vac-task-extractionlisted

Generates a task-extraction-prompt.md for a specific voice agent build. This prompt is executed at runtime (after each call) to analyze transcript + call summary and produce CRM tasks for human follow-up. Use at step 8 of the VAC pipeline (see PIPELINE.md), after vac-test. The output file contains ONLY the bare prompt, copy-paste ready; applied defaults and notes are reported in chat, not written into the file. Default output is a prose title+description task; only when the briefing prescribes or clearly implies a structured multi-field task does it also emit a conditional extraction-schema.json (the per-campaign structured-output schema your runner uses to shape the output). Trigger whenever building a voice agent and a post-call CRM task extraction is needed, or whenever the user asks to create a task extraction prompt, post-call processing, CRM task generation, or transcript analysis for a voice agent build.
callflows/voice-agent-kit · ★ 1 · Data & Documents · score 74

Install: claude install-skill callflows/voice-agent-kit

# VAC Task Extraction Produces a `task-extraction-prompt.md` that runs after every call to extract CRM tasks from transcript + call summary. The prompt is agent-specific: its task logic, information requirements, language, and tone are derived from the agent's briefing. ## Prerequisites (input check) Before starting, check that these files exist in the agent directory (`../voice-agents/<agent>/`): | File | Comes from | If missing | |---|---|---| | `briefing.md` | `vac-intake` (step 1) | Run `vac-intake` first | If an input is missing, do NOT improvise and do NOT carry on with placeholders: name the missing upstream step and ask the requester or workspace admin whether it should be run now. (On top of that, this skill enforces three MUST-HAVE preconditions on the briefing's content, see below.) ## Pipeline Position Step 8 (see `PIPELINE.md`) — runs after `vac-test`. It needs only `briefing.md` as input, so it can technically run any time after `vac-prompt` is finalized; the canonical position is step 8 because live-test findings can sharpen the no-task logic. Input: `briefing.md`. Output: `task-extraction-prompt.md` (the bare prompt only). ## What you're building A **system prompt for an LLM** that will be executed at runtime. At runtime, this LLM receives: - Lead data (may be partial): `vorname`, `nachname`, `unternehmen`, `telefonnummer` - Full call transcript - Call summary It outputs a **JSON array** of task objects — or an empty array if no follow-up is needed: