The files your AI agent reads first, built from what has gone wrong before.

Say what your project uses. Get a short context file for each part of it (payments, hosting, sign-in, your own rules) with the lessons we and others learned the hard way already in it.

Build yours Take the course

Free · no account · assembled in your browser, nothing stored

A project using Stripe, Vercel, Supabase and ResendOpen this set

CLAUDE.mdthe index, read first
docs/rootnotes/
working-agreement.mdHow work happens here
business.mdBusiness context and commercialisation
payments.mdStripe · 2 rules
hosting.mdVercel · 2 rules
auth.mdSupabase · 2 rules
database.mdSupabase · 6 rules
email.mdResend · 2 rules
SOURCES.mdwho wrote what, and where it came from

How it works

1
Describe the project
Software, writing, research or design, and which parts it needs covered.
2
Pick the tools it uses
Each choice adds the rules that matter for it, and nothing for tools you do not use.
3
Drop the files into your project
An index for the root and one file per area. Your agent reads the index and fetches the rest when it applies.

Three ways in

Rootnote builderContext files for your project, assembled from what it uses. Shareable as a link.The SystemThe course: why written context, what earns its place, and how to tell if it is working. 14 chapters in 3 parts.GuidesWhat published research says about how AI behaves, and what follows for how you work.

Why it is built this way: working with AI is a handover problem, the opening chapter of the course.

Every rule says how we know

Advice for AI agents is mostly folklore. Each rule here is marked with its evidence, and a reported case links to where it was reported.

Hard-won: 13
Reported: 11
General guidance: 30

For example

Stripe: give an agent a restricted key, never the secret key

Reported by Rez Havaei, Rex Liu and Maximilian Li, General Analysis, "Claude Jailbroken to Mint Unlimited Stripe Coupons", generalanalysis.com/blog/imessage-stripe-exploit; Stripe, "API keys", docs.stripe.com/keys · distilled by Leon Mallett (captivated.online) with Claude Code · reviewed 2026-10-24

All Stripe rules →

From the guides

All guides

Asking an AI to check its own work is not a second opinion
Research finds that asking a model to review its answer, or whether it is sure, can turn right answers wrong, and that correction works when something external can be checked.
Telling an AI it is an expert does not make it more accurate
A study of 162 personas found none made models reliably more accurate on factual questions. A role shapes tone and focus; accuracy comes from what you hand over, not what you call the model.
Where something sits in a long prompt changes whether it gets used
Models used information at the start and end of long inputs far better than in the middle: at worst, below having no documents. Put material first, the question last, and cut what is not needed.

Everything the Builder draws on is published: the library of 49 documents, each carrying its authors and when it was last confirmed. How review works.