When should you use it?
- Your support bot draws on answers agents saved from solved tickets
- You are about to widen what the bot can retrieve
- Customers have received replies that mention other people or accounts
Not for: Help-center articles that were written generically from the start; Bots that only quote official documentation
What do you need, and what do you get?
You need
- An export of every saved answer the bot can retrieve
- Your help-center links and current policies, to check conflicting answers
- Someone who can rewrite and re-save answers in your support tool
You get
- A risk-ranked list of saved answers with what is wrong in each
- A generic rewrite for each risky answer
- Conflicting answers merged, with policy questions for the team
- Rules for writing future saved answers
How does it work, step by step?
- Export the setPull every saved answer your bot can retrieve into one file.
- Run the reviewUse the prompt below to flag personal details, account links, claims of work done, dates, unstated assumptions and conflicts.
- Fix the worst firstRewrite or delete the most dangerous answers first, keeping the part that applies to everyone.
- Re-indexAfter each rewrite, make sure your tool recalculates the search index for that answer, or the old version can still be matched.
- Change the save habitShare rules with agents, starting with one test: would this still be right if the bot sent it to another customer?
The prompt
Our own prompt, inspired by Pierre-Laurent Medori and written to work in Claude, ChatGPT and Custom agent. Values in lime are placeholders.
Fill in the blanks (optional; the copied prompt uses your values)
Below are saved support answers that our AI assistant can reuse for any customer.
{saved_answers}
Read each one as if the bot were about to send it to a different customer. For each, list anything that would be wrong in that case: names, emails, domains or IDs; links into one customer's account; statements that someone did or tested something; dates or deadlines; statements made on a customer's behalf; and assumptions about one setup.
Also find answers to the same question that disagree.
Return a list sorted from most to least risky with: the answer's ID, the problems, a rewritten version that works for anyone (use menu paths instead of account links, and cover each case if the right reply varies by plan or setup), and any policy question we need to settle. Mark answers that are already fine as keep.Inspired by Pierre-Laurent Medori: A saved reply is written for one customer. Your support bot can send it to anyone.. The prompt above is written by playbooks.bot in our own words; nothing from the original is quoted.
What can go wrong?
- Pasting saved answers into an AI tool shares customer data; use an approved tool or remove details first.
- If the search index is not refreshed after a rewrite, the old wording can still be what gets matched; check it.
- Two answers that disagree may reflect an unclear policy; ask the team instead of picking one.
- Do not delete answers in bulk; most contain a useful core worth rewriting.
Our take
A sharp, honest engineering write-up about a risk many teams have not noticed. The save button in a help desk turns a reply to one person into knowledge for everyone, and the post shows exactly what that drags along: names, links into one account, things done once, statements made on someone's behalf, dates that expire and assumptions about one setup. It also reports real numbers from the cleanup and the bot's use afterward, including where the evidence is still unclear. The post has no prompt, so we wrote one for the audit. We tested it on six synthetic saved replies for a fictional software company, including a signed reply with a customer's domain, a relative deadline, two answers that contradicted each other and an appeal letter in one person's name. It ranked the appeal letter as the top risk, proposed generic rewrites with menu paths, merged the contradiction into one answer with both plan cases and a question for the team, and left the clean answer alone.
Where the time goes back: One focused review prevents the slow, embarrassing cleanup that follows a bot sending one customer's details or promises to another.
FAQ
Why are saved replies risky?
Did loose matching cause harm?
Is the prompt from the post?
Source and credit
Inspired by Pierre-Laurent Medori. We describe the idea in our own words, link to the original, and wrote the prompt ourselves; nothing is quoted.
- Inspired by
- A saved reply is written for one customer. Your support bot can send it to anyone.
- Creator
- Pierre-Laurent Medori
- Published on
- DEV Community
- Original date
- Oct 14, 2026
- License
- Not openly licensed (the poster keeps all rights), so we quote nothing and link to the original
- Originally for
- Claude
- AI dry run
- Oct 22, 2026: passed. Dry-run tested by our AI test agent on sample input. Method from the post; prompt is ours.
AI dry run on sample data; not run with live connected accounts. The playbooks.bot test agent filled the prompt with synthetic inputs, produced the output and checked it against the page's stated outputs and guardrails. - Verified
- Oct 23, 2026: source link resolves to the original, the credited creator published it, our excerpt or prompt matches it, and the license allows our use.
Are you the creator? Claim or correct this page.