Unit 2 · Set it up once
Standing instructions
The principle
A standing instruction holds only what stays true when today's task disappears — everything else is today's request wearing a permanent badge.
By the end: You can sort your usual context line by line into a standing instruction and a one-off request, then run a real task with the request reduced to one sentence.
The situation
In February you finally wrote it down. Twenty lines of saved instructions: your role, your readers, the 250-word limit, the house words, the rule about not inventing numbers. For three weeks everything came back in the right shape. One of the twenty lines said the priority this quarter is consolidating the two northern branches.
It is now June and the consolidation finished in April. The supplier email opens with a paragraph about consolidation. The agenda you asked for lists it second. When you ask for next quarter's risks, consolidation is one of them. You start adding "ignore the consolidation line" to what you type, which is the preamble coming back one sentence at a time.
Nothing is broken. You told it to care about something, it kept caring, and you had no rule for which lines were allowed in there.
The principle
A standing instruction holds only what stays true when today's task disappears — everything else is today's request wearing a permanent badge.
That is a test you can run on a line in about two seconds. Delete the task in front of you. Read the line again. Still true? Standing. Not true, or not relevant, or true only this once? It goes in the request.
Two things follow. A standing instruction is a boundary, not a biography — two kinds of work with contradicting standards are two workspaces, not one instruction with "when I'm writing X, do Y" branches inside it. And every line is read on every request, so a line that changes nothing competes for attention with the lines that do. The one-off half carries what is genuinely specific — this file, this week, this exception — and it gets shorter over time while the standing block barely moves.
Worked example
Shown here in ChatGPT.
The block from the last lesson, unchanged. Now sort it.
Who you are writing for ..... regional ops lead, 14 branches, building
products. Readers act the same day.
Default output .............. plain text, under 250 words, movements
first, no preamble, no closing offer
House vocabulary ............ branch = location. account = customer.
unit = a pallet, never a single item.
Handling numbers ............ only figures present in what I give you;
otherwise "not in source"
Never ....................... the three prohibitions, verbatimEvery line there survives the deletion test. The work is in the six that do not, and in where each one actually belongs:
| Line someone wrote | Where it goes | Why |
|---|---|---|
| "This quarter we are consolidating the northern branches" | One-off | It survives the deletion test, but only until April — and a standing instruction has no expiry date |
| "Make this one more formal, it's going to the board" | One-off | "This one" is the giveaway |
| "The attached CSV has one row per branch" | One-off for now | It describes a file, not a rule — unless that file is always the input, in which case it belongs to the source |
| "Write plainly. No consultancy register" | Standing | True of every output you will ever ask for, and checkable: you can point at the sentence that breaks it |
| "Do the arithmetic in a spreadsheet, not in prose" | Standing | A tool preference is a standard, not a task. It fails only if some work here wants the opposite |
| "Check the branch reference list before you answer" | Neither | A reminder to attach a file every time is setup you have not done. Connect it once — next lesson |
Row one is the interesting one. It passes the truth test and still fails, because standing instructions are never re-read once they work. Anything dated stays out.
In your tool
The saved instruction exists nearly everywhere. What differs is its scope — what it silently applies to.
| Tool | What a saved instruction applies to |
|---|---|
| Claude | Project instructions apply to every chat started inside that Project, and to nothing outside it |
| ChatGPT | Custom instructions apply across the account; project instructions apply inside one project and sit on top of them |
| Gemini | A Gem applies only to chats you start from that Gem, so the boundary is whichever Gem you open |
| Microsoft 365 Copilot | A published custom agent carries the persistent instruction and applies to whoever it was shared with — on a standard seat an administrator publishes it rather than you |
Product surfaces checked 2026-08-04.
Try it
Take the longest preamble you wrote in the last lesson, for your Unit 0 task. Put S, O or D beside every line — standing, one-off, delete. Paste the S lines into your tool's saved instruction for one workspace.
You are done when the standing block contains no date, no project name, no "this one", and no description of a specific file — and when one real task runs from a single sentence naming the job and still comes back sendable.
Common failures
- The dated line. A quarter, a project, a current priority. It reads as useful context on the day you write it, is quietly wrong six weeks later, and nobody re-reads a saved instruction that is working.
- The personality brief. "You are a seasoned consultant who thinks strategically." Nothing in that sentence can be checked, so nothing in it can come back wrong, so nothing in it changes the output — except by spending attention your format lines needed.
- One workspace for everything. The instruction holds both "under 250 words, plain" and "full prose with sources cited". The model splits the difference and you get 600 words of half-sourced prose.