AI for Everyday Work: Writing, Research, Meetings & DataYour AI workspace: projects, voice, files and agents · Lesson 4 of 17
Projects, memory and custom instructions
Video lecture
Projects, memory and custom instructions
The narrated lecture is in production
Every chapter is scripted and ready. Browse the chapters and read the full transcript now — the video will appear here when it’s published.
Chapters
Transcript of the narration, chapter by chapter.
0:00 Projects, memory and instructions
How many times this month have you typed who you are into an AI chat? I'm a marketing coordinator. Use American spelling. Keep it short. If the answer is more than a few, you're doing work your assistant can do for you. In this lecture you'll learn the three layers that make an assistant feel like it knows you: custom instructions, projects and memory. You'll see a simple example and a realistic freelancer setup, you'll watch me build a project from scratch, and you'll leave knowing exactly what goes in each layer, and what should never go in any of them.
0:44 Why set this up
Why bother? Three reasons. First, better first drafts. When the assistant already knows your role, your audience and your rules, the first answer is closer to usable. Second, less typing. Your everyday prompts shrink to the task itself. Third, and this one matters for anyone with clients, separation. When each client's files and rules live in their own workspace, you never get client A's tone, or worse, client A's prices, in client B's email. Set up well, these features turn a clever stranger into something much closer to a trained assistant.
1:24 The filing cabinet analogy
Here's an analogy. Custom instructions are your ID badge. Everyone who works with you sees it, every time. So it should say who you are and how you like to work, nothing secret. Projects are drawers in a filing cabinet, one per job. Inside each drawer: the rules for that job, the reference files, and the conversations. Open the drawer, and everything relevant is right there. Memory is a board of sticky notes the assistant writes as it learns about you. Handy, but sticky notes pile up and go out of date. So every so often, you clear the board. In ChatGPT these are custom instructions, Projects and memory. Claude has preferences and styles, Projects and memory. Gemini has saved info, Gems and Gemini Notebook. Copilot has Notebooks.
2:20 What goes where
So what goes where? On the badge, your custom instructions: your role, the kind of work you do, spelling, preferred format, and a rule like never invent numbers, write check instead. In the drawer, the project: instructions that apply only to that job, reference files that rarely change, like a style guide, a glossary of product names, and one excellent past example as the quality bar. And for memory, let the assistant learn, but review it once a month. Delete anything outdated, anything sensitive, anything wrong. And here's the key rule. Client material never goes on your badge or your sticky notes. It goes in that client's drawer, in a work approved account.
3:09 Example 1: a weekly report project
Let's start simple. Say you write a weekly operations report. You create a project called weekly ops report. The instructions say who reads it, the structure you want, and never invent numbers. You upload last month's best report as the example. Now each week you open the project, paste the new figures and say draft this week's report. It follows the structure, uses the right terms, and wherever you didn't supply a number, it writes check instead of guessing. The first week, you'll tweak the instructions once or twice. By week three, it's a five minute job.
3:51 Example 2: the freelancer's three clients
Now a realistic business setup, with illustrative details. Daniela is a freelance virtual assistant in Austin with three small business clients: a yoga studio, a bookkeeping firm and a catering company. She creates one project per client. Each has a one page brief, services, tone, key contacts by role, and what she may and may not promise, plus the client's email signature and a few approved sample messages. Her global instructions only describe her. I'm a virtual assistant, American spelling, flag anything that commits the client to money or dates. Now when a catering inquiry arrives, she opens the catering project and says draft a reply. The draft sounds like that client, uses their menu names, and flags the date for her to confirm. Three clients, zero mix ups.
4:47 Watch me do it
Let me build one with you. I click new project and name it social posts for the bakery. In the instructions, I paste the template from the lesson. Purpose: help me write three social posts a week. Audience: local families and office workers. Always follow the attached brand notes and use the example as the quality bar. Never invent prices or opening hours. Format: caption, three hashtag options, and a suggested image. Next, I upload three files: brand notes, a glossary with our product names, and our best performing post. Now I test it. Write Monday's post about the new cardamom buns. And look, it uses our real product name, our tone, and it adds check price, because I didn't give it one. In a plain chat, the same request invented a price.
5:45 Memory hygiene in practice
Let's talk about memory hygiene, because it's the layer people forget. Once a month, ask your assistant a simple question: what do you remember about me? Or open the memory page in settings. Then do three things. Delete what's outdated, like a job you've left or a project that's finished. Delete what's sensitive, like anything about health, money, or other people's personal details. And move anything client specific out of memory and into that client's project, where it belongs. If you ever want a completely clean answer, with no memory and no history, use a temporary or incognito chat. It won't use or update memory. Five minutes a month keeps the assistant accurate about you, and keeps your private details where you want them.
6:39 Common mistakes
Here are the mistakes to avoid. One giant project for everything. The context gets muddled and the assistant blends rules from different jobs. Keep one project per stream of work. Stale files. If last year's price list is in the project, the assistant will happily quote it. Replace files when they change. Memory that's never reviewed. After a few months, it's full of old jobs and half true assumptions. Put a monthly reminder in your calendar. And confidential client files in a personal account's project. Projects are convenient, but they don't change which terms apply. Client work goes in a work approved account.
7:24 Recap and try this now
Let's recap. Custom instructions are your badge: who you are and how you like answers, applied everywhere. Projects are drawers: one per recurring job, with rules, reference files and an example of great output. Memory is sticky notes: useful, but audit it monthly. Here's your try this now. Pick the one task you do every week. Build a project for it using the template in the lesson text, upload two or three reference files, and run your next real task inside it. Then count how many follow ups it took to get a usable draft, compared with a plain chat. If it's fewer, you've just bought back time every single week.
