# Voice configuration and tuning Turn the user's business material into a testable voice configuration. Work inside their chosen project and use their actual tools. Use the shared developer/runtime Skill distinction in the main entrypoint. ## Establish a useful baseline Inspect existing instructions, runtime Skill text, tool schemas, approved business documents and a few failed conversations. Preserve the working configuration, version and evaluation results before changing it. If nothing exists, start with a minimal draft. Clarify only consequential missing facts; label unknown policy and unavailable capabilities instead of inventing them. Read [design.md](design-methods.md) for business decomposition and prompt/tool contracts. For an existing failing flow, start from the failed decision node, not a complete prompt rewrite. Prefer small configuration patches and explicit flow coverage. Evaluate model choice and latency targets against the application’s own test cases. ## Separate procedure from facts Use Prompt for identity and global constraints, Skill for SOP steps, and Knowledge for factual policies/QA. Read [knowledge.md](knowledge.md) when the task includes factual reference material. Keep the existing custom model workflow; do not create a second independent Agent resource. Search accuracy and corpus rewriting are separate from API integration unless requested. ## Produce configuration that can be used - Keep global instructions short: identity, scope, speaking style, precedence, shared tool truthfulness and ending rules. - Put domain-specific procedures in versioned runtime Skills with clear triggers, required fields, branch/stop conditions and source references. Use actual supported tool names; a written procedure is not an executable tool. - For LatenTX, read https://api.latenth.com/llms.txt and the current Realtime/custom-model references. A plain call needs only model + instructions. Runtime Skills use `session.latentx`; tools use `session.tools`. Saved custom models are optional reuse, with draft → validate → publish and real returned version IDs. - Saved custom-model configuration selects `base_model` from `GET /voice/v1/catalog`. Use that model's supported languages and voices; do not generate separate `llm`, `asr` or `tts` component selections. Keep session examples distinct from custom-model request bodies. - Start from [assets/session.example.json](../assets/session.example.json) for a fictional, no-tool service desk. It is a valid session.update session object, not a custom-model PUT payload. Adapt it to the user's real material; don't publish the example as their business policy. - Read [evaluation.md](evaluation.md). Produce a small representative evaluation set with expected next actions and prohibited claims, then run only the requested/authorized test scope. ## Handoff Deliver the changed instructions, runtime Skill files, tool/MCP contract mapping, session or custom-model configuration, source/uncertainty notes, and evaluation cases/results. For a narrow change, deliver only the affected pieces plus regression evidence. Mark drafted, schema-validated, text-tested and human-audio-tested states separately. Preserve unrelated fields and resource references when updating a config; publish only within the user's authorized scope. Do not transfer private customer recordings, exact business policies or personal details into public examples. This Skill can prepare/configure tools but does not authorize sending messages, placing calls, issuing refunds, dispatching services or changing production traffic. Respect existing authorization; do not add repetitive approval steps where the user has already authorized the action.