Skip to content
PromptsBuddy
All episodes
The PromptsBuddy Podcast cover art
Season 1 · Episode 04August 21, 202648 min

Podcast episode

The support team that writes its own AI policy

The PromptsBuddy Podcast

Why support prompts are policy documents, how they get reviewed, and the one rule that stops agents from promising ship dates.

Audio drops soon

The support team that writes its own AI policy

Why a support prompt is a policy document

Audio drops soon — subscribe and it lands when it goes live.

0:00—:—

Chapters

  1. 00:00Why a support prompt is a policy document
  2. 06:40The week agents promised three ship dates
  3. 14:05Building the AUTHORITY field
  4. 23:30What the team refused to automate
  5. 31:15Measuring de-escalation without a survey
  6. 41:50What to tell a team starting today

Show notes

A mid-size support desk can handle thousands of tickets a month. When a third of replies start as a model draft, every one of those drafts should go through the same review process as a change to the refund policy — because that is what the prompt is.

The moment it got serious

Six weeks into one rollout, three separate agents promised customers a ship date for a feature that had no ship date. None of them invented it maliciously. All three used a prompt that said “be helpful and reassure the customer,” and the model filled the gap the way models do.

The fix was not a training session. It was a field in the prompt — an explicit statement of what the agent is authorised to offer — and a rule that anything outside it is a bug in the prompt, not a mistake by the agent.

What stays human on purpose

Cancellations. Anything involving a death or a medical situation. Anything where the customer used the word “lawyer.” The value of a human reply in those moments is not the words. It is that a person read it, and the customer can tell.

On measurement

Skip the vanity survey. Count how many threads reach a fourth reply, and how many get escalated to a human lead. If both go down after you add authority fields and refusal rules, you are measuring something real. Still: do not over-claim causation.

What to tell a team starting today

Write the authority list before the tone guidelines. “Be warm” without “never promise dates” is how you get warm lies at scale.

Mentioned in this episode

Prompts from this episode

SupportintermediateStaff pick

The support reply that de-escalates without grovelling

Drafts a response to an angry customer that acknowledges the real complaint, commits to something specific, and never apologises for things that aren't your fault.

ClaudeChatGPT3 vars
Open prompt
WritingbeginnerStaff pick

Rewrite until it stops sounding like AI wrote it

A diagnostic-then-edit pass that names the specific tells in a draft — the tricolons, the hedges, the empty transitions — and fixes them without flattening your voice.

ClaudeChatGPT3 vars
Open prompt