Build a Reusable AI Assistant for the Task You Do 20 Times a Week
If you use AI for the same kind of task over and over (triaging support emails, rewriting job posts, turning notes into tickets) you’ve probably noticed you keep typing the same setup instructions at the start of every chat. “You’re a support rep for a software company, our tone is friendly but concise, never promise a refund without approval…”
There’s a better way. Almost every major AI tool now lets you save that setup once and reuse it. Let’s build one.
What we’re building
A small, single-purpose assistant that already knows its job before you say a word. You give it the standing rules once; from then on you just paste in the new email, note, or draft and it does the task your way.
Under the hood this is just a The standing instructions a model reads before it ever sees your message: its “job description.” You normally never see it, but the custom-assistant features let you write your own. It’s the difference between briefing a temp every single morning and hiring someone who already knows the role. that you control, saved and attached to every conversation automatically.
Step 1: Pick one narrow job
The mistake is building a “general assistant.” Those are useless because they’re vague. Build something narrow enough to describe precisely:
- Good: “Rewrites our raw meeting notes into Jira tickets in our format.”
- Too broad: “Helps with product management.”
Narrow tasks have clear rules. Clear rules make a good assistant.
Step 2: Find the feature in your tool
Every major tool has this. They all call it something different, which is the main reason people assume they don’t have it.
Paths and plan requirements below were accurate in July 2026. These menus move, so treat them as “look for something like this” rather than gospel.
ChatGPT: “GPTs”. Sidebar → Explore GPTs → + Create in the top right. You’ll get a Configure tab with Name, Description, Instructions, Knowledge (file uploads) and Capabilities (web search, image generation and so on). Needs a paid plan. ChatGPT also has Projects, which bundle instructions and files with a thread of conversations: use a Project if the work is ongoing, a GPT if it’s a repeatable job.
Gemini: “Gems”. Profile menu → Gem manager → New Gem at the top right. Instructions, Knowledge files, and a default tool, then Save. Gems are available on every Gemini plan, free included, which makes this the cheapest place to try the idea if you’re not sure yet.
Claude: “Projects”. Left sidebar → Projects → New Project, then fill in the project’s instructions. If you’d rather set something across every conversation instead of one job, that’s your initials at the lower left → Settings → the “Instructions for Claude” box.
Microsoft 365 Copilot: “Agent Builder”. Creating one needs a Microsoft 365 Copilot licence, or a tenant with pay-as-you-go enabled. Copilot’s licensing for sharing is genuinely tangled: the rules changed in 2026 and the official guidance still reads inconsistently on whether colleagues need their own licence to use what you built. If you’re on Copilot, ask IT before you promise a teammate access.
Step 3: Write the standing instructions
Write them as if onboarding a new hire for exactly this one task. A reliable structure:
- Role: “You turn my rough meeting notes into well-formed tickets.”
- Rules: the non-negotiables. “Every ticket has a title, an acceptance-criteria list, and a priority. Never invent details I didn’t give you.”
- Format: show the exact output shape you want, with a filled-in example.
- What to do when unsure: “If a note is ambiguous, list your questions at the bottom instead of guessing.”
That last rule matters more than it looks. It’s the main defense against a When a model states something with total confidence that isn’t actually true, because it’s built to produce a plausible-sounding answer, not to look something up. Telling it to ask instead of guess when information is missing is one of the simplest ways to cut this down. : you’re explicitly giving it permission to say “I don’t know” instead of filling the gap with something invented.
Step 4: Attach reference material it should always have
If the task always needs the same background (your style guide, a glossary, a list of your team members, last quarter’s goals) attach those files or paste them into the instructions. Now they’re baked in. You’ll never re-paste your style guide again.
Some tools call this “knowledge” or let you upload files to the project. Behind the scenes it’s a lightweight version of When the tool automatically pulls relevant chunks from documents you’ve given it and slips them into the conversation as needed, so it can reference a 40-page handbook without you pasting the whole thing each time. , and it means the assistant can reference your documents without you copying them in over and over.
Step 5: Test it with three real cases, then fix the instructions
Run three real inputs through it. When it gets something wrong, don’t correct it in the chat: go back and fix the standing instructions, so the fix sticks for every future run. This is the key mindset shift:
You’re not correcting an answer. You’re editing the job description.
A few loops of this and the assistant stops making the same mistakes entirely.
Step 6: Name it, use it, and share it carefully
Give it an obvious name (“Notes → Tickets”) so future-you finds it in two seconds: you’ll pick it from a list every time you use it, so the name is the interface.
Using it afterwards is the part nobody explains: you don’t get a new app. You start a conversation inside the thing you made: pick it from your GPTs, Gems or Projects list, and everything you set up is already loaded. If you start a normal chat instead, none of it applies, which is the most common reason people think their assistant “stopped working.”
Before you share it, reread what’s in it. Sharing hands over the instructions and the attached files. A style guide is fine; the customer list you uploaded to make the examples better is not. Check what you attached in Step 4 before you send the link, and if your workplace hasn’t written down what may go into these tools, that’s the thing to sort first.
Once it’s clean, hand it to a colleague. Now the whole team does the task the same way, and you’ve quietly standardized a process without writing a policy doc.
The takeaway
Any task you re-explain more than a few times a week is a candidate for this. The upfront cost is fifteen minutes. The payoff is never typing that setup again, and getting more consistent results, because the rules stop depending on whether you remembered to include them today.
Sources: menu paths and plan requirements verified July 2026 against OpenAI’s help centre (opens in a new tab), Google’s Gemini Apps help (opens in a new tab), Claude’s help centre (opens in a new tab), and Microsoft Learn on sharing agents (opens in a new tab). These interfaces change often, if a path doesn’t match what you see, trust your screen. The Copilot licensing caveat reflects genuinely inconsistent official guidance as of writing, not a gap in the reading.
Related: How to Write a Recurring Report With AI in 10 Minutes: the same “save the recipe” idea, applied to reporting. How to Make an AI Agent Without Writing Any Code is the step up from this: when the job spans several steps and several apps and you’d rather not be the one carrying things between them.