AI Guides › Workbench
By Nigel Guy · 6 min read
Most people ask ChatGPT for "a product description" or "a LinkedIn post for my business" and get something that could belong to any competitor. The model is not being lazy. You gave it a task and no facts, so it filled the gaps with the most average version of your industry. Rewriting the output by hand afterwards feels productive while it fixes the wrong end of the problem.
The rule: before any copy gets written, make the model interview you, then lock the answers into a brief it must reuse.
The mechanism is one opening instruction, a short batch of questions, and a saved brief. This guide gives you the instruction, the places you can keep the brief, and the traps. It does not promise "forever": a chat has limits, and what persists depends on where you save the brief.
The same brief can be stored in four places. They behave differently, so pick on purpose. Details below are from OpenAI's help centre as checked on 2026-10-04; menus and plan limits change, so look at your own screen if something differs.
| Place | What it does | Cost / limits (check current) | Best for | Catch |
|---|---|---|---|---|
| Opening message in a chat | Applies to that one conversation | Included in every plan, including free | Trying the brief before you commit | Gone when you start a new chat |
| Custom instructions | Applied to all your chats once enabled | Available on all plans; 1,500 characters on Free and Go, 5,000 on Plus, Pro, Business, Enterprise and Education | Short facts that are always true about you | Applies to every chat, including unrelated ones |
| Project instructions | Applies only inside one project | Projects are included in qualifying plans; file limits are 5 (Free), 25 (Go, Plus), 40 (Pro and business plans) per project | A brief per business, client or product | Overrides your global custom instructions inside that project |
| Saved memory | ChatGPT keeps details it is told or decides to keep | Included; controls vary by plan and region | Light preferences | You do not fully control what it keeps or how it is applied |
For plan prices, check the £ price on OpenAI's pricing page at checkout. I have not stated figures because they could not be verified on the day of writing.
Paste this at the very start of a new chat, before you ask for anything.
You are a senior copywriter and brand strategist who works from facts, not assumptions.
Context: I run [BUSINESS_NAME], which [WHAT_YOU_SELL] for [WHO_BUYS]. I will later ask you for [KIND_OF_CONTENT_YOU_WANT, e.g. product descriptions, emails, social posts].
Before writing anything, interview me. Your goal is to collect enough specific facts that the copy could not be mistaken for a competitor's.
Rules:
1. Ask no more than [NUMBER_OF_QUESTIONS, e.g. 10] questions, in groups of five at most, and wait for my answers between groups.
2. Cover: who the customer is and what they are worried about; what we do differently and what proof exists; price positioning; tone, including words we never use; real examples of copy I like or dislike; and anything that must not be claimed.
3. Prefer questions that need a concrete answer (a number, a story, a phrase) over yes/no questions.
4. If my answer is vague, ask one follow-up before moving on.
5. Do not write any copy, and do not invent details about my business.
When I say "done", summarise everything as a one-page brief under these headings: Customer, Offer, Proof, Tone, Banned words, Do-not-claim list, Open gaps. Mark anything I did not tell you as "unknown". Before you reply, check that every line in the brief came from my answers.
Fill in the square brackets. The cap on questions matters: without it some models keep asking, and you stop answering well.
Give facts you can stand behind: the exact complaint a customer made, your real price band, the three words you would never use. Paste a past email you like. If you do not know an answer, say "unknown". An invented answer here becomes confident copy later.
When the brief appears, read it and correct it. Then run a test in the same chat:
Using only the brief above, write [ONE_SMALL_PIECE, e.g. a 60-word product description for PRODUCT_NAME]. Then list which brief lines you relied on, and flag anything you had to assume.
The "which lines" check is the useful bit. It shows you whether the model is using your facts or reverting to generic filler.
Copy the brief out of the chat and put it somewhere persistent.
The brief becomes stale. Prices change, you drop a product, your tone shifts, and the model keeps writing to the old brief with total confidence. Put a review date at the top of the brief and re-run the interview when it passes.
A second trap is clashing instructions. Project instructions override your global custom instructions inside that project, and memory can add context you forgot you gave. If output sounds odd, ask ChatGPT what it remembers about you, and look at the Sources that can appear under a reply.
It will not make the model know things you did not tell it. It will not remove the need to edit and fact-check. It will not stop drift in a very long chat, so for long jobs restate the brief or start a fresh chat in the project.