Write a one-page business profile covering what you sell, who buys it, your price ranges, your tone and the things you never say, then save it where the AI reads it automatically: project instructions in ChatGPT or Claude, a Gemini Gem (becoming a skill from November 2026), or Copilot's custom instructions. Add two or three real examples of good work for each repeated task.
What you're building has three parts: the profile itself, a short instruction block for each task you repeat, and a handful of test prompts that show whether the context is working. One caveat before you start. Context only helps while it's current. A profile holding last year's prices produces confident, wrong quotes, so it needs an owner and a date at the top.
What the AI assumes when you tell it nothing
A general AI assistant has read a vast amount of text about businesses in general and nothing about yours. When your prompt leaves a gap, it fills the gap with the most typical answer. Ask it to reply to someone wanting three bedrooms repainted and you'll get a polite, plausible email that offers a free quote "within 24 hours", guesses at a price per room, and promises a start date. If you never quote without a site visit and you're booked up for three weeks, every one of those details is wrong.
Most chats also start blank. Some tools carry memory between conversations, but memory is personal, patchy and not something you can audit easily. The fix is to write your context down once, in a place the tool reads every time, so nobody has to retype "we're a small decorating firm, we don't do wallpaper stripping on ceilings" at the start of each chat.
Layer one: the business profile
This is the document every AI task in your business draws on. Keep it to one or two pages. Write it in plain statements rather than adjectives: "we reply within one working day" tells the AI something; "we're responsive" doesn't. Here's the structure, filled in for an illustrative painter and decorator:
BUSINESS PROFILE (owner: office coordinator last checked: 1 Sept 2026)
WHAT WE DO
Interior and exterior painting and decorating for homes and small
commercial premises. Wallpaper hanging. Minor plaster repairs before
painting. We do NOT do: structural plastering, spray-painting kitchens,
removing old lead-based paint (we refer these on).
WHO OUR CUSTOMERS ARE
Mostly homeowners (about 70%), plus letting agents and small shops.
Homeowners care about tidiness and dust sheets; agents care about
turnaround between tenants.
HOW WE PRICE
Always after a site visit or photos, never from room counts alone.
Typical ranges (update quarterly): single bedroom repaint $___-$___,
full house exterior $___-$___. Materials itemised separately.
Deposit: 25% to book.
HOW WE TALK
Plain, friendly, short. First names. No exclamation marks.
We say "decorator", not "painter technician". We don't say "no job
too small".
RULES
Never promise a start date; say "our current lead time is about
___ weeks" and let the office confirm.
Never give a price in writing without a visit or photos.
Always ask for photos and room sizes on first contact.
THINGS THAT CHANGE (check dates)
Current lead time: 3 weeks (as of 1 Sept)
Seasonal note: exterior work booked April to October only.Notice the "things that change" section at the bottom, each with a date. That's where stale context does the most damage, so it gets its own place and gets checked most often.
The headings stay the same for other businesses, but what goes under them changes a lot. A two-person wedding florist's profile would put most of its effort into rules and changing facts, because that's where its enquiries go wrong:
- How we price: "Minimum wedding spend $1,200. We price from a mood board and guest count, never from a photo of someone else's wedding. Deposit 30%, balance six weeks before the date."
- Rules: "Never confirm a date is free; the calendar is checked by a person. Never promise a specific flower, because availability depends on the season and the market that week. Say 'we'll suggest the closest match in season'."
- Things that change: "Dates fully booked: 12 Jun, 19 Jun, 4 Sep (as of 15 Aug). Peonies available roughly May to early July only."
The "never promise a specific flower" line matters more than anything about tone. Without it, an AI asked to reply to a bride who wants peonies in October will cheerfully say they'd be lovely.
Layer two: task instructions with real examples
The profile says who you are. Task instructions say what good looks like for one job you repeat: replying to enquiries, writing quote cover notes, summarising site visits, drafting replies to reviews. Each one needs four parts: the goal, the input you'll paste in, the format of the output, and two or three real examples of that output done well.
The examples do more work than any description. Giving the AI a few finished samples to imitate (sometimes called "few-shot" prompting) teaches tone, length and structure faster than a paragraph of rules. Pick examples your best person wrote, remove customer names, and keep them short.
TASK: Reply to a new enquiry
GOAL: Get photos and room sizes so we can price, and set honest
expectations on lead time. Don't quote a price.
INPUT: I'll paste the customer's message.
