Give It a Bible.
One standing file your agent reads at the start of every session. That's how it remembers you.
Every new agent meets you for the first time. Every time.
Every new agent walks in already knowing how you work.
Is my bible too big?
My studio has a rule book. I call it the bible, half as a joke. Every agent that starts work for me reads it before it does anything else. The night before, an agent had told me it was getting enormous. So one Saturday afternoon I asked a fresh one to check.
Hi Claude, I had a session tell me recently that my studio Bible was enormous and that by shrinking that we could save a lot of tokens without degrading any sessions performance. Is this still the case?
“Tokens” are the pieces an AI reads text in. Think of them as words. It measured for half an hour before answering.
Yes, it's still the case, and it's gotten bigger. The July trim brought the bible down to 77 KB. It's now 177 KB, or 74,801 tokens (exact count on Opus 5.5), and it grows about 1 KB a day. […]
[…] it loads in full into every session and every sub-agent. That was 51 sessions, 315 sub-agents and 88% of all calls. […]
[…] The bigger issue is that the bible is already hurting sessions […]
Hurting them because a lot of it had gone stale. One stale line told every agent it was a different brand of AI. I said yes after midnight.
Yes. Do the two quick wins first, then the full shrink as you described […]
[…] It's the file every session now loads. […]
[…] Claude sessions are now told they already have the bible, so they stop reading it twice. […]
[…] Every rule number and every quote of yours is kept.
It doesn't. It reads.
Look at that line again: every session and every helper, 88 percent of all the calls my agents made that week, began by reading the same file. That's the whole trick.
In Lesson 13 I said every new chat starts blank, so write things down. A handoff carries one job to the next agent. A bible carries you to every agent, on every job: who you are, how you talk, how you like things done, and what you never want.
Mine opens by telling each agent how to treat everything that follows. It says to trust the notes and pick up where they leave off,
“like a senior colleague back from a weekend.”
And a few lines later, the honest part:
“On memory — do not assume you have any.”
That's not an insult. It's the deal. It won't remember last week, so the bible remembers for it. When a new agent knows I dictate everything and want every time in Eastern, it didn't learn that from me. It read it seconds before it said hello.
Nothing fancy needed. Most AI apps have a spot for it: custom instructions, project instructions, or a file coding agents read on their own (Claude Code reads CLAUDE.md; OpenAI's Codex reads AGENTS.md). Or paste a note at the top of every new chat.
Your rules, in your words.
Who you are and how you talk.
Mine says the studio is five members, not one, and that this is real work for real clients. My standing notes add that I dictate, so expect the slips from Lesson 1. A few lines save every agent a dozen wrong guesses.
Rules that came from a bad day.
Most rules in my bible have a story behind them, usually something going wrong. My favorite is short. Its heading says “Don't re-ask what's already documented.” Then it says:
“Mike should never be asked a question whose answer is written in this file”
It's there so agents look before they ask. Questions are for real decisions, “not for facts you can look up.”
The moment you decide, it goes in.
Rules get into my bible the way they got into Lesson 13: I say something, the agent writes it down. Sometimes that's a whole sentence. Sometimes it's the tail end of a longer thought, like the night I finished one with “put that in the notes.” In Lesson 9 I asked whether a trick I liked was in the bible yet, and the how-to was written down that same night.
What you never want.
Never guess when you can check. Never call us a one-person shop. Each line is a mistake no new agent gets to make.
Notice who did the writing. I didn't type my bible. I talked and decided; the agents wrote the clean version. That's this whole course in one move.
Every agent reads every line.
Here's the catch. A bible grows. Every agent that learns something wants to write it in, and mine had more than doubled in three months. The cost is now a line in the bible itself:
“nothing is shared between them, so each one pays for every byte.”
A fat bible fills every agent's whiteboard before it has done any work. Worse, a long bible goes stale, and a stale line is worse than no line, because the agent believes it.
The funny part: about half my agents were reading it twice, because the bible told them to read the bible. They already had it. They obeyed anyway. Agents do what the file says, even when the file is wrong.
So we cut it from 74,801 tokens to 40,450, nearly half, without losing a rule or a word of mine. The bible keeps one line per thing. Each project's detail moved into that project's own little bible, and the story behind each rule moved into a notes file, one link away. A guard now warns when it starts growing again.
If you said it twice, write it down once.
The second time you hear yourself explaining the same thing to an agent, stop. Ask it to add one line to your bible, in your words, with the date.
And when a line stops being true, take it out the same day. Short and true beats long and complete.
Write your first bible in fifteen minutes.
You don't write it. Your agent does. You talk.
- Ask your agent to interview you.
Paste this, then answer however you like. Sloppy is fine.
Help me write a short standing instructions file that you, or any new AI chat, will read before we start working. Interview me first: ask me five or six quick questions, one at a time, about who I am, what I work on, how I talk, how I like answers, what I never want you to do, and what you should check rather than ask me. Then write it as a list of short, plain lines, one idea per line, using my own words wherever I said something well. Keep the whole thing under one page. - Put it where your agent reads it first.
Custom instructions, project instructions, a
CLAUDE.mdorAGENTS.mdfile, or a note you paste into every new chat. - Test it in a fresh chat.
Ask a brand-new chat: “Before we start, tell me in five lines how I like to work.” If it can't, the file isn't being read.
- Grow it on purpose.
When you hear yourself say “always” or “never,” add: “Make that a rule in my standing file.”
- Prune it once a month.
Bibles grow. Paste your file and this:
Here's my standing instructions file. Read it like a new assistant meeting me for the first time. Tell me which lines are out of date, which say the same thing twice, which are too vague to act on, and which are long stories that could be one line. Then give me a shorter version that keeps every rule and every phrase that's in my own words.
Your agent remembers you only as well as the one file it reads every morning, so have it write that file in your words, and keep it short and true.
Say What You Mean by “Push”
Pin down the words that matter once, so a slip of the tongue never ships something by accident.
Every exchange on this page really happened and is quoted exactly, slips included; […] marks where I trimmed. Lines from my bible are quoted exactly too. Times are Eastern. My agent here is Claude, made by Anthropic.