Stop retyping yourself

Count the times this week you’ve typed some version of: plain English, no bullet points, I work in logistics, our customers are hospitals, don’t invent numbers.

Every one of those is a standing fact about you or about your job. None of them change between Tuesday and Wednesday. You’re paying for them by the message.

Most tools have somewhere to put them. Almost nobody who wasn’t shown it uses it.

---

Three layers, and what each one is for

Where it lives	Applies to	Best for
In the message	This one answer	Anything about today’s specific task
A project or space	Every conversation inside it	The job: audience, product, source material
Account-wide settings	Everything you ever send	You: how you want to be written to

The middle layer goes by different names. Projects, spaces, Gems, custom GPTs. The idea underneath is the same in all of them: a container that holds instructions and reference files, and hands them to every conversation you start inside it.

Some tools also have an automatic memory that decides for itself what to keep from your conversations. That’s a fourth thing, and it’s covered below.

---

The test for which layer

One question sorts almost everything:

Is this true of me always, or only for this job?

Always goes account-wide. Only-for-this-job goes in the project.

Get it backwards and you get strange results months later. Put “write in our brand voice, upbeat and punchy” in account-wide settings and it will show up in the email you send to your solicitor.

---

What belongs account-wide

Keep this short. It’s read every single time, including on jobs it has nothing to do with.

The useful content is how you want to be dealt with, not what you work on:

HOW TO WRITE TO ME
Plain English, short sentences. No preamble and no closing summary
that restates what you already said.
Answer first, then the reasoning if the reasoning is needed.
If my request is ambiguous, ask me one question instead of guessing.
If something isn't in the material I gave you, write "not stated"
rather than estimating it.

WHAT I DO
[your role, your industry, the kind of people your work goes to]

WHAT I ALREADY KNOW
[the background you don't want explained back to you every time]

That last section pays for itself faster than the rest. Without it, a specialist gets an introductory paragraph about their own field at the top of every answer.

Two lines above are worth keeping whatever else you change. “Ask me one question instead of guessing” turns silent gap-filling into a question you can answer. “Not stated” turns an invented number into a visible hole. Both are the standing version of habits from steps two and three.

---

What belongs in a project

Everything specific to one body of work:

• Who the audience is, in the detail step two asks for
• The vocabulary your organisation uses, and the words it doesn’t
• Reference documents, attached to the project rather than re-uploaded
• Rules that apply to this work and would be wrong elsewhere

The attached-documents part is the one people underrate. If you re-upload the same style guide, price list, or product spec every few days, that file belongs in the project, once.

---

What this is not

It isn’t learning. Nothing improves through use, and the tool doesn’t come to know you.

What’s happening is more ordinary: you wrote a note, and the note gets read at the start of every conversation. The same way a temp reads the card taped to the monitor. Useful, entirely dependent on the card being accurate, and it does nothing at all if you write something wrong on it.

That’s also why the previous guide still applies. Standing instructions get handed over at the top of a conversation, which puts them in exactly the position that a long thread crowds out.

---

When it breaks

Everything is coming back oddly and you can’t work out why. Check your standing instructions first, before you rewrite the prompt. This is the failure mode of the whole feature: the cause is invisible, sitting in a settings screen you last opened in March, and it never occurs to you to look there. When a tool starts behaving strangely across every conversation at once, the cause is almost always something that applies to every conversation at once.

It ignores standing instructions on long jobs. Same crowding described in step four. For the one rule that really matters, repeat it in the message, at the end.

Your project voice is turning up in unrelated work. A job-specific rule ended up account-wide. Move it.

A colleague runs your exact prompt and gets a different answer. Standing instructions are per person, and they’re invisible in the prompt you shared. When you hand a prompt to someone else, hand over the relevant standing instructions with it, or paste them into the prompt itself.

Automatic memory has stored something wrong. These features save things you said in passing, including things that were true that week and aren’t now. Find the list, read it, delete what’s stale. Worth doing occasionally even when nothing looks broken.

You set them and nothing changed. Some tools apply standing instructions only to conversations started afterwards. Open a new one and try again.

---

What to take from this

Ten minutes, once, and then a short review whenever something feels off.

The trap is that this feature is genuinely useful and completely silent. Everything you put in that box shapes every answer you get, and none of it appears on screen while you work. Write it carefully, keep it short, and know where the box is when the answers start going strange.