Set up a reusable programme brief in ChatGPT or Claude

Keep a programme's reusable writing rules in project instructions and its changing facts in dated files. Start the next chat inside that project, check the memory settings, and compare the answer with the sources before it goes into a steering pack.

Use this guide to set up a standing brief in ChatGPT or Claude. It includes a fictional programme example and a short note on Copilot Notebooks.

The instructions follow the vendors’ help pages, rechecked on 21 September 2026. Availability depends on your plan and workspace settings; interface labels may differ.

Who this is for

Use this when the same programme keeps coming up in your chats and you are tired of explaining it again. Keep its risks, assumptions, issues and dependencies (the RAID log), planned switch-over date and reporting rules together. If you are starting with a single task, begin with the guide to a one-shot prompt for real work.

If you need to keep two connected Google accounts apart, see Connect a second Google account to ChatGPT without mixing work and home. That guide covers account selection and memory settings.

Give instructions, files and memory different jobs

Project instructions hold behaviour: who you are writing as, who reads it, the output shape, and what the model must not do. OpenAI says project instructions apply inside that project and override your global custom instructions. Anthropic says project instructions apply to all chats in that project.

Keep changing facts in dated files: the RAID log, last week’s note, current constraints and the planned switch-over date. In Claude, add these to project knowledge so they are available across the project’s chats. Do not rely on a detail buried in an earlier conversation being recalled accurately.

Check memory separately from the files. ChatGPT offers default and project-only memory; shared projects use project-only memory and cannot switch back. Anthropic describes separate project memory for Team and Enterprise accounts where an owner has enabled it. Keep the source documents current whichever setting you use.

Where do Custom Instructions fit?

Keep Custom Instructions for tone and format only. Put programme facts in the Project that holds the files. OpenAI's Custom Instructions Help says those preferences apply across chats once Enable customization is on, and updates land immediately on existing chats. That is useful for British English, short answers and a house format. It is a poor place for a live RAID list, a cutover date or a named customer, because every chat can inherit it.

OpenAI’s Projects in ChatGPT guidance says project instructions override global Custom Instructions inside that project. Keep stable writing preferences in Custom Instructions, project-specific rules in project instructions, and changing programme facts in files. Check the first answer against those files.

Set the brief in ChatGPT

OpenAI’s Projects in ChatGPT help page describes projects for signed-in users, subject to plan availability and workspace settings. The steps below follow that guidance.

If you already have a useful chat, OpenAI documents drag-to-project or Move to project. Chats created with a GPT cannot be moved. Temporary chats cannot be added. After a move, the chat inherits the project's instructions and file context.

• Click New project in the sidebar. Name it after the live programme, not Work.
• Open the project’s three-dot menu, choose Project settings and add the writing instructions. These override your global custom instructions inside this project.
• Upload the fact files. OpenAI's current file caps on that page: Free 5 files per project; Go and Plus 25; Edu, Pro, Business and Enterprise 40. You can upload 10 files at once. Users can create an unlimited number of projects.
• Under Memory, choose Default memory or Project-only memory, then Save. Project-only memory limits chat context to that project, and OpenAI says ChatGPT Work is unavailable inside it. Choose the setting that fits your task. Shared projects use project-only memory and cannot switch back to default.
• OpenAI says changes to the memory setting can take a few hours to take effect. Do not treat Save as instant.
• Start a new chat in the project and ask for this week’s output. If the RAID log is out of date, replace it with the current version before asking.

Set the brief in Claude

Anthropic’s How can I create and manage projects? page lists projects for all users, including free accounts. Follow these steps to add the brief and its source files.

