Give ChatGPT helpful instructions without overloading every answer
"Keep it brief" is a useful preference until it turns a decision note into three vague bullets. Give ChatGPT room to match the job.
What you'll make
two or three lines of custom instructions that help across your work, and a quick test on two different tasks to show they don’t strip out detail you need.
What you'll need
ChatGPT in your work account, permission to change its personalisation settings, and two practice tasks: a short rewrite and a structured note. About 15 minutes. If you can’t change settings, test the wording at the top of a chat first.
Example result.
Use plain British English and explain unfamiliar terms when they affect an action.
Be concise, but include the detail needed for the task's purpose and checks. Follow the requested format rather than forcing every answer into bullets.
Keep stated facts, uncertainty and commitments intact. Flag missing information instead of filling it in.1. Write preferences that should survive the next task
Custom instructions are preferences ChatGPT follows across chats: plain language, spelling, how to show uncertainty. They’re a standing brief, not a replacement for the current job.
Pick two or three stable preferences. “Use British English” helps across many jobs. “The launch is on 12 November” is a changing fact and belongs in the task. “Always use five bullets” is a format rule that will damage a table, a calculation or a letter.
Weak: “Be extremely concise. Always use bullet points. No more than 100 words.” It reads like good discipline, and it’s why your decision notes keep losing their conditions.
Write each preference with room for exceptions, as in the example result. Leave out personal or confidential details the preference doesn’t need. Then copy your old wording somewhere safe, open Settings › Personalization, paste the new lines into Custom instructions and save.
Why this works: “concise within the requested structure” lets a one-line rewrite stay short without letting a decision note drop its conditions.
What you should see: your new instructions saved, and a copy of the previous wording you can restore.
2. Put the current requirements in the task
Start a new chat. State the source, audience, result and depth for this job. Put dates, prices, policies and approval status in the task, next to the source. A preference can’t tell ChatGPT which version of a document is the current one.
Use the source below to create [output] for [audience]. This task needs [required sections or level of detail]. Keep the writing concise within that structure. Preserve facts and uncertainty, and list missing information separately. Use only the source for factual claims.
Source:
[current source][output] and [audience] name what you want and who it’s for. [required sections or level of detail] might be “Proposal, Conditions and Open questions”. [current source] is the pasted text.
Why this works: a required structure stops “concise” from swallowing the explanation.
What you should see: a response with every section you asked for, at a length that suits the job.
3. Test a short job and a detailed job
Run a one-sentence rewrite and the structured task, and read both against the same preferences. Does the short one sound natural? Does the detailed one contain every required section and qualification? Don’t rely on ChatGPT saying it followed your instructions. Read the answer.
Why this works: an instruction that only fails on long tasks won’t show up if you only test short ones.
What you should see: a short answer that’s still accurate, and a detailed answer that’s complete.
Check the result
- The short answer keeps the status and conditions in the source.
- The detailed answer has every required section, including open questions.
- Nothing in either answer is a fact the source doesn’t contain.
- Your instructions say nothing that only applies to one task.
The usual failure: brevity eats a condition. “Subject to approval” vanishes, or the open questions section is quietly dropped, because the standing instruction said to keep things short. Spot it by checking the detailed answer against the sections you asked for.
The answer is missing [required section or qualification]. My preference for concise writing doesn't remove that requirement. Restore it using the source, keep the other sections, and flag any information the source doesn't contain.[required section or qualification] is what disappeared. If the same failure keeps coming back, simplify the custom instruction rather than adding exceptions around it.
Worked example: concise doesn’t mean incomplete
Fictional example. The support team lead is considering a six-week trial of replacing the Thursday stand-up with written updates. The source says the support manager must approve it, nobody has been named to collect the updates, and no start date or tool has been chosen.
The short task asked for one sentence. The answer: “A six-week trial of written updates in place of the Thursday stand-up is proposed, subject to the support manager’s approval and a named owner.”
The detailed task asked for a note with Proposal, Conditions and Open questions. Under the old “Be extremely concise. Always use bullet points.” instruction, it came back as:
- Replace Thursday stand-up with written updates for six weeks.
- Starts next Thursday.
- The failure. The conditions and open questions sections are missing, and “starts next Thursday” is invented.
- The repair. With the new instructions saved and the correction prompt naming “the Conditions and Open questions sections”, the note came back with Proposal (six-week trial of written updates), Conditions (manager approval, a named owner) and Open questions (start date, tool, who reads the updates).
- Same checks, both lengths. The one-sentence version and the full note keep the same facts and the same unknowns. Only the shape changes.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Optional training: override a preference for one task
Your instructions say “Use bullet points where helpful”. Today you need a customer apology letter, where bullets would read like a form. Override it in the task, not in Settings:
For this letter, ignore my bullet-point preference and write three short paragraphs: what went wrong, what we've done, and what happens next.A one-line override is quicker than editing your instructions and remembering to change them back.
Optional training: keep team guidance with the team's work
Custom instructions suit your own preferences. Guidance for a workstream, audience or set of sources belongs with that work. In ChatGPT, projects can carry their own instructions, so a shared house style lives in the project rather than in someone’s personal settings. Use ChatGPT Projects to keep one piece of work out of your other chats and Review assistant memory and choose what to keep cover both. Settings › Personalization › Base style and tone changes how ChatGPT sounds, not what it can do.
Optional training: test one change at a time
Keep the same two practice tasks whenever you revise your instructions. Change one preference, start a fresh chat and compare against the same checks. “The detailed note now keeps approval status” is a checkable observation. “This instruction makes ChatGPT accurate” is a much bigger claim than two tests can support.