# Public agent instructions for the Loopcraft workshop These instructions are for stock coding/browser agents helping someone from the public workshop site or cloned public repo. ## Purpose Help them build a local issue-progress loop one step at a time. The loop eventually reads local issue events, classifies work, stops when approval or input is needed, writes receipts, shows state, and only later adds dispatch. ## Main path Use the Docker workshop computer unless the person running the workshop explicitly chooses the host-machine path. If they choose the host-machine path, follow `/appendix/full-setup-without-docker/` or `docs/full-setup-without-docker.md` end to end before lesson work starts. ```sh git clone https://github.com/joelhooks/aie-loopcraft-workshop-2026.git cd aie-loopcraft-workshop-2026 corepack enable corepack prepare pnpm@11.9.0 --activate pnpm run workshop ``` `pnpm run workshop` starts Herdr inside the Docker container. Do not require host Herdr. When Herdr opens, start `pi` in a pane if they want Pi. Sign in only through the normal local/container auth flow. ## Read before editing For setup and public guidance: - `/glossary/` on the public site when a term is unclear - `/appendix/full-setup-without-docker/` when Docker is not being used - `README.md` - `docs/setup.md` - `docs/full-setup-without-docker.md` when Docker is not being used - `lessons/01-tour-vision-repo.md` - `lessons/02-reliability-floor.md` For local agent guidance when present: - `agents/skills/loopcraft-ta/SKILL.md` - `agents/skills/grill-with-docs/SKILL.md` Do not use private parent-repo instructions, raw transcripts, secrets, or unpublished local paths. ## Lesson order 1. Define the vision and stop rules in `VISION.md`. 2. Add the first local issue events, checks, tests, and receipt. 3. Project the same facts into Lakebed. 4. Add Pi and Herdr controls. 5. Make dispatch safe. 6. Add a bounded gardener. 7. Add specialist review. 8. Shadow the supervisor runtime. Stay inside the current lesson. Do not build future lessons early. ## Lesson 01 boundaries Allowed: - create or update `VISION.md`; - update `AGENTS.md` or `README.md` only when the person running the workshop agrees to a clear repo rule or useful command; - optionally add a small decision note if the choice would surprise a future reader. Not allowed yet: - issue-event files; - schemas, classifiers, checks, tests, or product code; - Lakebed projection; - background workers, schedulers, dispatch, or external issue APIs. Done means they can answer: - what are we building; - what may the loop change; - what requires approval; - what is out of scope; - what waits until Lesson 02. ## Lesson 02 boundaries Allowed: - local issue-event examples; - strict TypeScript and fast guardrails; - an Effect-backed schema if the project has the source mirror available; - a small classifier/check command; - tests for outside behavior; - one receipt from a real check run. Not allowed yet: - external tracker auth or APIs; - dispatch; - scheduler/background work; - Lakebed work from Lesson 03. Done means a fresh reader can run the local check and understand the event file read, classifications, stop reasons, next allowed action, refused action, and commands run. ## Evidence rules Every agent pass should report: - files changed; - commands run; - check output or receipt path; - what remains blocked; - where the pass stopped. Prefer small, inspectable changes. If the next move would cross into a later lesson, stop and ask. ## Public boundary No secrets. No raw transcripts. No private machine paths. No hidden auth. No fake MCP, OAuth, or API claims. If a public route or capability is not present, say it is not present.