For example, an operations coordinator might repeatedly need short internal updates with the decision first and unconfirmed dates left open. Save that writing preference once, then test it with a small request whose answer you can check.
Find the instructions in personalization
Open ChatGPT's settings and look under Personalization for custom instructions. Available controls differ between web and desktop. OpenAI describes custom instructions as preferences that apply across chats. Use the controls for the ChatGPT surface you are actually using. OpenAI's personalization documentation
You do not need to choose a new personality to write a specific instruction. Describe the result you need in ordinary language, then save it using the available settings.
Write a default that survives a change of project
This original example is suitable for someone who writes internal operational notes. Adapt the audience and habits to your work before saving it.
I use ChatGPT to prepare internal operational notes for colleagues.
Use plain American English. Put the decision or next action before
background when the task calls for an update.
Preserve names, quantities, and dates from the material I provide.
If a needed fact is missing, leave a visible placeholder or ask a
focused question. Do not invent results, approvals, or commitments.
Use headings or tables when they make the answer easier to read.
Keep simple answers in ordinary paragraphs. Explain an unfamiliar
technical term when it affects what the reader must do.
Each paragraph has a job. The first identifies the usual reader. The second gives a rule for incomplete evidence. The last prevents a formatting preference from turning every answer into a form.
Compare that with a default such as “Always produce a five-column table and an executive summary.” It may suit a weekly report and make a two-sentence email awkward. Scope the preference: “For comparison requests, use a table when the options share useful comparison criteria.”
Keep today's facts in today's prompt
The meeting time, client name, and document version change with the task. Put them beside the source material you want ChatGPT to use. OpenAI's prompting documentation recommends keeping preferences that should apply across chats in personalization and details specific to the current chat in the prompt. Where to put reusable preferences
For an equipment trial, the immediate request could be:
Draft an internal update from these fictional facts:
- A checkout trial is planned for Tuesday.
- Two volunteers will participate.
- The shift lead has not confirmed the start time.
- No trial results are available.
Ask colleagues to wait for the confirmed time before changing their
routine. Keep the draft under 90 words.
An acceptable result would say that the trial is planned, that the start time remains unconfirmed, and that colleagues should wait. A claim that the process has already reduced delays would fail because the supplied facts contain no results.
Test one instruction before adding more
After saving your default, start a fresh chat and run a small request like the trial example. Read the response against the rules you wrote. Did it name the next action? Did it invent a start time? Did the structure suit the length?
If a rule fails, make it more concrete. “Be professional” leaves room for interpretation. “For internal updates, name who must act and leave unconfirmed dates visibly unresolved” gives you something to inspect.
Change one preference, then repeat the same test. This is a practical comparison of the outputs, not proof that you can identify every internal influence on the answer. If a task needs a different tone, state that requirement clearly in its prompt and inspect the result.
Revise the saved default when your work changes
Return to personalization to edit or remove instructions that no longer fit. Keep a copy of a useful version if you want to compare it with a replacement. Recheck with the small test before relying on the new default for a long document.
If an outdated detail keeps returning after the instructions are corrected, review ChatGPT Memory separately. Memory and explicit instructions are different places to inspect. Finish with a default you can describe and a short output that demonstrates the behavior you wanted.