# Voice Extractor > Reads the indexed raw corpus and turns it into the atomic unit of a run: voice evidence cards, each with verbatim kept separate from interpretation, one confidence label, and a diffusing flag when a phrase lives in both user voice and creator voice. Also writes the forward voice map of what the market wants you to say. Triggers on: extract evidence cards, build voice map, mine the verbatim, turn sources into cards, what does my market actually say, separate quotes from interpretation, voice extraction. - Skill: `themattberman/voice-extractor` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add themattberman/voice-extractor` - Raw SKILL.md: https://api.skillmd.com/api/skills/themattberman/voice-extractor/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: TheMattBerman (https://skillmd.com/u/themattberman) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/themattberman/voice-extractor --- # Voice Extractor the extraction engine. it reads `raw/` plus `01-source-index.md` and produces the two artifacts the rest of the run depends on: the evidence cards (the atomic, traceable unit) and the forward-looking voice map. the core discipline: never smooth the market's words into your own. `verbatim` is the market's exact language, untouched. `interpretation` is your reading of it, in a separate field. if you cannot quote it, it is not an evidence card. this single separation is what keeps the kit from becoming another VOC summarizer that flattens "double-books room 3" into "scheduling concerns." ## When to use use when an indexed corpus exists (`raw/` and `01-source-index.md` present) and the operator needs evidence cards plus a voice map. invocable standalone or as Stage 2 of `frontrun`. do not use to map objections (`objection-mapper`), score angles (`angle-synthesizer`), or ingest sources (`source-intake`). ## Inputs -> `01-source-index.md` (the source rows and their reliability) -> `raw/` (the excerpt files every card must trace to) -> the scope from `00-intake.md` (market, business type, audience) for theme judgment ## Procedure ### 1. Read every source walk `raw/` source by source. for each, pull the candidate phrases that carry a tension: a fight, a frustration, a desire, a switch, a status play. flat, neutral, "it works fine" lines are not cards. if there is no tension, there is no card. ### 2. Write one card per piece of market language each card conforms to `voice-evidence-card.schema.json`. the field-by-field guide is in `references/extraction-schema.md`. the load-bearing fields: -> `verbatim`: exact, untouched. copy it, do not clean it. typos and all. -> `interpretation`: your reading. separate field. optional, but if present it must never be presented as the market's words. -> `confidence`: exactly one of `direct_quote` / `strong_inference` / `weak_pattern` / `do_not_use_yet`. the rules and examples are in `references/evidence-confidence.md`. -> `status_flags`: add `diffusing` when the phrase appears in BOTH creator voice and user voice. see the confidence reference. -> `voice_layer`: `user` (what the market feels) or `creator` (who shapes how the market talks). -> `tension`, `why_it_matters`, `do_not_claim`: all required, all one line. `do_not_claim` is the anti-overclaim field: what this evidence does NOT prove. every card's `source_id` must resolve to a row in `01-source-index.md` that resolves to a file in `raw/`. a card you cannot trace does not get written. ### 3. Cluster without flattening near-identical phrasings cluster, but clustering must not erase the sharpest verbatim. the rubric is in `references/voice-cluster-rubric.md`. the rule: preserve the sharpest verbatim as the cluster exemplar, count the recurrence, but never replace the exemplar with a smoothed label. "removed CSV export" and "they killed the export button" cluster, but you keep both verbatims and lead with the sharper one. ### 4. Emit 02-evidence-cards.json a JSON array of cards. it must validate against the schema: enum fields legal, `verbatim`/`tension`/`why_it_matters`/`do_not_claim` non-empty, `usable_for` at least one entry, `evidence_id` as `ev_001` ascending, `source_id` matching the index. ### 5. Write 03-voice-map.md (FORWARD) the forward-looking map: what the market wants you to SAY. organize into: -> **desires**: what they are actually trying to get (the outcome, in their words) -> **myths**: what they wrongly believe about the category or the product -> **enemy beliefs**: the specific wrong thing worth attacking (feeds the angle slate's `enemy_belief`) -> **buying language**: the words they use when they are close to a decision -> **emotional language**: the charged, un-inventable phrasing a copywriter would not dare make up every entry in the voice map cites the `evidence_id`s behind it. no broad label without the verbatim that proves it. hold the boundary: the voice map is FORWARD (what to say). objections, anxieties, workarounds, and trust gaps belong in `04-objection-map.md` (what to answer), which `objection-mapper` owns. never write the same cluster into both. ## Outputs | File | What it carries | |---|---| | `02-evidence-cards.json` | the atomic unit. array of cards validating against `voice-evidence-card.schema.json`. everything downstream cites these. | | `03-voice-map.md` | FORWARD: desires, myths, enemy beliefs, buying language, emotional language. every entry cites `evidence_id`s. | ## References - `references/extraction-schema.md`: the card field-by-field guide in prose, mirroring the JSON schema. - `references/evidence-confidence.md`: the four confidence labels plus the diffusing flag, when each applies, with examples. - `references/voice-cluster-rubric.md`: how to cluster near-identical phrasings without flattening, and the user-voice vs creator-voice extraction differences.