--- title: "Set context once: custom instructions, projects and memory" description: "The problem: repeating yourself If you use an assistant every day, you probably type the same background again and again: who you are, what your business…" url: https://optimizeall.com/learn/prompt-engineering-foundations/context-once-projects-and-memory updated: 2026-10-05 --- Prompt Engineering Foundations · Modern assistant features: modes, memory, research and media · lesson 8 of 16 · 12 min # Set context once: custom instructions, projects and memory ## The problem: repeating yourself If you use an assistant every day, you probably type the same background again and again: who you are, what your business does, your preferred tone. Modern assistants let you store that context so every conversation starts with it. This is the everyday version of what professionals call **context engineering**: deciding what information the AI sees before it answers. ## Three places to store context **1. Custom instructions (or personalisation).** A short standing note that applies to all your chats. Typically two parts: what the assistant should know about you, and how you want it to respond. **2. Projects (and similar workspaces).** Many assistants let you create a workspace with its own instructions and uploaded files. ChatGPT and Claude call these Projects; Gemini offers Gems; Microsoft Copilot has notebooks and agents in some plans. Names and features vary by plan and change often, so check your own assistant. Everything in the workspace (your brand guide, price list, FAQs) is available to every chat inside it. **3. Memory.** Some assistants remember useful facts from past chats. You can usually see, edit and delete memories, or switch the feature off. ## Writing good custom instructions Keep them short, factual and stable. Things that change weekly belong in the prompt, not here. ```text About me: I run a 4-person digital marketing agency in Manchester. Clients are local restaurants, gyms and salons. I write in British English. How to respond: - Be concise: lead with the answer, then detail if needed. - Use prices in GBP and UK examples unless I say otherwise. - When you are unsure of a fact, say so rather than guessing. - For anything I will publish, flag claims I should verify. ``` Avoid vague lines like "be amazing" or long lists of rules. Each line should change the answer in a way you want. ## Setting up a project that works A good project has three parts: 1. **A clear purpose:** one client, one product line, one recurring job (for example "Weekly social posts for Bloom Café"). 2. **Standing instructions:** the brief that applies to every chat in the project. 3. **Reference files:** the few documents the AI really needs. Example project instructions: ```text Purpose: draft social posts and replies for Bloom Café, a family café in Abu Dhabi. Always: - Use the brand voice guide (file: bloom-voice.pdf): warm, playful, never sarcastic. - Use prices from the current menu (file: menu-sept-2026.pdf) only. - Write in English; add an Arabic version when I ask. - Keep captions under 80 words with at most 2 emojis. Never: - Invent offers, prices or opening hours. If something is not in the files, ask me. - Mention competitors. ``` ## Choosing the right files More files are not always better. The assistant has to decide what is relevant, and outdated files cause confident mistakes. Keep files: - **Current:** delete last season's menu when the new one arrives. - **Focused:** the brand guide and the menu, not every document you own. - **Clearly named:** "menu-sept-2026.pdf" is better than "final_v3.pdf". ## Memory: helpful and risky Memory saves time, but it can also surprise you. An assistant might remember that you once asked for a formal tone and apply it everywhere. Review memories regularly, delete anything wrong or outdated, and switch memory off for sensitive topics. Never rely on memory for anything critical; put important facts in the prompt or project files. ## Worked example: an agency with many clients Omar runs a small agency in Riyadh with six clients. He used to paste each client's details into every chat and still got mixed-up tones. Now he has one project per client, each with the brand guide, product list and standing instructions. His custom instructions hold only his own preferences. Results are more consistent, and when a new team member joins, the projects double as onboarding documentation. ## Privacy and workplace rules Anything you put in custom instructions, projects or memory is stored by the provider. Before uploading client documents: - Check your organisation's AI policy and which plan it approves (business and enterprise plans typically offer stronger data controls). - Remove personal data you do not need, such as customer phone numbers and addresses. - Check whether your plan uses your content for model training and how to opt out if appropriate. ## Hands-on: set up your first project in 15 minutes 1. Choose one recurring task (weekly newsletter, client replies, lesson plans). 2. Create a project or workspace in your assistant (or, if yours has none, a saved text you paste at the start of each chat). 3. Write standing instructions using the Purpose, Always, Never structure above. 4. Upload one or two current reference files. 5. Test with three real tasks. Note anything the assistant got wrong and add a line to the instructions to fix it. ## How to know it is working Your prompts get shorter while the results get more consistent. If you still have to repeat a detail in most chats, add it to the instructions. If answers include outdated facts, clean up your files or memory. ## Video lecture: Set context once: custom instructions, projects and memory Video: [Set context once: custom instructions, projects and memory](https://www.youtube-nocookie.com/embed/pABxXy0kgOM) — The problem: repeating yourself If you use an assistant every day, you probably type the same background again and again: who you are, what your business… 12 chapters · about 8 minutes · captions and full transcript below. 1. Set context once 2. Why store context? 