Safe Prompt Patterns for Common Work

This is Unit Four of Generative AI at Work, on safe prompt patterns for common work. Unit Three built prompts from their elements. This unit adds the lines that make them safe, drawn from the M-CaReM-AI method's prompt patterns and the DAIR.AI guide. Carry one question through this unit: what turns a fast prompt into a safe one? The answer arrives with an unsafe prompt rewritten, one problem at a time. By the end you will be able to use the method's grounding and uncertainty lines. To decide what may go into a prompt. To apply safe patterns for summaries, extraction, rewriting and messages. To use the method's work patterns. And to answer from documents and chain prompts safely.

Five entries on the lines the method builds into prompts to stop them inventing.

The first grounding rule.

GAI 076 Only What Is Given

The M-CaReM-AI method builds a grounding rule into its prompts: use only what the pasted material provides or plainly implies, and if something seems plausible but is not given, leave it out. A plausible addition is the hardest error to catch, because it fits.

The rule against invention.

GAI 077 Do Not Invent

The method's prompts also say outright: do not invent new fields, states or entities beyond what the material supports. Asked to assemble a record definition, a model may add a plausible risk-score field nobody asked for. Models fill gaps by default, so the rule has to be stated in every prompt.

What to do with gaps.

GAI 078 Unclear Means Ask, Not Guess

Where the material is unclear or incomplete, the method sends the gap to an open-questions list instead of letting the model guess, so a summary can end with two questions for the policy owner rather than two invented answers. An open question costs a short conversation; a guessed answer can cost a wrong decision.

Marking uncertainty.

GAI 079 Mark It Uncertain

The method also asks the model to label any relationship or conclusion it is unsure of as uncertain, so the reviewer knows where to look first. A label of certainty from a model is not evidence, though, so the unmarked items still need checking.

How to treat the output.

GAI 080 A Starting Point, Not a Verdict

The method treats every AI output, even from a well-built prompt, as a starting point a person reviews and corrects by hand, like an impact list the analyst trims, extends and signs off on rather than forwarding as it came. A good prompt reduces the correcting; it never removes the review.

Five entries on what may never go into a prompt, and how to treat what does.

The pasting rule.

GAI 081 Never Paste What You Would Not Publish

The method's tool-governance rule is not to paste screenshots, identifiers or notes containing personal details into tools your organization does not govern. When in doubt, assume the provider may log or reuse whatever you paste, and remember that deleting the conversation afterwards does not undo what was sent.

Safe substitutes.

GAI 082 Synthetic and De-identified Examples

The method's alternative is de-identified or synthetic examples wherever possible, with sensitive work kept inside internal, governed AI services, such as replacing a real patient's name, dates and record number with invented ones. De-identification is harder than deleting names; a rare condition plus a town can still identify a person.

Text that carries instructions.

GAI 083 Pasted Text Is Data

NIST's profile warns of prompt injection: text a model reads can carry its own instructions, and a twenty twenty-three study found such inputs blur data and instructions. A web page pasted for a summary might hide a line telling the model to change its answer. Treat pasted and retrieved text as data, and check any output that acts on it.

Secrets do not belong in prompts.

GAI 084 No Secrets in Prompts

The DAIR.AI guide documents prompt leaking, where attacks coax a model into revealing its hidden instructions, including anything confidential in them, such as an internal discount code in a shared assistant's system message. Write every prompt, system messages included, as if its users could read it.

A team's shared prompts.

GAI 085 A Team Prompt Library

The method calls its prompt patterns skeletons to copy and adapt, kept in a shared, living library where people already work, with each template carrying an owner, a last-tested date and the tool it was tested on. An unowned template drifts, and the one everyone copies is often the one nobody has checked recently.

Four procedures for the most common work: summaries, extraction, rewriting and messages.

Summaries for a named audience.

GAI 086 Summarizing for an Audience

The method's summary pattern turns a document into a short read for a named audience. Name the reader and what they will do with it. Paste the full source. List the points the summary must cover. Set the cap and register, such as plain language under five hundred words. Add the grounding line, and list anything unclear as a question. Summaries fail most often by omission, which no reader can see from the summary alone.

Pulling facts into fields.

GAI 087 Extracting Into Fields

Extraction pulls specific facts into named fields, and the DAIR.AI guide shows that the exact output format makes it reliable. Name every field with a one-line definition. Give the exact layout. Allow an honest blank by telling the model to write not stated when a value is absent. Paste the source below a separator. Then spot-check results against the source. Without the not-stated option, a model often fills a field with a plausible value.

Rewriting text safely.

GAI 088 Rewriting Without Changing Meaning

Rewriting changes tone, length or reading level while the meaning must stay the same. State the transformation precisely. State what must not change: facts, figures, obligations, deadlines and defined terms. Ask the model to list every change that could affect meaning. Then compare the rewrite against the original sentence by sentence. The danger is softened obligations: a must that becomes a should has changed the rule while sounding the same.

Drafting a message.

