What standing instructions are and why a prompt library entry is not enough
A Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. writes from whatever is in front of it, and the guide to how language models generate text notes that tools wrap the model in an interface and add instructions the user never sees. In an assistant built by a business, part of that hidden text is the business's own: the standing instructions. They sit in front of the customer's question every time, alongside the passages the assistant retrieves from the business's documents.
A PromptThe instruction and material given to an AI language model, which works like a brief: the model knows language, not the audience or product, unless told. in a team's prompt library is different in one way that matters. It is run by a colleague who reads the output and decides what to do with it. Standing instructions are followed with no one reading the reply before the customer does. So they have to carry the decisions a colleague would otherwise make: what is in scope, what to say when the sources are silent, when to stop and fetch a person, what never to promise. A library entry can be a task and a format; an instruction set is closer to a job description, a policy and a style guide in one.
Instructions are one part of a working assistant, not the whole. The documents it answers from are prepared in the tutorial on preparing documents for an AI assistant; scope, handover and disclosure are decided in the customer service guide; and the method for scoring answers belongs to the guide to testing prompts and models. Where the instructions are entered, and any limit on their length, is in the tool's own documentation.
| Section of the instructions | What it says | Where the decision is made |
|---|---|---|
| Job and contact reasons | Who the assistant serves and which questions it handles | Customer service guide, contact reason list |
| Sources and the silence rule | What it answers from, how it cites, what it says when the sources do not cover the question | Customer service guide, grounding page |
| Hand-over topics | The subjects and signals that always go to a person | Customer service guide, vulnerable circumstances guide |
| Disclosure and handover lines | The words it uses to say what it is and to pass a conversation on | Disclosure guide, support channels guide |
| Tone | How it sounds, with approved and rejected examples | Support tone guidelines, brand voice guide |
| Never | Promises, data and actions it must not make, ask for or take | Customer service guide, claims rules |
| Tests, version and owner | A test conversation per rule, a version, a named owner | Testing guide, NIST framework |
Before you start
- The scoping decisions from the AI in customer service guide: which contact reasons are automated, assisted or kept human.
- The team's Contact ReasonThe main reason a customer contacted support, tagged by the agent from a fixed, defined list so volume and other service metrics can be read by cause. list, built as the guide to tagging support contacts by reason describes, with a few real past questions for each reason, stripped of personal details.
- The document library the assistant will answer from, prepared and owned.
- The support tone guidelines, or time to write them with the tutorial on writing customer service tone guidelines.
- The disclosure rule from the disclosure and data privacy guide.
- A test environment where the assistant can be run with draft instructions before customers see them.
- An owner: one person accountable for the instructions, with authority to change them.
The steps
Write the assistant's job and the contact reasons it handles
Open the instructions with a plain statement of who the assistant serves and what it is for: customers booking on the website, staff looking up policy, visitors with questions before they buy. Then list the contact reasons it handles, in the words of the contact reason list, and say that anything else is out of scope. Keep the list to the reasons the scoping exercise marked for automation. A short, explicit scope gives every later rule something to refer to, and it gives testers a clear line between a question the assistant should answer and one it should pass on.
Name the sources it answers from and what to do when they are silent
State that the assistant answers only from the supplied documents, names the article or page it used, and says plainly when the documents do not cover a question, then offers a person. This is the core rule the customer service guide gives, and the grounding and retrieval page explains why it matters: GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. narrows invented answers without removing them. Name the document sets by purpose so a reviewer can trace an answer, and say what to do when two sources disagree: stop and hand over rather than choose. Keep the documents themselves out of the instructions; facts that live in the instructions go stale where no document owner will see them.
List the topics it must hand to a person
Write the hand-over list as topics and signals, not as a vague request to use judgement. The customer service guide keeps complaints, refunds, credits and exceptions, legal, safety and regulatory matters, and customers in vulnerable situations with people. The guide to supporting customers in vulnerable circumstances describes the signs a support team watches for; turn them into plain descriptions the assistant can recognise, and accept that a cautious rule will hand over some conversations that did not need it. Add the triggers the customer service guide names: the customer asks for a person, the conversation goes in circles, or the customer is plainly upset.
Write the disclosure and handover lines
Write the exact words the assistant uses to say what it is, at the start of every conversation and whenever a customer asks. The disclosure guide gives the rule, and the European Commission's overview of its regulatory approach includes transparency for systems that interact with people; the FTC's business guidance holds that deception standards apply whatever the technology. None of this is legal advice. Then write the handover lines: what the assistant says when it passes the conversation on, and what it says when no person is available, offering a realistic alternative. The support channels guide asks that the person who takes over receives the full conversation, so the line can promise the customer will not have to repeat themselves only if the handover really does that.
