Prompts the instructions sent to the model — editable here, no
deployment needed
What is and is not decided here. These texts tell the model how to
transcribe a call and how to judge it. They do not decide the score:
every number is recomputed in Python from config/rubric.json, question
counts override whatever the model claims, and a transcript that runs past the end of
its own audio is rejected in code. So an edit changes judgement and wording — it
cannot change the arithmetic, and it cannot switch off the safety checks.
A saved change applies to the next run; calls already scored stay as
they are until they are re-run.
Transcription
3255 characters ·
built-in, never edited
Sent with the audio. It decides how the recording is turned into English text, who is labelled agent or customer, and the JSON shape the rest of the system reads.
Before you edit. This prompt has invented conversations before. It must never be told the call's expected length or the speakers' names, and it must keep telling the model not to transcribe audio it cannot hear. Recordings under 5 seconds are refused in code, and any transcript running past the end of its audio is rejected in code, whatever this prompt says.
These must stay in the text, because the code that reads the
reply looks for them:
"turns" · "speaker" · text_native · text_en · primary_language · "connected"
The opening instructions for scoring a call against the pitch. The rubric itself - steps, weights, objections - is generated from config/rubric.json and is not part of this text.
Before you edit. Arithmetic is not done by the model: scores are recomputed in Python from the rubric weights, and question counts override whatever label comes back. Editing this changes judgement, not the maths.
Rules the rubric, the bands and the
lexicon — what the numbers are computed from
These are not wording — they are arithmetic. The prompts above tell the
model what to look for; these files decide what a score is. Every one is checked
before it saves: invalid JSON, a weight that is not a number, a negative weight, a
repeated step id or weights summing to zero are all refused with the offending step
named, and nothing is written.
A saved change applies to new runs at once. Calls already scored keep their old
numbers until they are re-scored — that reads the transcripts already on disk and
costs no tokens:
./cli.py rescore
Pitch rubric
14243 characters · 2026-10-31 15:38
The pitch steps, their weights and the score bands, per pitch type. This is what a score is computed FROM - the model reports which steps happened, and Python does the arithmetic from these weights.
Before you edit. Changing a weight changes every future score, and makes new scores incomparable with old ones. Calls already scored keep their old number until they are re-scored (./cli.py rescore), which recomputes from the transcripts already on disk and costs nothing.
Objections
3108 characters · 2026-10-31 15:38
The objection types the scorer recognises and how each is judged.
Before you edit. Affects objection handling rates on the overview and campaign pages.
Analysability bands
1143 characters · 2026-10-31 15:38
How a recording is judged scoreable at all - full conversation, short or unclear, or no conversation. Not a judgement of the agent.
Before you edit. Loosening this pushes thin calls into scoring, where they score low for lack of material rather than for anything the rep did.
Campaign to pitch type
2983 characters · 2026-10-31 15:38
Which pitch rubric each campaign is scored against. An unmatched campaign is transcribed but excluded from pitch scoring, rather than scored against a guessed rubric.
Before you edit. Most calls route here off a naming inference that business has not signed off. Mapping a campaign to the wrong pitch type scores those reps against a script they were never asked to follow.
Sentiment lexicon
5325 characters · 2026-10-31 15:38
The words and phrases that move customer sentiment, with weights. Sentiment is computed from this list alone - no model is involved, so the same words always give the same answer.
Before you edit. Short common words are dangerous here: 'no' once matched inside know, now, nothing and afternoon and produced 54% false positives until matching was made word-boundary aware.
