Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Glossary term

Standing Instructions

The business's rules placed before every question an AI assistant answers: scope, sources, refusals, when to hand over and what never to promise.
StrategyconceptUpdated

Definition

Standing instructions are the business's own rules for an AI assistant, written once and placed in front of every question it answers, alongside the passages it retrieves from the business's documents. The tutorial on writing the standing instructions for an AI assistant builds them in sections: the assistant's job and the contact reasons it handles, the sources it answers from and what to do when they do not cover a question, the topics it hands to a person, the exact disclosure and handover lines, the tone, and a short list of things it never promises, asks for or does. Each rule gets a test conversation, and the whole set gets a version number, a date and a named owner.

The word "standing" is the point. These rules are not typed fresh for each task; they persist across every conversation until someone changes them on purpose, so earlier versions are kept and the tests are run again whenever the instructions or the documents are edited or the model behind them is replaced.

How they differ from a prompt in a library

A prompt kept in a team's library is used by a colleague, who reads what comes back and decides whether to use it. Standing instructions act without that reader: the customer sees the reply first. So they must carry the decisions the colleague would have made, such as whether a question is in scope, when to stop and fetch a person, and what is never offered. A library prompt can be little more than a task and a format; a set of standing instructions works more like a remit, a policy and a style sheet combined.

They also differ from the documents the assistant answers from. Prices, opening hours and policy details belong in owned, current documents, which the tutorial on preparing documents for an AI assistant gets ready; facts typed into the instructions age unseen, outside any document owner's view. When a trace shows a wrong answer came from an outdated passage, the fix is the document, not another instruction.

In practice: customer assistants, analysis queries and tone

Customer assistants. This is the main use. The rules are tested against conversations built from real past questions, one where the assistant should keep each rule and one designed to lure it into breaking it. When it breaks one, the instructions or documents change and the test stays. No model can be relied on to keep every rule every time, so reading a sample of live conversations continues after launch.

Analysis queries. The same idea appears when a team uses AI to query its data. The guide to AI for analysis and reporting turns the log of mismatches found in review into a list of standing instructions given to every query: which definitions, which governed tables, which time zone, which touch model.

Tone. Adjectives such as "friendly and professional" leave the model to guess. The instructions describe the tone in the terms the tutorial on writing customer service tone guidelines already uses, with approved and rejected replies taken from those guidelines. The AI applications hub lists the other task guides an assistant's owner is likely to need beside these.

Illustrative example — not a real result

A lettings agency's assistant stops promising late checkouts

Vaeqain Cottages, an invented holiday lettings agency, manages 120 cottages for their owners and runs an assistant on its booking pages. In its first 3 weeks the assistant tells 11 guests that a late checkout is fine, reasoning from a help article that says late checkouts are "often possible". No owner has agreed to any of them, and 4 clash with same-day arrivals. The operations lead traces the replies to the instructions, which said nothing about availability. Version 2 adds a rule that any question about dates, availability, early arrival or late checkout goes to a person, with a set handover line, and 6 new test conversations, 3 of them phrased as casual asides. All pass before release, and the next month's sample of 200 conversations finds no availability promises.

A Prompt is the general case, an instruction and material given to a model; standing instructions are the part that stays fixed in front of every conversation. A Language Model writes from whatever it is given, which is why the rules have to be explicit. Grounding is what the sources rule asks for, and it narrows invented answers without removing them. Human Review is the check a person makes before something is sent; an assistant answering customers directly has none, so its instructions take on that role. Brand Voice is the source the tone section draws on, so the assistant sounds like the rest of the business.

Common confusions

  • Standing instructions are just a long prompt. The format may look similar, but they run unattended and repeatedly, so they need scope, refusals, handover rules, tests, versions and an owner that a one-off prompt does not.
  • Facts belong in the instructions. Facts belong in the documents, where their owners keep them current. The instructions say where to look and what to do when nothing is found.
  • Good instructions guarantee good answers. Tests show how the assistant behaves with the current instructions on the day they run; reviewing real conversations shows whether that holds.
  • A failing test means the test is wrong. The tutorial's rule is the reverse: fix the instructions or the documents and leave the test as it is.
  • The tool maker's instructions are enough. The maker's part is general. This business's scope, documents, handover topics and promises come only from its own rules.

More terms to look up

Where this term is used

Related