Set the tone from the support tone guide with approved and rejected examples
Describe the tone in the terms the support tone guidelines already use, not new adjectives. Instructions that say "friendly and professional" leave the model to fill in what that means. Include a few short approved replies, including one or two from the team's Saved ReplyA pre-written answer a support agent inserts and then edits for the customer in front of them, keeping facts and tone consistent across the team. library, and a few rejected ones with a note on what is wrong with each: too formal, too chatty, an apology that blames the customer. The brand voice guide for AI builds a voice pack for drafting with the same approach; reuse its examples where they fit, so the assistant and the team's drafting tools sound like the same brandThe consistent personality a brand expresses in its words across channels, with tone adjusting to the situation and the audience..
State what it must never promise or ask for
Write a short list of things the assistant never does, each in one plain sentence. It never promises a refund, a credit, an exception or a delivery date beyond what a cited document states. It never asks for passwords, full card numbers or other details the business does not need in chat. It never gives legal, medical or safety advice. It never presents itself as a person. It never makes a claim about the business's products that the documents do not support, and the FTC's business guidance expects claims to be substantiated whoever, or whatever, makes them. Where the assistant can see account information, the customer service guide's rule is read-only access through defined integrations; say so in the instructions as well as enforcing it in the tool's settings.
Write a test conversation for every rule
For each rule in the instructions, write at least one short conversation that should make the assistant keep it and one that tempts it to break it. A question the documents answer, and one they nearly answer. A customer who asks whether they are talking to a person. A refund request phrased as a simple question. A sign of distress in the middle of an ordinary booking query. Build them from the real past questions gathered before you started, with personal details removed, and write the expected behaviour beside each. The guide to testing prompts and models sets out how to build a test set and a scoring rubric; this step makes sure every rule is covered by it.
Run the tests and fix the instructions, not the tests
Run every test conversation against the draft instructions in the test environment and score each against its expected behaviour. When the assistant breaks a rule, change the instructions or the documents, never the test: make the rule clearer, move it nearer the top, add an example, or fix the document that misled it. Re-run the whole set after each change, because a fix for one rule can loosen another. The site makes no promise that any model keeps every rule every time; the tests show how this assistant behaves with these instructions, and the sampling reviewA person reading a random sample of AI answers or generated copy on a fixed rhythm, plus every case that ended badly, and routing findings to the owners. the customer service guide describes keeps showing it after launch.
Version the instructions and name an owner
Give the instructions a version number, a date and an owner, and keep every earlier version. Record what changed and why, linked to the test result that prompted it. Re-run the test set whenever the instructions, the documents or the underlying model change. When an assistant that once passed its tests has been drifting for weeks, the page on AI output quality getting worse separates patched instructions from a model change, stale documents and thinner inputs. The NIST AI Risk Management Framework frames the cycle: govern with a named owner, map where the assistant acts and what it sees, measure with tests and sampling, and manage by changing the instructions when the evidence says so. Before launch, work through the AI customer service launch checklist, which ticks off the decisions these instructions now carry.
The applications hub lists the other AI task guides, and the document preparation tutorial is the companion job for the content the assistant reads. The academy checkpoint on grounding, document libraries and standing instructions tests what standing instructions carry that a prompt library entry does not, how a hand-over list is written and whether to fix the instructions or the documents when a test conversation fails.
A ferry operator writes its booking assistant's instructions (illustrative)
Tolvanji Ferries writes and tests its booking assistant's instructions
Tolvanji Ferries, an invented operator running 3 island routes, adds an assistant to its booking pages. The customer experience lead owns the instructions, version 1.0.
The scope covers 4 contact reasons: timetables, luggage and vehicles, accessibility on board, and how to change a booking online. The assistant answers from 3 named pages, the timetable, the luggage policy and the accessibility guide, and cites the page it used. It hands over refunds, disruption compensation claims, medical needs and any complaint. It opens every chat with a line saying it is an automated assistant, and never quotes a fare.
The team writes 26 test conversations, at least 1 per rule. The first run fails 3:
| Test | What happened | Fix |
|---|---|---|
| "What time is the last sailing on Saturday?" asked in July | Quoted the winter timetable | Old timetable retired from the library; instructions now say to state the dates the timetable covers |
| "Can I get my money back if it's cancelled?" | Summarised the refund policy instead of handing over | Refunds rule reworded to cover any question about money back |
| "Are you a real person?" | Answered the booking question and ignored the disclosure question | Disclosure line moved to the top and repeated as a rule |
After the fixes, all 26 tests pass. Version 1.1 records the 3 changes and the test results, and the launch checklist is completed before the assistant goes live.
Common mistakes
- Copying a prompt library entry and calling it instructions. A task and a format do not cover scope, silence, handover and disclosure.
- Putting facts in the instructions. Prices, times and policies belong in owned documents the assistant cites.
- Asking for "good judgement" instead of a hand-over list. Name the topics and signals that go to a person.
- Writing the disclosure line and never testing it. Ask the assistant whether it is a person and read the answer.
- Describing tone with adjectives only. Approved and rejected examples do more than "warm and professional".
- Changing the test to match the answer. A failing test means the instructions or the documents need fixing.
- No owner and no version. Nobody knows which instructions produced last week's conversations.
- Testing once at launch. Changes to the instructions, the documents or the model each need a fresh run.
