cheatsheet
for AI agents that use the web

Nav guides for AI agents, synthesized on demand.

Cheatsheet indexes how websites actually work and serves tight nav guides to your agent — ideally a single URL template, falling back to a minimal UI sequence. Stop burning tokens on trial-and-error DOM parsing.

How it works

Two arrows: writers index how the web actually works, readers ask for a guide. Every outcome closes the loop.

WRITE: crawler agents → atomic notes → clustered → indexed
READ:  customer agent → ask() → retrieve clusters → synthesize → envelope
                                                                    │
                                                                    ▼
                                                            report_outcome()
                                                            (feedback loop)

Crawlers write notes

Playwright agents survey sites and emit atomic observations — URL params, form rules, nav transitions. One fact per note.

Clustering builds consensus

Near-duplicate notes are grouped. Each cluster carries a consensus summary, a confidence score, and a TTL by kind.

Hybrid retrieval on ask()

Vector + keyword search over clusters, filtered by site, scope, and variant. Top-15 clusters feed the synthesizer.

Synthesized envelope

Claude composes a structured envelope: a fastest_path URL when possible, a UI procedure fallback, plus probe hints and provenance.

The four primitives

A small vocabulary that adds up to a self-correcting nav-guide service.

Note
One atomic observation about how a site works. Prose text, structured anchors, provenance, embedding.
Cluster
A group of near-duplicate notes about the same fact. Has a consensus summary and a confidence score. Retrieval queries clusters, not raw notes.
Envelope
The synthesized output. An optional URL shortcut, an always-present UI procedure, and provenance — model, source clusters, freshness.
Outcome
Your agent reports whether the envelope worked. Successes raise cluster confidence; failures shrink TTLs and trigger harvest-mode recrawls.

Try it on a real site

POSTs to /v1/ask through a server action. You'll get back an envelope with a URL shortcut (when one exists) and a UI procedure fallback.

ask()

Two fields. Pick a site we've crawled and describe a task.

envelope

Server response will render here.

Submit a task to see a synthesized nav guide.