cd /news/ai-agents/i-vibe-coded-a-cryptid-reporting-sta… · home topics ai-agents article
[ARTICLE · art-137874] src=dev.to ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

I Vibe-Coded a Cryptid-Reporting Station on Sanity (and the Workflow Rejects Bigfoot Blurs)

A developer built Cryptid Field Station, a Sanity-based monorepo application that lets hikers report cryptid sightings and routes each report through an AI-plus-human verification workflow, with obvious fabrications auto-rejected and borderline cases queued for human review. The project uses a single transition() function for all write paths, a real-time Ranger triage app built on @sanity/client listen(), and a seeded jackalope report that admits to being taxidermy as a pipeline control sample. The build was prompted end-to-end in the agentic IDE OpenCode, with the developer noting course corrections away from the early-access @sanity/workflow-engine and @sanity/sdk-react in favor of a data-modeled workflow and a custom real-time interface.

by read3 min views2 publishedSep 23, 2026

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.

── more in #ai-agents 4 stories · sorted by recency
── more on @sanity 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/i-vibe-coded-a-crypt…] indexed:0 read:3min 2026-09-23 ·