WitUS ecosystem · Wanderlearn

Help & onboarding

The Field Reporter is an agent that turns a raw capture — a location transcript, GPS, and photo references — into a publishable lesson draft. It researches the place, writes a cited draft, then critiques its own draft against a quality checklist and revises until the draft passes or it asks a human to take over. This page explains how to use it, whichever of the three readers below you are.

1. Browsing reports (no account needed)

The report views are public — you do not sign in to read them. Start at the report list and open any report to read the generated lesson.

  • Read the final lesson — the published draft, rendered with its citations.
  • Scrub the revisions — the side-by-side viewer shows each draft the agent wrote and how the critique changed it. This is the reflection loop made visible: watch a weak first draft become a strong final one.
  • Open the trace — when a report links to its LangSmith trace, you can see every step the agent took to produce it.

2. Creating a lesson (capture → generate → review)

Creating lessons is the cost-incurring part, so it requires signing in. Today the app is single-user: only the configured admin address can sign in (a magic link is emailed — no password). If you are not the admin, the sign-in screen offers a waitlist instead, for when this becomes a product you can use.

  1. Start a capture at New capture. Paste your location transcript, set the target audience, and pick which model writes the draft.
  2. Generate — the agent researches the location, drafts an objectives-first lesson, critiques it against the rubric, and revises until it passes (or escalates to human review after a few tries). This runs synchronously; give it a moment.
  3. Review — open the new report and use the revision viewer to see how each critique cycle improved the draft. If the agent escalated, the report is flagged for your review rather than published blindly.

You choose the model per run (Claude or Gemini) so you can compare draft quality side by side.

3. Running the app (operators)

You do not need to be the developer to run this. Everything below is configuration, not code.

  • Local run — npm install, copy .env.example to .env.local, then npm run dev. No Docker is required.
  • Keys — the agent needs one model key (ANTHROPIC_API_KEY or GOOGLE_API_KEY). Every other key is optional — its feature fails soft when the key is absent. Each variable is documented in .env.example.
  • Admin console — at /admin you review waitlist signups and (under Model configuration) set the default provider and per-provider models.
  • Tracing — set the LangSmith variables to send run traces to a dashboard. The app runs fine without them.
  • Sign-in delivery — the magic link is sent through Mailgun; the admin address and mail keys live in the environment.

For the full setup, scripts, and project layout, see the README.

FAQ

Why can't I sign in?
The app is single-user for now — only the admin address is allowed. Join the waitlist from the sign-in screen and you'll be first in line when access opens up.
Do I need Docker?
No. The app runs with Node and npm; nothing here uses Docker.
What is the revision viewer showing me?
Each pass of the agent's write-then-critique loop. Reading down the revisions shows the draft improving toward the rubric — or, if it never passes, why it was sent to human review.

← Back to home