Harness engineering
How to write instructions an AI will follow
You asked the AI to answer a customer enquiry and it came back with three cheerful paragraphs and a price you don't charge. Nine times out of ten, the problem is the brief it was given.
What makes AI instructions work?
Instructions work when they are written for someone who knows nothing about your business. The AI is quick and capable, but it has never met your customers, seen your price list or heard how you talk. Everything it needs has to be on the page.
Anthropic, the company behind Claude, gives the same advice in its prompting best practices: think of the AI as a brilliant but new employee who lacks context on your norms and workflows. You would never hand a new starter a sticky note saying "do the emails" and walk off.
One job per brief
A brief for answering enquiries should not also cover writing social posts and chasing invoices. Each job gets its own page. Short, separate briefs are easier for the AI to follow and much easier for you to fix when something goes wrong.
Say why, as well as what
"Keep replies under 120 words" works better as "Keep replies under 120 words, because most customers read them on a phone between jobs." Anthropic's guide says the same: explaining why a behaviour matters helps the AI aim at your real goal. A reason also lets it handle the cases your rule never mentioned.
Show it good work
Examples steer the AI more reliably than any description of tone. Anthropic calls examples one of the most reliable ways to steer format, tone and structure, and suggests three to five. Paste in real replies you were happy with, with names and details changed. The AI will copy their length, their warmth and their structure.
Tell it what to do instead
"Don't make up prices" leaves a gap. "If the customer asks for a price that isn't in the price list below, say we'll confirm it by the end of the day" closes it. Anthropic's guide puts it as a rule: tell the AI what to do, rather than only what not to do. Every never should come with an instead.
Not too vague, not too rigid
A brief can fail in two directions. "Be helpful and professional" tells the AI nothing. Forty if-this-then-that rules break the first time a customer writes something you didn't foresee. Anthropic's engineers call the middle the right altitude: clear enough to steer, loose enough to think. Aim for a page of principles, a few hard rules and good examples.
Weak lines and the stronger versions
| Weak line in a brief | Stronger line | Why it works better |
|---|---|---|
| Be friendly | Thank them by first name, then answer their question in the first sentence | Describes behaviour you can check, not a mood |
| Don't quote prices | Only quote prices from the list below; for anything else, say we'll confirm by 5pm | Gives the AI something to do instead |
| Keep it short | Under 120 words, because customers read on a phone | A number plus a reason handles cases the rule misses |
| Use our usual sign-off | Sign off: Thanks, Priya - Northside Dental reception | The AI can't know your usual anything |
| Handle complaints carefully | Never reply to a complaint; flag it for the practice manager | Takes a risky judgement away from the AI |
| Use your best judgement | If you are not sure, say so and flag the email for me | Turns a confident guess into a note you can check |
How to write a brief for one job
- 01
Name the job and who it is for
One line at the top. For a plumber: "You draft replies to new enquiries from homeowners in and around Leeds who need a plumber, usually for a leak or a boiler." The AI now knows who it is talking to. Not sure which job to brief first? See where to start with AI in a small business.
- 02
Describe a good result
What does a reply you'd be proud of contain? For example: thanks them by name, answers the question asked, gives the next free slot, ends with one clear next step. Write it as a short checklist. That checklist doubles as the test you mark the AI's work against later.
- 03
Write the rules it must never break
Keep this list short and specific to the job: never quote a price that isn't in the price list, never promise a same-day visit, never discuss another customer. Five firm rules beat twenty vague ones. Rules that matter a lot also need a check behind them, which we cover below.
- 04
Give it the facts it needs
Paste in the price list, the service area, the opening hours and the call-out policy, or point it at the document where they live. An AI without the facts will fill the gaps with guesses that sound right, which is most of why AI makes things up. Keep one current copy of each fact, so the AI never reads last year's prices. In a bigger business, Living Documentation adds checks that fail when a document falls behind.
- 05
Add two or three real examples
Pick replies you were proud of, covering different kinds of enquiry: a simple booking, a price question, an awkward one. Change the names and remove anything personal first. Label them clearly as examples, so the AI copies their shape and not their details.
- 06
Say what to do when it isn't sure
This is the line most briefs leave out. "If you are not sure, say so, and flag the enquiry for me." It turns a confident wrong answer into a short note for you to check.
- 07
Save it where the AI reads it every time
Most AI tools let you save standing instructions, in a project, a custom assistant or a settings page. Use that, so you never paste the brief by hand and every reply starts from the same page.
An example brief, start to finish
Here is a complete brief for one job at a small dental practice. The names are made up. It runs to under 250 words, which is plenty.
The job: you draft replies to emails sent to the reception inbox of Northside Dental, a two-dentist NHS and private practice. Most emails ask about appointments, prices or what to do about pain. Replies go to patients who are often anxious, so be calm and clear.
A good reply: thanks them by first name, answers their question in the first sentence, gives one clear next step, and stays under 120 words because most people read it on a phone.
The rules: only quote prices from the price list below; for anything else, say reception will confirm by 5pm. Never give clinical advice. If someone describes swelling, bleeding that won't stop or difficulty breathing, do not draft a reply: flag it as urgent for the practice manager. Never mention another patient. Never book or move an appointment; suggest times from the diary and let reception confirm.
When you are not sure: say so in a note at the top of the draft, and leave the reply for a person.
The facts: opening hours, the price list, the cancellation policy and the out-of-hours number, pasted below. Then three example replies the practice was happy with.
Why the clinical line is so strict
A reception AI that gives health advice is a risk no practice should take. The brief doesn't try to teach the AI where the line is. It moves the whole category to a person. That is a pattern worth copying: when a judgement is risky, don't describe it, route it.
