Make AI Work Your Way

The problem: repeating yourself

By now you've had enough AI conversations to notice a pattern. Every time you start a new chat, you re-introduce yourself. "I run a small clothing shop in Chattogram." "Please use simple language." "Don't give me ten options, give me one recommendation." The AI does exactly what you ask — and then a new chat opens with a completely blank memory, and you type the same preamble all over again.

That's not you doing it wrong. As we saw in Context & Memory, each chat is its own sealed room. But the tool makers know this is annoying, so every major chatbot now offers ways to make your preferences stick. There are three of them, from lightest to heaviest: custom instructions, memory, and projects (also sold as GPTs or Gems). Master these three and the AI stops feeling like a stranger you brief every morning and starts feeling like a colleague who already knows you.

Custom instructions: your standing note

Custom instructions (some apps call it "personalization" or "profile") are a short note the AI silently reads before every single chat you start. You write it once in settings, and it shapes every answer afterwards. The note usually has two halves:

• Who you are. "I run a small clothing shop in Chattogram with two employees." "I'm a nurse who works night shifts." "I'm job hunting in accounting." The AI uses this to pitch answers at the right level and pick relevant examples.
• How you like answers. "Always use simple language." "Answer in Bangla unless I write in English." "Be direct — one recommendation, not a menu of options." "Keep answers short unless I ask for detail."

Where to find it: open your AI app's Settings, then look for something named Personalization, Custom instructions, or Profile. Every major tool — ChatGPT, Claude, Gemini, Copilot — has this drawer, though each hides it in a slightly different corner. It's typically one or two text boxes; type your note, save, done. From the next chat onward it applies automatically.

The effect is bigger than it sounds. Tell it once that you prefer plain language and Bangla, and you never again get a wall of jargon in English. Tell it once about your shop, and "write me a Facebook post for Eid" comes back already sounding like a clothing shop in Chattogram rather than a generic corporation.

Memory: facts that stick on their own

Custom instructions are things you write down. Memory is the tool's own notebook: many chatbots now quietly remember facts that come up in conversation — "user's daughter is in class 8", "user is vegetarian", "user's shop sells women's clothing" — and reuse them in later chats without being asked.

This is convenient and slightly unsettling in equal measure, so know your controls. In settings (usually under "Memory" or "Personalization") you can typically:

• View the list of everything it has remembered about you — worth doing once; it's an odd little mirror.
• Edit or delete individual memories ("forget that I was planning a Cox's Bazar trip").
• Turn memory off entirely, or start a temporary chat that saves nothing.

A sensible habit: let memory keep harmless useful facts (your profession, your language preference), and delete anything sensitive it picks up. Health matters, money matters, family conflicts — if you discussed them once, that doesn't mean you want them echoed back in an unrelated chat next month. The full picture of what to share at all is in Privacy Basics.

Projects and custom bots: a workspace with its own brain

Custom instructions apply to everything, which is exactly why you should keep them general. But some of your work is repetitive in a more specific way. Suppose you answer customer messages for your shop every day. Each time, the AI would ideally know your product list, your prices, your delivery policy, and your friendly-but-brief tone. That's too much baggage for a global note — and far too much to paste into every chat.

This is what projects are for (ChatGPT calls its shareable version GPTs, Gemini has Gems, Claude has Projects — same idea, different labels). A project is a workspace with two superpowers: its own standing instructions, and attached files that every chat inside it can read. Chats inside the project start pre-loaded; chats outside it are unaffected.

Creating one takes about five minutes, and the steps are nearly identical everywhere:

1. Create and name it. Look for "New project" or "Create a GPT/Gem". Name it for the job: "Shop Assistant", "CV & Job Applications", "Thesis Helper".
2. Write its instructions. This is the same skill as custom instructions, but specific: "You help me reply to customer messages for my clothing shop. Use the attached product list for prices. Tone: warm, brief, always end by offering delivery info. Reply in the language the customer used."
3. Attach files. Upload the product list, the price sheet, the return policy — whatever this job always needs. PDF, Word, spreadsheet all work in most tools.
4. Use it. Open the project, paste a customer message, and the reply comes back knowing your stock and your tone — no preamble at all.

Other project ideas from ordinary life: a "Family Health Notes" project holding your parents' prescription lists (for organising questions before a doctor visit — not for diagnosis); a "Job Hunt" project with your CV attached so every cover letter starts from the real you; a "Landlord" project with your tenancy agreement so rent questions get answered against the actual contract.

The three layers, in one picture

Here's how the pieces stack. Every message you send sits inside up to three layers of standing context:

Which tool for which job

Writing standing instructions that actually work

A standing instruction is just a prompt that runs forever, so everything from Prompting 101 applies — with three extra rules born from the "forever" part:

• Keep it short. Five to ten crisp lines beat two rambling paragraphs. Every line you add is a line the AI must obey in every chat, including chats where it doesn't fit. "Always answer in a table" sounds clever until you ask for a poem.
• Keep it concrete. "Be helpful" does nothing — it's already trying. "Give one clear recommendation before listing alternatives" changes real behaviour. If you can't imagine an answer visibly changing because of a line, cut the line.
• Treat it as a living note. Whenever you notice yourself correcting the AI the same way again, open settings and add one line. Whenever a line starts causing weird answers, delete it. Ten minutes of gardening a month keeps it sharp.

Here's a fill-in template to start from — edit the bracketed parts and paste it into your settings:

About me: I'm a [your work / role] in [city]. I mostly use AI for [your top 2–3 uses, e.g. customer messages, emails, learning].

How I like answers:
- Simple, everyday language — no jargon without a one-line explanation.
- Answer in [Bangla / English / the language I write in].
- Be direct: give your best single recommendation first, alternatives only if I ask.
- Keep answers short by default; I'll ask when I want detail.
- If my request is unclear, ask me one clarifying question instead of guessing.

And here's a starter template for your first project or custom bot — this one is written for a shop assistant, but swap the bracketed parts for any recurring job:

You are my assistant for [the job, e.g. replying to customer messages for my clothing shop].

Your materials: use the attached files ([product list / price sheet / policy doc]) as the source of truth. If something isn't in them, say so — don't invent prices or stock.

Your style: [warm and brief / formal and precise]. Reply in the language the customer used. Always end with [e.g. delivery and payment info].

When I paste a [customer message / document / question], respond with [a ready-to-send reply / a draft for me to review].

Key takeaways

• Custom instructions are a note the AI reads before every chat: who you are, how you like answers. Write it once in Settings → Personalization.
• Memory is the tool's own notebook of facts about you — view it, prune it, and delete anything sensitive.
• Projects (GPTs / Gems) bundle standing instructions plus attached files for one recurring job — a "Shop Assistant" that actually knows your stock.
• Rule of thumb: one-off task → normal chat; recurring preference → custom instructions; recurring task with materials → project.
• Keep standing instructions short, concrete, and updated whenever you catch yourself repeating a correction.
• Personalization changes tone and context, not accuracy — keep checking important output.

Next up: teaching the AI to go out and gather information for you — Research & Deep Dives.