• Hover the left side and click Projects, or go to claude.ai/projects.
• Click + New Project. Give it a name and description. Anthropic notes that Claude will not have access to those details, so do not hide the brief there.
• Click Set project instructions, add the writing rules, then choose Save instructions. They apply to all chats in the project.
• Use + on the right to add documents to project knowledge. Put the current RAID log and other source files here so they are available across the project’s chats.
• Free accounts can create a maximum of five projects. This help page does not give a matching cap for paid plans. Team and Enterprise plans can set visibility and share with people in the organisation, where admins allow it.
• Claude memory is on by default for Free, Pro and Max on web, desktop and mobile. Team and Enterprise accounts need an owner to enable it. Anthropic’s guidance for those team plans describes separate project memory. Removing a chat from a project moves it into non-project memory; it does not erase it.

Copilot Notebooks

Copilot Notebooks also support writing instructions. Microsoft’s Provide custom instructions for your Microsoft Copilot Notebook page gives two routes: More options (...) > Instructions in the Copilot app, or the arrow beside the notebook name > Copilot instructions in OneNote. Save the rules there. Licence and rollout requirements still apply. This is another documented setup route, not a comparison of the tools.

A constructed standing brief

The following brief is fictional. It is not client material or output from a session I ran.

Constructed programme: Northbridge Payments Migration. Fourteen weeks to cut an in-house payments file over to a hosted processor. Friday steering pack for the sponsor, programme board and workstream leads.

Put the writing rules below in project instructions. Add the factual material as files in ChatGPT or project knowledge in Claude. Then open a new chat in the project and ask for the week’s steering note using your current notes.

What to put in files, not in the instructions box (all constructed):

• RAID log dated 25 August 2026.
• Last Friday's steering note.
• Cutover window: weekend of 17–18 October 2026, freeze from 10 October.
• Named constraint: card scheme certification is a dependency this programme does not own.

Check the first answer against the files

Use the checks in How to check AI output before using it at work. The sample below is fictional; it shows the problems to look for in your own result.

Constructed first output: "Northbridge remains on track for the 17 October cutover. The supplier delay is a minor issue and should clear next week. Sarah Chen owns the scheme certification action, due 3 October. Risk score is now amber. Next week we will confirm the freeze with finance."

Before sharing, check names and numbers, look for risks that have been softened, and read the whole note yourself. Use the full eight checks for material that needs formal review or sign-off. Saving the brief reduces repeated typing; you still need to check each answer.

• Task fit. You asked for a steering note with movement, open reds and decisions needed. This is a cheerful status line. It skipped the rehearsal decision the board asked for.
• Invented facts. Sarah Chen, 3 October, an amber score, and "on track" do not appear in the constructed files or notes.
• Missing context. The freeze from 10 October and the rule that scheme certification is a dependency, not a task this programme owns, are absent.
• Softened or inflated risk. A supplier block on test data became a minor issue that should clear. That is the dangerous one for a steering pack.
• Wrong owners and commitments. It handed scheme certification to a named person and a date nobody set.
• Tone and audience. “Remains on track” sounds reassuring, but the supplied notes do not support it. A steering pack needs a status the reader can trace to the evidence.
• Verification route. You cannot trace the owner, the date or the amber score to a file. They do not go out.
• Sign-off. You own the final version. Read it once as yourself before your name goes on the pack.

Keep the brief current

Replace outdated files when the facts change, and check that similarly named versions are not giving the model conflicting instructions.

Your plan and workspace settings can limit file uploads, sharing and memory. Claude Free allows five projects. Check those limits before moving a programme’s working material into a project.

After changing ChatGPT’s memory setting, allow for the few hours OpenAI says it may take to apply. Check the setting before relying on project-only memory.

Sources

Projects in ChatGPT, OpenAI Help Center. Instructions, files, memory and plan limits. Checked 21 September 2026.

How can I create and manage projects?, Anthropic Help Center. Project knowledge, instructions, free-plan limits and memory guidance. Checked 21 September 2026.

Provide custom instructions for your Microsoft Copilot Notebook, Microsoft Support. Setup routes and availability. Checked 21 September 2026.

ChatGPT Custom Instructions, OpenAI Help Center. Global writing preferences and when changes apply. Checked 21 September 2026.

The eight checks are from How to check AI output before using it at work. The one-shot prompt method sits on How to write AI prompts for real work.