OUTPUT: An email under 120 words. Thank them, ask the two or three
questions we need, state our current lead time from the profile,
sign off as "The office team".
GOOD EXAMPLE 1:
"Hi [first name], thanks for getting in touch about the hallway and landing.
To give you an accurate price, could you send a few photos of each
area and a rough idea of ceiling height on the landing? Our current
lead time is about three weeks, and we'll confirm dates once you're
happy with the quote. The office team"
GOOD EXAMPLE 2:
[second real reply, names removed]
CHECK BEFORE SENDING: no prices, no fixed start dates, no promises
about paint brands we don't stock.Here's the task block at work. The customer writes: "Hi, we've just bought a two-bed flat and want the whole place painted before we move in on the 20th. Can you give me a price?" With the profile and task block loaded, the draft might come back like this (illustrative):
Hi [first name],
Congratulations on the new flat. To give you an accurate price, could
you send a few photos of each room and the rough room sizes? It also
helps to know whether you'd like ceilings and woodwork done as well as
walls. Our current lead time is about three weeks, and we'd be happy
to work around your move-in date.
The office teamMost of it is right: no price, photos requested, the real lead time, the right sign-off. The problem is the last clause of the paragraph. "Happy to work around your move-in date" reads to a customer as a promise that the job will be done by the 20th, which nobody has checked. The profile's rule bans promising start dates, but the AI didn't treat a finish date as the same thing. The fix goes in the profile, not the chat: "Never promise a start or finish date, or agree to work around a deadline; say the office will confirm what's possible." One added line, and every future draft follows it.
If your team already keeps prompts in one place, these task blocks belong there too; building a shared prompt library covers how to organise them so people actually find them.
Where to put the context in each tool
Every major assistant now has a place for standing instructions and reference files. Use the one in the tool your team already opens every day, rather than choosing a tool for this feature alone. In ChatGPT that means a project, not a custom GPT: OpenAI is retiring GPTs, no new ones can be created, and existing ones stop running on 11 December 2026. For a fuller comparison, see how ChatGPT and Claude projects compare with Gemini Gems.
| Tool | For short, always-true facts | For the profile, task blocks and files | Worth knowing |
|---|---|---|---|
| ChatGPT | Custom instructions in personalisation settings, which apply to all your chats | A project: its own instructions plus uploaded files, applying only to chats inside that project | Custom instructions have a length limit, so keep the long material in project files |
| Claude | Personal preferences in your profile settings, which apply to every chat | A project: project instructions plus project knowledge files | Free accounts can create up to five projects; sharing projects with colleagues needs a Team or Enterprise plan |
| Gemini | Instructions for Gemini in settings (labelled Saved info on some accounts) | A Gem: written instructions plus files added under Knowledge, uploaded or picked from Drive | Google Docs and Sheets added to a Gem are picked up when you update them, so one master Doc stays current |
| Microsoft 365 Copilot | Custom instructions in the personalisation section of Copilot's chat settings | A Copilot Notebook with your files added as references | Notebook references must be files in your organisation's OneDrive or SharePoint, or uploads; links to outside content can't be added |
Two official pages worth bookmarking if you're setting these up: Anthropic's explanation of what Claude projects are and Google's guide to using Gems in Gemini Apps. Setting labels move around between updates, so if a menu name above doesn't match what you see, search the vendor's help centre for "project instructions" or "custom instructions".
A practical pattern for a small team: one project, Gem or Notebook per recurring area of work (enquiries, quotes, marketing), each holding the same master profile plus the task blocks for that area. Personal custom instructions then carry only individual preferences, such as "I prefer bullet points".
The reason for splitting by area shows up when you don't. A garden landscaping firm kept every task block in one project. Its profile said "prices only after a design visit", and its social-media task block said "mention that patios start from $4,000". Within a fortnight, enquiry replies were telling customers who asked about a whole new garden that "our patios start from $4,000", which some read as the price of the job. The AI hadn't malfunctioned; it had two instructions in the same place and blended them. Moving the marketing block into its own project stopped it at once, and the profile gained a line saying price examples belong in marketing copy only.
Standing context versus the details of one job
People often try to cram everything into the standing context, then wonder why answers get muddled. Split it in two. Standing context is true for every job this month: who you are, how you price, how you talk, your rules. Job context is true for this one piece of work only: the customer's message, the site-visit notes, the photos, the colour the client picked.
