity deploy (custom evidence input below)
Try it: file a sighting at /report (lat/lng optional but the weather check loves them), then watch it land in the Ranger Board's submitted lane and move through review.
web/ - Next.js 14 guide + workflow API routes (/api/sightings, /transition) EvidenceGifInput (paste a trail-cam video URL, get a preview + capturable 6-frame GIF-strip)ranger/ - real-time triage app (live Sanity subscription, own UI)workflows/sighting-review.ts - the review process as data (stages, guards, effects)seed/ - 7 cryptids, 12 sightings with mixed fates, idempotent importer
IDE: OpenCode (agentic, autonomous) powered by Muse Spark - prompted end-to-end, no interactive Studio clicking for the build itself.
Prompts that worked: "scaffold a monorepo with these four folders and make every app build before touching Sanity" (got three green builds in a row); "every write path goes through one transition() function" (kept agent + ranger honest); "reads must work with an empty dataset" (the baked-in field cache, which saved the demo before seeding).
Prompts that didn't: asking for the App SDK with @sanity/sdk-react from memory - I couldn't pin the exact provider API without docs in hand, so I course-corrected to a real-time custom app on @sanity/client listen() with its own UI, which is what the bonus actually asks for ("your own interface, instead of another read-only frontend"). The workflow engine went the same way: rather than fight the early-access @sanity/workflow-engine setup, I modeled the process as data (status + verification docs, cookbook-mapped in workflows/), which runs on any plan today.
Where it got stuck: an npm cache corruption on esbuild killed the Studio install (fixed with --prefer-online retry); the Vercel CLI hung on update-notifier checks in a non-TTY shell (fixed with NO_UPDATE_NOTIFIER=1); a TS slug-shape mismatch between raw seed JSON and the Cryptid type (fixed by mapping slug.current at the fallback boundary).
Past the Studio: yes to both bonuses. The Ranger app is a custom real-time surface with live updates and approve/reject/publish actions - not Studio, not read-only. The workflow is data: the agent moves drafts forward (triage, auto-verify ?90, auto-reject ?20) and the ranger approves through identical transitions, all audit-logged side by side.
yy3ugxmv (dataset production) cryptid (classification, danger 1-5, traits, habitat), sighting (cryptid ref, geopoint, evidence URL+type, workflow-owned status, cached weather), expedition (cryptid/sighting refs for joint field pushes), credibilityScore are readOnly in Studio - the workflow owns them, humans use the board; evidence type drives conditional rendering (video player vs image); the jackalope is debunked by design as a pipeline control.
Built autonomously in OpenCode; transcript available on request. Key decision log: monorepo-first, cache-first reads, one-transition-function rule, honest 503s when the write token is absent.