Stop re-explaining yourself
If you find yourself typing "I'm a marketing coordinator at a dental clinic group, use American spelling, keep it short" into every chat, your assistant has features built for exactly that. There are three layers, and each has a different job.
| Layer | What it holds | Scope | Where to find it |
|---|---|---|---|
| Custom instructions / preferences | Who you are and how you like answers | Every chat | ChatGPT: Settings > Personalization; Claude: profile preferences (plus Styles); Gemini: saved info and Gems; Copilot: personalization settings where your admin allows |
| Projects (workspaces) | Files, instructions and chats for one client, product or goal | Only chats inside that project | ChatGPT Projects, Claude Projects, Gemini Gems with files or a Gemini Notebook, Copilot Notebooks, a Notion page with Notion AI |
| Memory | Facts the assistant learns over time | Across chats (and, in some tools, separately per project) | ChatGPT and Claude memory settings; Gemini's personal context; review and delete entries |
Names and plan availability change often; check your tool's help center if a menu has moved.
Custom instructions: short, stable, not secret
Good custom instructions describe you and your preferences, not a specific task:
About me: I coordinate marketing for a group of three dental clinics in Texas.
I write patient emails, social posts, staff updates and monthly reports.
How to respond:
- American spelling. Plain English at an 8th-grade reading level for patient content.
- Lead with the answer, then detail. Bullets over long paragraphs.
- Never invent prices, statistics or medical claims; write [CHECK] instead.
- When I ask for copy, give two options with different angles.Avoid client names, passwords, or anything you wouldn't want in every conversation.
Projects: one workspace per recurring job
A project bundles instructions, reference files and related chats. Create one for each recurring stream of work: a client, a product launch, your weekly reporting, a course you're studying.
A strong project has:
- A clear name ("Client: Al Noor Hotels", "Weekly ops report").
- Project instructions with the rules that apply only here (tone, banned claims, approval steps, disclosure needs).
- Reference files that rarely change: brand guide, product sheet, glossary, last quarter's report, a "good example" of the output.
- A naming habit for chats inside it ("2026-09 newsletter draft").
In ChatGPT, projects can have project-only memory so what the assistant learns there stays there. In Claude, project knowledge files are available to every chat in the project, and memory can be kept separate per project. In Gemini, a Gem is a saved custom assistant with instructions and optional files; a Gemini Notebook (the product formerly called NotebookLM) answers only from the sources you add.
Memory: useful, but audit it
Memory lets the assistant carry facts across chats ("prefers tables", "works in hospitality"). You can usually ask "What do you remember about me?" and view, edit or delete entries in settings. Review it monthly:
- Delete outdated facts (a former employer, a finished project).
- Delete anything sensitive (health, finances, other people's details).
- Correct wrong assumptions.
- Keep client-specific context in a project, not global memory.
When you want a clean slate, use a temporary or incognito chat, which does not use or update memory.
Worked example: a freelancer's three projects
Daniela, a freelance virtual assistant in Austin, supports three small business clients. She creates one project per client with a one-page brief (services, tone, key contacts by role, what she may and may not promise), the client's email signature and a sample of approved messages. Her global custom instructions only describe her ("I'm a virtual assistant; American spelling; flag anything that commits the client to money or dates"). Now "draft a reply to this booking inquiry" inside the right project produces an on-brand draft on the first try, and client details never mix.
Hands-on: build your first project in 15 minutes
- Pick one recurring task you do at least weekly.
- Create a project (or Gem, or Notebook) named after it.
- Paste these project instructions and adapt them:
Purpose: help me produce [the weekly client update / social posts for X / monthly report].
Audience: [who reads it and what they care about].
Always: follow the attached style guide; use the attached example as the quality bar;
list any facts I should verify at the end.
Never: invent numbers, quotes or dates; promise discounts, deadlines or refunds.
Format: [the exact structure you want].- Upload two or three reference files: a style guide or brand notes, one excellent past example, and a glossary of product names.
- Test it with a real task. Compare the result with what you got in a plain chat.
Before (plain chat): "Write this week's client update" produces a generic update with made-up metrics.
After (inside the project): the same request produces an update in the client's format, using the terms from the glossary, with "[CHECK: sessions number]" wherever you didn't supply data, plus a short "verify before sending" list.
Pitfalls
- One giant project for everything; context gets muddled. Keep one per stream.
- Uploading confidential files into a personal account's project.
- Stale reference files: replace last year's price list, or the assistant will quote it.
- Letting memory accumulate for months without review.
How to measure success
For your project's task, count follow-up rounds before a usable draft. A good setup gets you there in one or two rounds instead of four or five.
Key takeaways
- Custom instructions describe you and your preferences; keep them short, stable and non-confidential.
- Projects (and Gems, Notebooks) bundle instructions, reference files and chats for one recurring job.
- Memory carries facts across chats; review and prune it monthly, and keep client context inside projects.
- A good project includes a quality-bar example and a glossary, and tells the AI to flag facts to verify.
Check your understanding
Quick questions to lock in the lesson. They don’t count towards your certificate.
Put it into practice
Build one project for a weekly task with instructions and two or three reference files, then compare follow-up rounds with a plain chat.
Enrol for free to save your progress
Reading is always free. Enrol to keep your place, take the final assessment and earn a verifiable certificate.