Standing context lives in the project, Gem or Notebook. Job context goes into the chat itself, pasted or attached at the moment you need it, and stays there. That split has three benefits. The standing context stays short and consistent. Customer details don't linger in a place every future chat can read. And when an answer is wrong, you can tell quickly whether the fault was the standing profile (fix it once, for everyone) or the job details (fix the brief for this job).
Mixing the two causes a particular kind of error that's hard to trace. Say a decorator uploads one customer's site-visit notes to the project files to get a quote written, and forgets to remove them. A week later, a cover note for a different customer mentions "the damp patch in the back bedroom we discussed", because the AI now treats those notes as part of the business's standing knowledge. The customer has no back bedroom. If an answer ever contains a detail from another job, check the project's files before blaming the model; a stray upload is the usual cause.
A useful habit is a short job brief pasted at the top of the chat, in the same order every time:
JOB: Repaint, two bedrooms and landing
CUSTOMER TYPE: homeowner, first-time customer
WHAT THEY SENT: [paste message or visit notes]
WHAT I NEED FROM YOU: quote cover note, under 150 words
ANYTHING UNUSUAL: high landing ceiling, needs a towerFive lines take thirty seconds to fill in, and they stop the AI inventing the details it wasn't given.
Worked example: a painter and decorator's enquiry replies
Take an illustrative eight-person painting and decorating firm: the owner, an office coordinator and six decorators. The coordinator answers around 40 enquiries a week and started using an AI assistant with no context, pasting each message in and asking for a reply.
Before any context, about seven in ten drafts needed heavy rewriting: they quoted per-room prices, offered start dates and used a breezy tone the owner disliked. Each reply took around nine minutes including the rewrite, roughly six hours a week.
The coordinator spent an afternoon writing the profile above, then an hour on the enquiry task block with two of her own best past replies as examples, and loaded both into a project. After two weeks of use, heavy rewrites dropped to roughly one reply in seven, mostly unusual jobs such as listed-building work. Average time fell to about four minutes, saving over three hours a week. The one thing that went wrong was telling: in week three the lead time grew to five weeks, nobody updated the profile, and eleven replies promised three. That's why "things that change" now gets checked every Monday.
What to leave out of the context
- Customer personal data. Names, addresses and phone numbers don't belong in standing instructions. Add them to a single chat only when the task needs them, and only on a plan whose terms you've checked. Classifying your business data before using AI tools helps you draw the line.
- Anything you'd be embarrassed to see leaked. Margins, staff issues and supplier disputes rarely improve everyday drafts.
- Old documents "just in case". A 2023 price list sitting next to the current one gives the AI two answers to choose between.
- Adjectives without examples. "Professional but warm" means different things to every reader, human or machine. Show a reply instead.
- Contradictions. If the profile says "never quote without a visit" and a task block says "give a ballpark", the output will wobble between them.
Testing whether the context is working
Write five test prompts that cover your normal work and your trickiest cases, and run them every time you change the profile. Score each answer: correct facts, right tone, followed the rules. If you have a written house style, test against it; a house style guide your team and AI both follow makes that scoring far less subjective.
TEST PROMPTS (run after every profile change)
1. A routine enquiry: "Can you paint my living room next week?"
2. Price pressure: "Just give me a rough price, I don't need a visit."
3. Out of scope: "Can you spray my kitchen cabinets?"
4. Out of season: "We'd like the outside done in December."
5. Tone trap: an angry message about a late start.
PASS: no price given, correct lead time, refers out-of-scope work,
explains the exterior season, stays calm and short on number 5.A scored run is quick to record. After the week-three lead-time update in the worked example above, the coordinator's results sheet might read:
| Test | What came back | Result | Fix |
|---|---|---|---|
| 1. Routine enquiry | Asked for photos; said "about five weeks" | Pass | None |
| 2. Price pressure | Explained why a visit or photos are needed; no figure | Pass | None |
| 3. Out of scope | Offered to "look into" spraying the cabinets | Fail | Profile now says "we refer spraying to a specialist" and names what to say |
| 4. Out of season | Explained exterior work runs April to October | Pass | None |
| 5. Tone trap | Calm, but 180 words and apologised three times | Partial | Task block now says "one apology, under 100 words" |
Three passes, one fail, one partial, and ten minutes to fix both. Run the same five again after the fix; a test that passes once and fails the next time needs a firmer rule, not a lucky rerun.
