Voice AI technology

Writing Prompts for Better Call Summaries: Templates and Tests

How to write a prompt that turns a call transcript into a short, accurate note: who it is for, what to include, rules against invention and how to test it.

Call editorial team · Published · 6 min read

An AI summary can save a team minutes after every call, but only if the summary is useful. A vague paragraph that repeats the greeting does not help the person who must call the customer back. The difference between a useful summary and a useless one is usually the prompt: the instruction you give the model about what a summary is for.

A support agent typing a note after a phone call
Illustrative image

This guide shows how to write that instruction for call summaries, with templates you can adapt, common mistakes and a way to test. It goes with AI Call Summaries: Cutting After-Call Work and, for prompting in general, Prompting an LLM Phone Agent.

Start with the reader

A summary is written for someone. Decide who and what they will do with it. A team lead scanning many calls wants a line or two. An agent about to call back wants the caller's request, the promised action and the deadline. A compliance reviewer wants the facts without opinion. One prompt cannot serve all three equally, so choose the main reader and write for them.

Questions the summary should answer

  • WhoThe caller and what is known about them.
  • WhyThe reason for the call in plain words.
  • What happenedWhat the agent or system said or did.
  • OutcomeResolved, message taken, callback needed or handed over.
  • Next stepWhat must happen, by whom and by when, if stated.

The anatomy of a good prompt

Parts of a summary prompt

  1. Role and readerSay what the summary is for and who will read it.
  2. InputExplain what the model receives: a transcript with speaker labels.
  3. Output formatList the fields, order and length.
  4. RulesState what to use, what to avoid and how to mark doubt.
  5. ExamplesShow one or two short, good outputs.

Here is a plain-language skeleton you can adapt. It is written as instructions, not as code.

Rules that make summaries trustworthy

Use only what was said

Models tend to fill gaps with plausible details. Say clearly that the summary must come from the transcript and that missing items should be marked not stated. This single line prevents many of the worst errors. For the broader issue, read Preventing Made-Up Answers in Voice AI.

Mark doubt

Transcripts contain recognition errors, especially for names and numbers. Instruct the model to flag figures or names that look uncertain, for example by writing the item followed by unverified. A flagged number sends the reader to the recording; a confident wrong number sends them to the wrong place.

Keep it neutral

Ask for facts, not interpretations. Summaries that label a caller as rude or confused can be unfair and could be seen by the customer if they request their data. Describe behaviour only when it matters to the outcome.

Set a length

Without a limit, summaries grow. A fixed number of lines or a rough word limit keeps them scannable. Longer calls may justify a longer note, so you can set a limit that scales.

Name the language

State the language of the summary. If calls happen in several languages but staff read in one, say so explicitly, and test with calls in each language.

Different summaries for different jobs

Two styles of summary

Quick scanFollow-up note
ReaderA manager reviewing many callsAn agent about to act
LengthOne or two linesFour to six short lines
ContentsTopic and resultRequest, details, promise, next step
Prompt emphasisBrevity and a fixed vocabularyCompleteness and exact details

You can keep both: a short line for the call list and a fuller note for the person doing the follow-up. Many teams also add tags, such as complaint, booking request or billing, chosen from a list you define. Giving the model a fixed list of allowed tags avoids dozens of near-duplicates.

Testing and improving the prompt

Prompts are improved by trying them on real material. Build a small set of transcripts that covers your call types: a simple question, a message with details, a complaint, a call with a misheard name and a call that ends without a result. Run the prompt on all of them after each change and compare the outputs.

What to check in each summary

  • Is the reason for the call correct?
  • Are names, numbers and dates right or marked unverified?
  • Is anything stated that the transcript does not support?
  • Is the outcome clear?
  • Is the next step clear, with a person or time where stated?
  • Is the length within the limit?
  • Would the follow-up person need to listen to the call anyway?

Sample the results by hand even when things look fine. A regular check that marks summaries as accurate, partly accurate or wrong gives you a running measure of quality and shows when a change made things worse.

Before and after

A weak prompt says summarise this call. It tends to produce a paragraph such as the customer called about an issue and the agent helped them, which tells the reader nothing. A stronger prompt produces something like this, for a hypothetical call about a delayed order.

The second version took a few more lines of instruction, yet it saves the next person from replaying the call. It also shows the value of the phrase not stated: the summary does not invent a delivery date that nobody gave.

Privacy and handling

Summaries are personal data too, since they describe a caller and what they said. Keep them as short as the purpose allows, keep sensitive details out where you can, and apply your retention rules to summaries as well as to audio and transcripts. See Privacy by Design for Voice AI.

Summaries in Call

In Call, the call log shows a summary at the top of each call, with the transcript below, and offers to draft a note, a follow-up email or a suggested outcome from the call. Drafts written by a connected model are marked as such and may need correcting, so read them before using them. In the Automatic review area, managers can see the wording that is sent to the model, edit it and save it as a new version, or go back to the default. A sample of recent summaries can be marked as accurate, partly or wrong, which feeds a running view of summary accuracy. You can also have the summary of a finished call emailed to your own account address. Use the structure in this guide as the starting point for your own wording, and see Call Summaries in Call for the product steps.

Frequently asked questions

What should a call summary include?

The reason for the call, what was said or done, the outcome and the next step, with names and numbers only if they were clearly stated.

How long should a summary be?

Long enough for the reader to act without replaying the call, and short enough to scan. A few short lines is a common target; set a limit in the prompt.

How do I stop the model inventing details?

Tell it to use only the transcript, to write not stated for missing items and to mark uncertain names and numbers. Then check a sample by hand.

Should the summary be in the caller's language?

Write it in the language your team reads, and say so in the prompt. Test with calls in each language you receive.

Can I use different prompts for different calls?

Yes, for example a short scan line and a fuller follow-up note. Keep the instructions consistent and version them.

How do I know if the summaries are good?

Review a regular sample against the transcript, marking each as accurate, partly accurate or wrong, and compare the results after any change to the prompt.

Review call summaries and their wording in Call

In the demo console you can try incoming calls in a simulation.

Related guides