20-Minute Customer Call Triage Agent: Full Recipe

Drop this into your agent stack today. Stack is whatever you already use for transcripts plus Claude with the customer-call-notes skill (or any general-purpose LLM with a tight system prompt). Output goes to whatever discovery ledger you already keep.

Part 1: The Brief File

Before you run the agent, write a short brief file (50-150 words). The agent reads it as context before processing the transcript. Without the brief, the agent reverts to generic synthesis.

# Brief: [Customer Name], [Date of Call]

## Account context
- Tier: [Enterprise / Mid-market / SMB]
- Stage: [New / Active / Renewal coming / At risk]
- Champion: [Name, role]
- Last meaningful interaction: [date, what happened]

## Product context
- What we are currently considering shipping:
  1. [Bet one, one sentence]
  2. [Bet two, one sentence]
  3. [Bet three, one sentence]

## What I want from this call
[One sentence, specific. Not "everything." Something like: "I want to know whether the bulk-export friction is real enough to justify a prototype this sprint."]

The brief is reusable. Keep one brief file per account. Update it after every call.

Part 2: The System Prompt

You are a customer call triage agent. Your job is to take a transcript of a customer call and produce four structured outputs.

CONTEXT:
- You will be given a brief file describing the customer, the product context, and what the PM wants from the call.
- You will be given a call transcript (with speaker labels if available).

YOUR OUTPUTS (return all four in a single response, in this exact order):

1. VERBATIM QUOTE BANK
   - 8 to 15 quotes from the customer.
   - Verbatim, not paraphrased. If you cannot find a verbatim quote, do not include it.
   - Each quote includes the timestamp (if available) and a one-line tag describing the topic.
   - Organize by topic, not by chronological order.

2. THEME ROLLUP
   - 3 to 5 themes the call surfaced.
   - Name each theme as a noun phrase, not a category. Example: "Bulk export friction blocking weekly reporting" not "Product feedback."
   - For each theme, include: frequency (number of quotes), confidence (how clear the customer was: low / medium / high), and which active bet (from the brief) it connects to (if any).

3. RECOMMENDED NEXT ACTION
   - One sentence.
   - Concrete enough to put on the PM's calendar this week.
   - Bad action: "Follow up with the customer." "Investigate further." "Schedule a deep-dive."
   - Good action: "Ship a 90-minute prototype of the bulk-export flow and send it to [champion name] by Thursday."
   - If you cannot produce a concrete action, say "No action recommended" and explain why in one sentence.

4. DISCOVERY LEDGER ROW
   - A single row formatted for the PM's discovery ledger.
   - Columns: Account | Date | Themes (comma-separated) | Top quote | Confidence | Linked bet | Owner | Next action | Action deadline
   - Use the data from outputs 1, 2, and 3.

SELF-CHECK BEFORE YOU RETURN:
- Are all quotes verbatim? (If you paraphrased, replace with verbatim.)
- Do the theme names read as specific patterns, not generic categories?
- Is the recommended action specific enough that the PM could put it on the calendar?
- If any of these fail, fix before returning.

Return all four outputs. Do not include preamble or commentary.

Part 3: The User Prompt

Two parts. The brief file, then the transcript.

<brief>
[Paste brief file here]
</brief>

<transcript>
[Paste transcript here]
</transcript>

That's it. Send.

Part 4: The Eval Rubric

Run this on the first 10 calls. Then weekly.

Check	What good looks like	Threshold
Quote accuracy	Pick 3 quotes, find them at the timestamp in the transcript, verbatim	95% verbatim across 10 calls
Theme legibility	Show theme names to a senior PM who didn't sit on the call. They can describe what the customer cares about in 30 seconds.	Passes on 8 of 10 calls
Action specificity	The action could be put on a calendar this week with no further interpretation	100% (rewrite any vague action)

Below threshold on any check, re-tune the system prompt or change the model.

Part 5: The Batch Run Recipe

For clearing a backlog of unprocessed calls (the highest-ROI use).

1. Export all unprocessed transcripts to a folder.
2. Generate or update brief files for each account represented.
3. Run the agent on each transcript with its matching brief.
4. Read only the recommended actions and the theme rollups. Skip quote banks for now.
5. Cluster themes across calls. Pull the 3 to 5 themes that recur most.
6. Pick the highest-recurring theme. Ship a prototype against it within the week.

Time budget for a 40-call backlog: 90 seconds per call agent runtime (60 minutes total), plus roughly 5 minutes per call of recommended-action review (3.3 hours), plus 30 minutes of cross-call theme clustering. Total: about 4 hours of human time.

Part 6: Common Failure Modes

• Cross-talk in the transcript. Use a transcript tool with speaker diarization (Otter, Fireflies, Gong). Verify speaker labels before running.
• Customer diplomacy taken at face value. The agent will tag "we'd love to use it more" as a positive signal. The PM has to know it usually means "we are not using it much." Add an instruction to the system prompt: "When the customer is being diplomatic, flag the underlying signal, not the surface statement."
• Generic theme names. If themes come back as "product feedback" or "feature request," your brief is too thin. Add specifics about active bets and recent decisions.
• Vague actions. If you keep getting "investigate further," add to the system prompt: "If you cannot produce a concrete action, say 'No action recommended' and explain why in one sentence." This is better than a vague action that looks like progress.

Part 7: Discovery Ledger Schema (Sample)

| Account     | Date       | Themes                                    | Top quote                                                | Confidence | Linked bet         | Owner | Next action                                                              | Deadline   |
|-------------|------------|-------------------------------------------|----------------------------------------------------------|------------|--------------------|-------|--------------------------------------------------------------------------|------------|
| Acme Corp   | 2026-05-27 | Bulk export friction, Slow Monday rollup | "We export every Monday morning and it takes 40 minutes." | high       | bulk-export-v2     | Falk  | Ship 90-min prototype of bulk-export flow, send to Sarah by Thursday    | 2026-05-30 |

Use this schema or whatever your team uses. The point is that the agent's last output drops into the row without re-typing.

---

That's the recipe. The agent is 90 seconds. The brief is reusable. The eval is three checks. The batch run is the afternoon that pays for the whole stack.