ChatGPT Custom Instructions, Memory and Projects for Marketers

#01What goes where?

OpenAI's memory article says memory can draw on past chats, saved memories, custom instructions, files in Library or Space and connected apps such as Gmail, depending on plan and region. It also says memory does not keep every detail. That is why exact facts belong in project files.

#02What should your custom instructions say?

Keep them about you, not about a brand. Here's a template that fits the 1,500-character limit.

I'm a [role] at a [B2B/B2C] company selling [category] to
[audience]. When I ask for copy: be direct, use British/US
English, short sentences, no hype words (revolutionary,
seamless, unlock). When you state a fact that isn't from my
files, say where it comes from or mark it [CHECK]. When a
task is ambiguous, ask one question before answering. Default
to tables for comparisons.

#03How should you set up a brand project?

Create one project per brand or client. In Project settings, add instructions and set the memory option. Then upload a short set of files.

PROJECT INSTRUCTIONS: [Brand]
Audience: [who, their job, what they're measured on].
Voice: [three rules plus one example sentence].
Product facts: use only the attached fact sheet.
Proof points we may cite: [list]. Nothing else.
Never claim: [list, e.g. "guaranteed savings", named customers
without approval].
Competitors: refer to them as Competitor A-D in drafts.
Output defaults: headings in sentence case, UK spelling.

Good files to attach: a one-page fact sheet, a voice guide with three published examples, two or three personas built from real interviews, and an approved claims list. Keep each short. Long brand books dilute what matters.

#04When should you use project-only memory?

Use it for client work, regulated products and anything you'd not want leaking into other chats. OpenAI's projects article lists what changes: saved memories are not used in the project, chats can reference only chats in the same project, and outside chats can't reference it. Shared projects are always project-only.

One trade-off: ChatGPT Work isn't available in a project-only memory project. If a project needs Work (lesson 11), use default memory and keep sensitive material out.

#05When does memory help, and when does it hurt?

Memory helps with small preferences: your time zone, how you like tables, the brands you work on. It hurts when it carries an old fact into a new context, such as last year's pricing. When an answer seems to know something it shouldn't, look for the Sources line under the response; OpenAI says it can show the past chat, saved memory or file that shaped the answer. For one-off sensitive work, a temporary chat stays out of history and isn't used for training while it remains temporary, though it can still use existing memories and instructions unless you switch personalisation off.

Our view: most marketers put too much in memory and too little in project files. If a fact has to be exact, it belongs in a file you can update.

#06Worked example: Ledgerline's projects

Ledgerline is a fictional accounts-payable tool. Details are illustrative.

Ledgerline set up three projects: "Brand content" (default memory, used with Work for decks and docs), "Customer research" (project-only, holding interview notes) and "Launch Q1" (shared with the product marketing manager, so project-only by default). Each had the same fact sheet. Custom instructions held only personal style rules.

When the pricing page changed, they updated one fact sheet and re-uploaded it to all three. Within a week a draft in "Brand content" still quoted the old tier name; the Sources line showed it came from a saved memory. They deleted that memory.

#07What should you check before the next lesson?

• Write custom instructions under 1,500 characters that say nothing about a specific brand.
• Create one brand project with instructions, a fact sheet and a voice guide.
• Decide which projects need project-only memory and which need ChatGPT Work.
• Claude users can compare with Claude projects and brand voice.

Next: Custom GPTs, plugins and skills for your team. Or return to the ChatGPT for Marketers track.

#08Common questions

What is the difference between custom instructions and memory in ChatGPT? Custom instructions are rules you write that apply to every chat. Memory is context ChatGPT picks up from past chats, saved memories and other sources. Use instructions for stable preferences and check memory regularly.

Do project instructions override custom instructions? Yes. OpenAI says project instructions apply only inside that project and override your global custom instructions.

Should a client project use project-only memory? Usually yes. Project-only memory stops chats in that project from drawing on chats elsewhere, and vice versa. Shared projects use it automatically. Note that ChatGPT Work is not available in a project-only memory project.