🍪 CRUMB: the guided-tour / AI-review layer

Tours as markdown, in. A guided walk, out. CRUMB is PROOF's twin: where PROOF renders plans-as-markdown as a board that never writes, CRUMB renders tours-as-markdown as a guided projection that writes only through the app's own door, and never submits. Published as @tjakoen/crumb and live: it's the guided-tour frame running on tjakoen.github.io today.

What's built

Core, routes, loader, crumb check, and from-timeline (turning grain's audit trail into a review tour) are all built and shipping in 0.1.4. crumb-live.js drives the lamp + popover client-side. What's still open is tracked in PLAN.md.

A tour can also end by asking the reviewer what the walk could not check, and hand back a prompt to paste into a session:

## prompt
- ask: reads-wrong | Anything here that reads wrong?
- ask: next | What should change?
- template: Continue the {title} review.\nWrong: {reads-wrong}\nNext: {next}
- handoff: https://claude.ai/new?q={payload}

Nothing is submitted: the answers compose into text in the browser, and where that text goes is the tour's own URL template to declare.

Quickstart

bun add @tjakoen/crumb
bunx crumb check tours/     # lint a tours/ folder

Mount it as a read-only layer alongside your app:

import { createCrumbRoutes } from "@tjakoen/crumb/routes.ts";

const crumbRoutes = createCrumbRoutes({ toursDir: "./tours" });
// mount crumbRoutes(pathname) inside your own request handler

Read next

• Docs: the full write-up.
• Live: the guided tour running in production.
• PLAN.md: the canonical plan (design, build pieces, what's left).

🤖 Built with Claude, and every tour is a projection, never a fork. I decided what "reviewed" has to mean, Claude typed the parser that enforces it. I don't prompt and pray, I prompt and prove. How I actually work with AI, receipts and all →