An AI operating system template for Claude
The folder tree is the easy half and people over-invest in it. The part that decides whether this works is three files that have nothing to do with your notes: the standing instructions, the routines, and the corrections.
The three files that make it a system
| File | Its job | Why it matters |
|---|---|---|
| Standing instructions | Read at the start of every session. Where things go, what is never done without asking, how to address you. | Without it you re-explain your system every session, and the assistant makes a different judgment call each time |
| Routines | Each recurring job written once and invoked by name: weekly review, morning brief, follow-up sweep | A routine in a file improves permanently. A routine you retype degrades every time. |
| Corrections | Every steer you give, written back as a rule | This is the only mechanism that makes the system better rather than merely consistent |
Notice that none of these are notes. They are the operating layer, and they are what separates an assistant that happens to have file access from a system.
The tree, briefly
hub/ CLAUDE.md the standing instructions, read every session skills/ one file per routine, invoked by name projects/ one index file per project: status, decisions, next action people/ one file per person: history, what is owed either way decisions/ what was chosen, and why reference/ reading, and derived views that are never hand-edited
That is deliberately close to the tool-agnostic template, because the structure should not depend on the assistant. What is specific to running it on Claude is the top two lines.
What goes in the standing instructions
Keep it short, and keep it to things that are true on almost every turn. Where a new piece of material gets filed. Which acts are never taken without your approval, which for most people means anything that leaves the machine: a sent email, a payment, a published post. How to tell you something is uncertain rather than guessing. Where the routines live.
The failure mode is a standing instruction file that grows into a manual. Anything situational belongs in the routine that needs it, not in the file read on every single turn, or the important rules get diluted by the incidental ones.
A rule worth adopting early: when you correct the assistant, do not just accept the fix. Write the correction into the instructions in the same session. Otherwise you will give the same steer next month, and the month after.
What goes in a routine
Everything the job needs to run without you re-specifying it: which files to read, in what order, what the output looks like, and what it must never do. A weekly review routine that says "read every project's index file, list what changed, flag anything untouched for three weeks, draft nothing outbound" produces the same quality of review in a distracted week as in a good one. That consistency is most of the value.
Concrete versions of the common ones are at the AI weekly review, the daily brief and project status tracking.
Why this shape and not a chat window
Because all three files only pay off if the assistant can read and write the folder directly. In a chat interface you are the one carrying information between the conversation and the record, which is the maintenance tax that kills systems. The reasoning in full is at a personal OS built with Claude Code, and the tool comparison at Claude Projects versus Claude Code.
The template and every prompt that fills it
The book builds all three files and the four tiers above them, hands-on. Every prompt is printed, and you write no code.
Get the bookSee it running
Standing instructions, routines and corrections, driving a record kept current from real email and meetings. Try the demo.
See the demo