What context does Callmine give the LLM for each call?

The full call transcript, call metadata (date, duration, participants, rep identity), and any HubSpot deal context Gong has linked — deal stage, amount, owner, company, and custom properties. Your prompt is applied on top.

How long should my prompt be?

Whatever the question requires. A short prompt (one paragraph) works for simple categorization. Complex multi-dimensional analyses run 200-500 words. Past 1,000 words, results often degrade — the LLM weights the latest instructions more than earlier ones.

Can I ask for structured JSON output?

Yes. Specify the output schema in your prompt — JSON keys, expected types, allowed enum values. Callmine extracts the JSON from each per-call response and surfaces it in the report's JSON payload download.

What's the executive summary directive?

An optional textarea below the deal-level prompt. It tunes the cross-call roll-up — what themes to highlight, what counts to surface, how to phrase the top-line insight. Leave it blank to use Callmine's default.

How do I write a custom analysis prompt?

Analyses·May 31, 2026·2 min·By Callmine Team

Write a custom prompt at /new in plain English. Add an executive summary directive to tune the roll-up. Test on 10 calls first.

TL;DR

At /new, choose Write your own prompt. Write the question Callmine should answer for every call in scope, in plain English. Optionally add an executive summary directive to tune the roll-up. Start with 10 calls to validate before scaling up.

Steps

01Open /new and click Write your own prompt on the start-mode chooser. The deal-level prompt textarea appears.
02Write the question Callmine should answer for every call. Be specific about the output you want — categories, scores, named themes, or a JSON shape if you'll process the result downstream.
03Optionally expand Executive summary settings and write a directive that focuses the cross-call roll-up. Leave blank to use Callmine's default schema-contract prompt.
04Pick a model. Per-plan options are gpt-4o-mini, gpt-4o, and gpt-5.4-nano. Default is gpt-4o-mini for cost; upgrade to gpt-4o when output quality matters more than per-call cost.
05Set max_calls to 10 and pick a recent date range. Click Run. Read the report, iterate on the prompt, re-run on the same window. Scale up once the 10-call run produces the output you want.

Callmine /new page with the deal-level prompt textarea visible after selecting Write your own prompt — After picking Write your own prompt, the deal-level prompt textarea appears alongside the run controls.

What the LLM sees per call

For each call in scope, Callmine sends the LLM:

·The full transcript text, speaker-attributed.
·Call metadata — start time, duration, participants, the rep who ran the call.
·HubSpot deal context, if linked — deal stage, amount, owner, company, custom properties Gong has logged.
·Your prompt, appended after the context.

The LLM responds once per call. Callmine collects every response, extracts structured output, and feeds the set into the executive summary stage.

Prompt-writing tips

·Ask for one thing per call. Multi-axis prompts ("classify by X and Y and Z") work but degrade faster than focused prompts.
·Specify the output format. Without it, the LLM picks a format, and it varies call to call. Schema-typed JSON gives the cleanest reports.
·Test on 10 calls before scaling. Prompts that produce great output on three hand-picked calls sometimes fail on the long tail.
·Iterate, don't rebuild. Re-run with the same date range and a modified prompt — the diff is much easier to spot.