Why it never books
The brief says the AI suggests times and reception confirms. That limit is only a request in the instructions. To make it stick, the AI should not have permission to change the diary at all, which is the subject of give AI only the tools it needs.
Where to save your instructions
Save the brief in the tool's own place for standing instructions, so the AI reads the same page before every job. A brief pasted into one chat is gone when the chat ends, and a new chat starts from nothing.
In Claude
A Claude project holds its own instructions and a knowledge base of documents, such as your price list, which Claude uses in every chat inside that project. Anthropic explains them in what are projects.
In Gemini
Google calls a saved set of instructions a Gem: you create and save one to shape Gemini's answers for a particular job. Google's help page on using Gems in Gemini Apps shows how.
In ChatGPT
ChatGPT has settings-level custom instructions and projects that keep files and instructions together. Use a project per job rather than one set of instructions for everything, for the same reason you keep one brief per job.
Inside your own software
If a developer is building the AI into your systems, the brief goes in as the system or developer message. OpenAI's prompt engineering guide describes developer messages as the system's rules, given priority over what a user types, and recommends a few input and output examples.
Where AI instructions go wrong
Most bad briefs fail in the same few ways. Each one is easy to spot once you know to look.
Writing for yourself
"Use the usual sign-off" means something to you and nothing to the AI. Anthropic suggests a simple test: show the brief to a colleague with no context and ask them to follow it. If they would be confused, the AI will be too.
Rules that fight each other
"Always be brief" in one place and "always explain our full guarantee" in another leaves the AI to pick. Read the rules together once a month and remove the clashes.
A brief that grows until nobody reads it
Every fix adds a line, and a year later the brief is six pages. Long instructions get followed less well. Anthropic's own guidance for its coding tool suggests keeping each instructions file under 200 lines, because longer files reduce how closely they are followed, in its page on how Claude remembers your project. Merge, trim and move rarely needed detail into a separate document.
Real customer details in the examples
It is tempting to paste last week's real emails in as examples. Change the names, addresses and anything identifying first. UK data protection law expects you to use no more personal data than the job needs, as the ICO explains under data minimisation.
Fixing answers instead of the brief
When a reply comes out wrong, it is tempting to correct that one reply and move on. The same mistake will come back next week. Change the brief, and that mistake has a rule against it from then on. The full habit is in every AI mistake should become a rule.
Expecting the brief to do everything
Instructions shape what the AI tries to do. They don't stop it doing something it has the access to do, and they don't check its work. Anthropic's own documentation says written instructions are treated as context, not enforced settings. Those are separate parts of the set-up: limits on its tools, and a check before anything reaches a customer.
How do you test a brief before trusting it?
Run it on real past work before it touches new customers. Take twenty emails from last month, let the AI draft replies, and mark each one against your checklist of a good result. Fix the brief, run them again, and only go live once most drafts need little or no editing.
Use last month's real emails
Real work catches what invented examples miss: the customer who writes in capitals, the one who asks three things at once. Twenty is enough to see a pattern. Remove personal details if you are pasting them into a tool that isn't already part of your business set-up.
Try the awkward ones on purpose
Add a few you know are hard: a complaint, a request for a discount, a question you can't answer. Anthropic's guide to building evaluations makes the same point for developers: test against your real mix of work, and don't forget the edge cases.
Set a pass mark before you look
Decide what counts as good enough first, for example 18 of 20 drafts usable with a light edit, and none that would embarrass you. Deciding afterwards lets you talk yourself into a weak result. How to trial AI on one job sets out a fuller version.
What it costs in time
Example figures for one job: an hour or two to write the first brief from real examples, an hour to test it on twenty emails, and ten minutes a week for the first month adding rules as mistakes appear. Most owners spend longer than that each week fixing replies from an AI that was never properly briefed.
Questions
08Asked and answered
Usually one page for one job: long enough to hold the rules, the facts and a few examples, short enough that you can read it in two minutes. Longer briefs tend to be followed less closely, so merge rules that overlap and move rarely needed detail into a separate document. The idea behind it is explained in what harness engineering is.
Not quite: a prompt is a one-off request, while standing instructions are the brief the AI reads before every request for that job. You write them once, save them in the tool and improve them each time something goes wrong. Anthropic's prompting best practices cover the techniques in more depth.
Yes, two or three real examples of good work steer the AI more reliably than any description of the tone you want. Pick different kinds of job, such as a simple booking, a price question and an awkward one, and label them clearly as examples. Remove customer names and personal details first, as our safety rules explain.
Check the brief first, because the rule may be vague, buried, or contradicted by another rule. Rewrite it with a reason and an instead, then test it on the job that went wrong. If it keeps happening, the fix is a check rather than a louder rule, the habit described in every AI mistake should become a rule.
Only partly, because instructions guide the AI but don't remove its access. An AI told never to send email can still send one if it has permission to. To be sure, limit what it can touch in the first place, such as drafts only, as covered in give AI only the tools it needs.
Mostly yes, because a clear brief with real examples works on any capable model. Each tool saves it in a different place: projects in Claude and ChatGPT, Gems in Gemini. Test on a handful of real jobs after any switch. Choosing between them is covered in which AI model a small business should use.
Point the brief at one current document for each fact, such as the price list, instead of pasting copies that go stale. When a price changes, you change one file and every job reads the new one. Our Knowledge Index keeps a map of every document in the business and how fresh it is.
Yes, every product we set up comes with written instructions and rules for its job, at a fixed price. We write them from your real examples, test them on past work and hand them over in your own accounts, so you can read and change them. See how it works for the three steps.
This is general guidance, not legal or financial advice. For your own situation, speak to a qualified adviser.