The useful result
Produce policy-grounded reply drafts for five hard cases, each quoting a numbered policy line or escalating.
- Produce policy-grounded reply drafts for five hard cases, each quoting a numbered policy line or escalating.
A support chatbot or a helpful draft that says “Of course, we can refund that” when your policy says 14 days is more than rude, it is a promise. Language models fill gaps with plausible answers, which is exactly wrong when the answer is a rule.
Anthropic’s guidance on reducing hallucinations suggests three habits that fit support work: give the model permission to say it does not know, ground its answer in quotes from the supplied text, and tell it to use only the documents you provide. This lesson turns that into a repeatable reply routine.
The examples below are illustrative exercises, not recorded results from any tool. Sources were checked on 24 October 2026; menus, limits and prices change, so confirm details on the linked pages.
Before you start
- Your real refund, shipping or service policy as plain text
- Five tricky customer messages (names removed)
- A person who decides exceptions
Review evidence
Documentation-based review · 2026-10-24
Primary documentation and official guidance checked on 24 October 2026. Examples, checks and exercises are original illustrations, not recorded tool tests. Cover artwork is a conceptual illustration.
Number your policy lines
Paste the policy and give every rule a number, such as P1, P2, P3. Keep one rule per line. Numbering lets the model, and you, point to exactly which rule a sentence relies on. Remove anything that is out of date before you start.
Check your result
Every rule has a number, and there are no old or conflicting versions in the text.
Make every reply quote a rule, or escalate
Instruct the model to use only the numbered policy, to quote the rule number beside each commitment and to answer “Needs human” when the policy does not cover the case. Ask it to keep the tone warm without promising anything the policy does not say.
You draft customer support replies using ONLY the numbered policy below. Do not use outside knowledge. After each commitment write the rule number in brackets, like [P2]. If the policy does not cover the situation, write exactly “NEEDS HUMAN: [what is missing]” instead of a reply. Warm, plain tone, under 120 words.
Policy:
[paste numbered policy]
Customer message:
[paste message]Example / expected result
Illustrative example; not a recorded model result.
Policy P1 “Returns within 14 days”, customer asks on day 20. Expected: a polite decline citing [P1], or NEEDS HUMAN if an exception is possible.
Check your result
Every promise in a draft cites a policy number, and uncovered cases return Needs human.
Test five awkward cases and keep the good replies
Try the prompt on five difficult messages: outside the window, a damaged item, missing details, an angry message and a question the policy never answers. Fix the policy where a reply exposed a gap, and save the approved replies as your starting bank. A human still reads each one before sending.
Check your result
Each of the five cases returns either a cited reply or NEEDS HUMAN, and no reply promises something outside the policy.
When it goes wrong
If the policy is vague, replies will be vague. A model can still misquote a rule, so check the cited number against the text. Do not let it decide exceptions, refunds or legal matters, and do not paste private customer details into tools you have not checked.
Make it your own
Write a five-line policy for a fictional shop and test three customer messages: a late return, a damaged parcel and a question the policy does not answer. Mark each reply as cited or escalated.
Useful next resources
Sources and review
Last checked Oct 24, 2026. Features and account limits can change. Verify important results against your original material.
Did this help you complete your task?
Tell us what needs improving. No account or private output required.
Report incorrect information or a broken linkGenius Forges Editorial
Practical AI lessons with checked sources, labelled examples and human verification.