If a test fails, fix the profile rather than patching the prompt in the chat. A patch helps once; a profile fix helps every chat after it. Where your business needs the AI to answer from a large body of documents rather than a single profile, a company knowledge base AI can answer from is the next step up, and writing prompts that sound like your business goes further on tone.
Keeping the profile current without it becoming a chore
Keep one master copy of the profile in your shared drive and treat everything loaded into AI tools as a copy of it. Give it one owner. Check the "things that change" section weekly, a two-minute job, and review the whole profile quarterly or whenever prices, services or staff change. In Gemini, attaching the master Google Doc to a Gem means updates flow through; in ChatGPT and Claude, re-upload the file after each change and delete the old one from the project so the two versions can't conflict.
Questions about feeding AI your business details
Can I just switch on memory instead of writing a profile?
Memory helps with personal preferences, but it's a poor home for business facts. You don't control exactly what it keeps, it belongs to one person rather than the team, and it can hang on to an old price or a one-off exception as if it were the rule. Put business facts in a document you control, loaded into a project, Gem or custom instructions, and let memory handle individual habits.
How long should the business profile be?
One to two pages is plenty for most small businesses: roughly 400 to 900 words. Longer profiles tend to contain contradictions and stale detail, and the AI gives every line similar weight. If you have more reference material, such as full price lists or product specifications, keep it in separate files attached to the project rather than in the profile itself.
Is it safe to upload our price list and templates?
Price lists and templates are usually internal rather than confidential, so they're reasonable to use on a business plan that doesn't train on your content by default. Check the plan's data terms first. Keep customer names, addresses and anything covered by a confidentiality agreement out of the standing context altogether; add those only to individual chats when a task genuinely needs them.
Further reads
- How to Set Up Claude Projects as a Shared Team Assistant — Set up one Claude project the whole team can share.
- How to Build Gemini Gems for Repeat Business Tasks — Build Gems for the tasks you repeat every week.
- How to Use ChatGPT Projects to Keep Client Work Separate — Use ChatGPT projects so each client's context stays in its own place.
- How to Build a Product Glossary AI Must Use in Every Draft — Stop AI renaming your products and services.
- Does ChatGPT Remember Conversations? Memory Settings for Business — What ChatGPT's memory keeps and how to control it.
- Why AI Gives Different Answers Each Time, and How to Fix It — Why the same prompt varies, and how context steadies it.
- How to Get Started With AI in Your Small Business: First 7 Steps — Seven steps that take about a month, from a five-day time log to a keep-or-drop decision, with a yoga studio's numbers at each stage.
- How to Check AI Is Doing Good Work, Not Just Fast Work — A rubric template, a 30-minute weekly sampling routine, a re-runnable test set and the warning signs, shown through a garden centre's plant-care replies.
- How to Train Staff to Use AI in a Small Business — A three-layer training plan with learning outcomes, a 75-minute session agenda, a copyable error-spotting exercise and a language-school example.
- How to Run a SWOT Analysis of Your Business With AI — Six stages and copyable prompts for a SWOT built on your own figures and reviews, with a podiatry clinic example and a TOWS step that turns it into actions.
- Reduce Owner Dependency: Use AI to Capture What Only You Know — Get what's in your head into a form staff can use: an interruption log, an AI interview prompt, decision rules with thresholds, and a shared assistant.
- The Limits of AI: What It Still Gets Wrong in a Small Business — Eight things AI still gets wrong in a small firm, where each one bites, the warning signs, and the specific check that catches it.
- Do You Need a Lot of Data to Use AI in Your Business? — How much data each common AI task needs, from three example emails to two years of sales history, and why scattered data is the bigger problem.
- Is AI Too Complicated for Non-Technical Business Owners? — Four levels of AI use, from typing into a chat box to custom builds, with honest learning times and five tests for when to bring in help.
- How to Learn AI as a Business Owner: A 30-Day Self-Study Plan — A day-by-day 30-day plan for owners to learn AI on real work, from first prompts to one working automation, with free courses and a day-30 self-test.
- How to Keep Up With AI in 30 Minutes a Week — A timed weekly routine for owners: scan your own tools' release notes, filter hard, test one change on real work, and log what you decide.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Anthropic's Claude help pages on projects and personal preferences; Gemini Apps Help on using Gems and adding Knowledge files; OpenAI help pages on ChatGPT projects and custom instructions; Microsoft Support pages on Copilot custom instructions and adding references to Copilot Notebooks.