Bolna Voice Prompt Engineering

This skill produces production-ready prompts for Bolna voice agents running on GPT-4.1 mini. The format is rigid by design — GPT-4.1 mini executes these prompts at call-time, and any drift in structure, variable notation, or language ordering degrades reliability. Treat the rules in this file as load-bearing.

What this skill covers

Two workflows, both triggered by the user:

1. New prompt from a use-case description. The user describes a calling scenario (recruitment screening, appointment booking, NPS, renewal, lead qualification, etc.) and any preloaded variables they have. Output: a full prompt following the exact structure below.
2. Fixing an existing prompt from a transcript. The user pastes a call transcript (sometimes alongside the current prompt) and points at problems. Output: either the patched section(s) or a full updated prompt, with a short change-note. See references/transcript-workflow.md for the diagnostic approach.

If agent gender, primary language, or the primary call goal is unclear from the user's brief, ask one focused clarifying question before drafting. If those three are clear, draft the whole prompt straight away — don't pepper the user with questions for things you can reasonably infer (call direction, role title, branch coverage).

Before drafting: always load the static modules

Two subsections in Section 1 — # Conversational Naturalization (Acknowledgements & Flow Softening) and # Pronunciation and Script Normalisation — must be copied verbatim into every prompt. Read references/static-modules.md and paste the contents into the new prompt. Do not paraphrase, summarise, shorten, or "improve" these modules. They are tuned and changes break agent behaviour.

The mandatory section structure