3. Three places for context 4. Writing custom instructions 5. A project that works 6. Choosing files 7. Memory: helpful but review it 8. Privacy first 9. No projects feature? Use a saved note 10. Example 1: weekly meal plans 11. Example 2: an agency with 5 clients (illustrative) 12. Recap ## Lecture transcript ### Set context once If you use an AI assistant every day, you have probably typed the same background dozens of times. I run a small agency. I write in British English. Keep it short. In this lecture you will learn how to store that context once, using custom instructions, projects and memory, how to write instructions that actually change the output, which files to upload, and how to stay on the right side of privacy rules. By the end, your prompts will get shorter while your results get more consistent. ### Why store context? Why does this matter? Because repetition is where quality quietly slips. On a busy day you forget to mention the budget, or the brand's tone, or the rule about never promising delivery dates, and the answer drifts. Stored context works like a well-briefed assistant who already knows the house rules. You do not re-explain the rules every morning; you just give today's task. The result is shorter prompts, fewer mistakes, and answers that stay consistent across days and across team members. ### Three places for context There are three places to store context. Custom instructions are a short standing note that applies to all your chats, usually in two parts: what the assistant should know about you, and how you want it to respond. Projects, or similar workspaces, are for one recurring job, like one client or one product line. They hold their own instructions and reference files. ChatGPT and Claude call these projects, Gemini offers Gems, and other assistants have their own versions; names and features vary by plan. And memory, where the assistant remembers useful facts from past chats, which you can usually view, edit or switch off. ### Writing custom instructions Good custom instructions are short, factual and stable. For example: I run a four-person digital marketing agency in Manchester, clients are local restaurants, gyms and salons, I write in British English. Then how to respond: lead with the answer, use prices in pounds and UK examples, say when you are unsure of a fact, and flag claims I should verify before publishing. Each line changes the answer in a way you want. Avoid vague lines like be amazing, and keep things that change weekly out; those belong in the prompt. ### A project that works A good project has three parts. A clear purpose, such as weekly social posts for a family café in Abu Dhabi. Standing instructions in a simple structure: purpose, always, and never. Always use the brand voice guide and the current menu prices; keep captions under eighty words. Never invent offers or opening hours; if it is not in the files, ask me. And a few reference files, only the ones the AI truly needs. ### Choosing files Files deserve special care, because more is not better. The assistant has to decide what is relevant, and outdated files cause confident mistakes. Keep files current, and delete last season's menu when the new one arrives. Keep them focused; the brand guide and the menu, not every document you own. And name them clearly; menu September twenty twenty-six beats final version three. Omar, who runs a small agency in Riyadh, set up one project per client with its own brand guide and products. The tone mix-ups stopped, and new team members now use the projects as onboarding material. ### Memory: helpful but review it Memory is helpful and a little risky. An assistant might remember that you once asked for a formal tone and quietly apply it everywhere. So review your memories every few months, delete anything wrong or outdated, and switch memory off for sensitive topics. Never rely on memory for critical facts. Put those in the prompt or project files, where you can see them. ### Privacy first Everything you store is held by the provider, so think before uploading. Check your organisation's AI policy and which plan it approves; business and enterprise plans typically offer stronger data controls. Remove personal data you do not need, like customer phone numbers and addresses. And check whether your plan uses your content to improve models, and how to opt out if that is appropriate. These habits keep you compliant with data protection laws in the UK, the Gulf and beyond. ### No projects feature? Use a saved note A common question is, what if my assistant does not have projects? You can get most of the benefit with a saved context note. Keep a short document with your standing instructions, purpose, always and never, and paste it at the start of each new chat for that job, followed by the day's task. It is a little more manual, but it gives the same consistency. And if you work in a team, share the note, so everyone briefs the assistant the same way and gets results in the same voice. ### Example 1: weekly meal plans A simple example first. Every week, you ask your assistant to help plan family meals. Every week, you repeat the same facts: four people, one child who will not eat spicy food, a halal kitchen, and a budget. Put those facts into custom instructions, or a small meal-planning project, once. Now your weekly prompt can be just: plan this week, we have chicken, lentils and spinach. The answers stay consistent, and you stop forgetting to mention the spicy food rule. Stored context is simply the facts you would otherwise repeat forever. ### Example 2: an agency with 5 clients (illustrative) Now a business scenario, with illustrative numbers. Kamran runs a four-person agency in Lahore with five retainer clients, each producing about twenty social posts a month. Before projects, writers pasted each brand's details into every chat, and mixed up tones about once a week. He creates one project per client, each with purpose, always and never instructions, the brand voice guide and the current offers sheet. Writers now start every task inside the right project. Over a month, that is around one hundred posts where nobody retypes the brief, and the tone mix-ups largely stop. Every quarter, Kamran deletes outdated offer files and reviews memory. These figures are illustrative, but the structure scales to any small team. ### Recap To recap. Custom instructions hold stable facts about you and how to respond. Projects hold instructions and a few current files for one recurring job. Memory saves time but needs regular review. And privacy rules still apply to everything you store. Try this now: pick one recurring task, create a project or saved context with purpose, always and never instructions, upload one current file, and test it on three real tasks. Add a line for anything it gets wrong. Next lesson: prompting with files, photos and your voice. ## Key takeaways - Custom instructions store who you are and how you like answers across all chats. - Projects (and similar workspaces) hold standing instructions and reference files for one recurring job. - Keep stored context short, current and focused; outdated files cause confident mistakes. - Review memory regularly and follow workplace rules before uploading client or personal data. ## Try it Create one project or saved context for a recurring task: write Purpose, Always and Never instructions, upload one current file, and test it on three real tasks. - [Previous: Quick answers, thinking modes and deep research](https://optimizeall.com/learn/prompt-engineering-foundations/quick-thinking-and-research-modes) - [Next: Prompting with files, photos, screenshots and voice](https://optimizeall.com/learn/prompt-engineering-foundations/prompting-with-files-photos-and-voice) - [All lessons of Prompt Engineering Foundations](https://optimizeall.com/learn/prompt-engineering-foundations)