Notes · 11 June 2026
Setting up the Context Kit, and the three traps that break it.
The free Context Kit is now live. It is six plain markdown files: the folder structure our paid setups are built on, extracted from a system that runs daily inside a real business. Since v1.1 the kit also carries a guided start: a read-this-first PDF and a setup prompt you paste in, which then builds the workspace and runs the interview itself. Since v2.0 it works the same in Claude or ChatGPT. The first note covered why the structure exists. This one covers how to set it up well, because the difference between a context system and a folder of old notes is entirely in the setup and the upkeep.
What the six files do
The kit opens with a Start Here file that walks you through everything below. The working parts are five: a workspace constitution, which becomes the AGENTS.md file your AI reads at the start of every session and holds the standing rules, with a three-line CLAUDE.md beside it that points Claude to the same place; a business context file, the living memory of where things stand, with a decisions log inside it; an entity context file template, one per client, supplier or project that matters; a working preferences file, which is how you want your AI to write and work; and a maintenance schedule, the routine that keeps all of it true.
Everything is markdown in a folder you own. It works in Claude Cowork or ChatGPT Work, and in Claude Code or Codex if you are technical. If you ever switch between them, or stop using both, the files are still yours and still readable.
The setup is an interview, not a form
The whole setup takes about 30 minutes, and the most important instruction in the kit is the one people skip past: do not fill the templates in by hand. Have your AI do the whole thing. The kit ships the exact prompt as a file (07 - Setup Prompt): give it your folder (connect it in Claude Cowork, or add it to a project in ChatGPT Work), paste the prompt in as your first message, and it builds the workspace, handles the renames and interviews you, one question at a time, until it can complete the files properly, showing you each finished one before saving.
Twenty minutes of answering honest questions beats an hour of staring at brackets. It asks the obvious follow-ups you would not think to write down, and the files come out in full sentences rather than half-filled placeholders.
Then check it worked. Open a fresh session on the same folder and ask: what do you know about my business, and where did you learn it? If the answer names your files and gets the facts right, the system is live.
The three traps
Almost every failed setup we have seen fails one of three ways, and none of them announces itself.
The double extension. The constitution only works if it is named exactly AGENTS.md. Most computers hide file extensions, so renaming the template can silently produce AGENTS.md.md, which gets ignored. Nothing errors; the file simply never gets read. Turn on show file extensions before you rename.
The word processor. Markdown files opened and saved in a word processor come back wrapped in formatting that breaks them. Edit them through your AI, or in a plain text editor if you must. The kit assumes your AI does the editing, which sidesteps the problem entirely.
The stale copy. There should only ever be one live copy of a context file. When the facts change, the same file is renamed to the new date, not saved as a fresh version next to the old one. Old copies left in the folder leak stale facts into your AI's answers with complete confidence, which is worse than no context at all. The history belongs in the decisions log inside the file, not in the folder listing.
The habit that carries it
The constitution in the kit instructs your AI to ask one question at the end of every working session: did anything material change today, and should it go into the context files? Your half of the habit is not to wave it through. It proposes the exact lines, you approve or amend, and the file's date gets bumped. Two minutes, every session. That single habit is why the system stays current instead of quietly rotting, and it is the part the maintenance schedule exists to protect.
One boundary worth restating: context files hold facts, terms, history and decisions. They never hold passwords, API keys or account numbers. Prices and payment terms are fine; anything that unlocks an account is not.
Getting it
The kit is free; request it on the homepage and the download link arrives by email straight away. If you would rather build the system with someone, that is exactly what the Working Setup is: a scoping call, a working session on your real files, and a finished setup handed over. The kit and the paid setup are the same architecture. The difference is who does the work.
Daniel · The Clarion