GAI 089 Drafting a Message

The course's message pattern has five steps. Say whether it is a reply, an announcement or a request. Give the context: who reads it, the relationship and the outcome wanted. Paste the message you are answering or the facts it must contain. Set constraints: length, tone, and what must or must not be promised. Ask for the draft plus a list of statements to confirm before sending. A draft message can make commitments, so read it as the recipient will.

Five procedures from the method's own work: scenarios, edge cases, change impact, high-risk flags and drafting from structured fields.

Realistic scenarios.

GAI 090 Generating Scenarios

The method's scenario pattern builds a small set of realistic situations from a definition or policy. Paste the rules the scenarios must follow. Ask for a few, such as three, involving typical use. For each, ask for the context, the key data values, and the rules that apply. Require at least one edge case. Then check every scenario against the rules, because a training example that teaches the wrong rule is worse than none.

The scenario that matters most.

GAI 091 Adding the Edge Case

The edge case shows where rules strain, so it gets its own check. Name the kind of edge you want: a boundary value, a missing field, a conflict between rules, or an unusual sequence. Ask which rule it tests. Check it could happen under your policy as written. Then ask a person who handles real cases whether they have seen something like it. Generated edge cases are often too ordinary or too unlikely; the practitioner makes one useful.

Assessing a change.

GAI 092 Assessing a Change

The method's impact pattern gives the model a proposed change and the list of things that depend on what is changing. Paste the reference list. Describe the change. Ask which items are most likely affected and why. Add the constraints: do not invent new policies, and mark uncertain links as uncertain. Then confirm each item with its owner. The model only knows the dependencies you pasted, so anything missing from your list is missing from its answer.

Flagging the high-risk areas.

GAI 093 Flagging High-Risk Areas

The method singles out money, safety, legal risk and customer trust as the areas where discipline pays most, and asks the model to flag them. Mark affected items that touch those areas. Route every flagged item to its owner for a decision, not only a review. Check the source policy directly for those items. Record who decided and why. A missing flag is not a clean bill of health; owners should scan the whole list.

Drafting from approved data.

GAI 094 Drafting From Structured Fields

The method's safest drafting use turns approved structured data into readable text, such as a notice built from a recorded decision and an approved template. Paste the fields and template. Ask for a draft filled from those fields only. Add the method's constraint: do not change the meaning of any legal phrase. Compare every figure, date and decision against the source. Then have the responsible person approve it. AI may draft the notice, but must not decide the outcome or interpret the law.

Three procedures for answering from documents, then two worked examples. The opening question is answered here.

Evidence before the answer.

GAI 095 Quote First, Then Answer

For questions about a long document, the DAIR.AI guide shows a two-prompt chain. First ask for the relevant passages, word for word, or a statement that none were found. Read the quotes and confirm they are real. Then ask the question with only those quotes, and require citations. Check that each claim traces to a quote. If the first step finds nothing relevant, the honest answer is that the document does not say.

Answering only from a source.

GAI 096 Answering Only From a Document

The guide's first defence against made-up answers is putting the ground truth in the prompt. Paste the authoritative source. Instruct the model to answer only from it. Ask for the passage behind each part of the answer. Say what to reply when the text does not cover the question. Then check the citations. Grounding lowers the error rate without removing errors, and research the guide draws on names traceability as an open problem.

Permission to say I don't know.

GAI 097 Letting the Model Say "I Don't Know"

The guide recommends telling the model it may admit it does not know, because without permission it tends to answer confidently anyway. State that I don't know is acceptable. Use a less varied setting for factual tasks where the tool allows one. Include examples of answerable and unanswerable questions. And treat each reported unknown as a question for a person. Permission reduces invention without guaranteeing the model knows its limits.

A checked chain, worked through.

GAI 098 A Summary Chain, Checked

A worked example, constructed for this course, follows a compliance officer summarizing a forty-page vendor contract for leadership in her company's governed tool. She first extracts the key clauses word for word. Checking them, she finds the renewal quote merged two sections, making renewal look automatic, and that the extraction missed an appendix on data handling. She fixes both before asking for a cited one-page summary. Both errors entered at the first step, and both were caught because quotes were checked before the summary was written.

An unsafe prompt, rewritten.

GAI 099 An Unsafe Prompt, Rewritten

A worked example, constructed for this course, rewrites a customer service prompt that pasted forty complaints, names, account numbers and a medication note into a free public chatbot with the instruction to fill in missing details. Five fixes follow. The work moves to the approved tool. Identifiers become placeholders and the medication note is removed. Fill in becomes use only what each complaint states, and list questions. The complaints are declared material, not instructions. And replies return as drafts with every promise highlighted. That answers this unit's question: safety comes from the tool, the data, the grounding, the data-not-instructions line and the review, not from the wording alone.

Look at what you can now do. You can use the method's grounding lines: use only what is given, invent nothing, ask instead of guessing, and mark uncertainty. You can decide what never goes into a prompt, and treat pasted text as data. You can run safe patterns for summaries, extraction, rewriting and messages. You can use the method's patterns for scenarios, impact and drafting from structured fields. And you can answer from documents safely and rewrite an unsafe prompt. Unit Five turns to verification.