transcript-corrector

name: transcript-corrector description: | Use when you have a transcribed text (meeting transcript, voice-note transcript, call transcript, dictation) that may contain ASR (automatic-speech-recognition) errors — phonetic substitutions, typos, grammar drift, punctuation loss — and you need a corrected version BEFORE downstream consumers (summarizers, ticket creators, action-item extractors) ingest it. Multi-lingual pt-BR + en-US. Cross-vendor AAIF compatible. Founding empirical case: Loom/equivalent transcribed "Nelcael Alves Ferreira" as "Nilson" in a Vek daily meeting; downstream consumers propagated the wrong name until corrected. This skill closes that gap. triggers: - corrigir transcrição - corrigir transcricao - normalizar transcrição - sanitizar transcrição - correct transcript - fix transcript - correct meeting transcript - apply transcript correction - check ASR errors - phonetic substitution check - participants whitelist check version: 1.0.0 allowed-tools: - Read - Write - Edit - Bash - Glob - Grep

transcript-corrector — Skill

Cross-vendor AAIF skill. Detects + corrects ASR-class errors in transcribed text (phonetic substitutions, typos, grammar drift, punctuation loss). Emits corrected text + side-by-side diff + per-correction confidence audit. Multi-lingual pt-BR + en-US. Atomic + generic (Goldilocks): works on any transcribed text, not only meetings.

Purpose (1-line)

Close the ASR-error gap between raw transcript and downstream consumers, so action items, decisions, and ticket assignments never propagate phonetically-mis-recognized proper names or systematic typos.

When to invoke

• Loom / Otter / Tactiq / Fireflies / Whisper transcript ingested into Confluence / GDrive / Notion / local file — about to feed into a summarizer, ticket extractor, or other downstream consumer
• Daily-meeting page authored by ASR + needs review before stakeholder distribution
• Voice-note dictation needing cleanup
• Empirical signal: speaker attribution looks "off" (rare name replaced by common one) or operator says "isso não é o nome certo"

When NOT to invoke

• Hand-typed text (operator wrote it; no ASR layer) — only the grammar pass would apply, and that's overkill
• Encrypted / PII-bearing content where corrections might leak personal data — escalate to operator + apply per-field redaction first
• Transcripts in languages other than pt-BR or en-US (v1.0.0 scope; extension via PDCA)

4-pass pipeline

Pass 1 — Participants whitelist check (Levenshtein + Metaphone)

For every detected speaker name OR @mention in the transcript:

1. If the name is in catalogs/participants-canonical.yaml → mark verified
2. Else compute Levenshtein distance to every whitelist entry AND phonetic-code distance (Metaphone for en-US, Portuguese-Metaphone heuristic for pt-BR; falls back to difflib.SequenceMatcher ratio if phonetics lib unavailable)
3. Score = combined (1 - levenshtein_normalized) * 0.5 + phonetic_match * 0.5
4. If score ≥ 0.85 → HIGH confidence → auto-apply correction
5. If 0.65 ≤ score < 0.85 → MEDIUM → flag-only in audit (no auto-apply); operator reviews + manually promotes (conservative-by-default per operator's Q2 draft-first preference in plan)
6. If score < 0.65 → LOW → leave original, emit :warning: in audit + suggest alternatives
7. Also consult catalogs/phonetic-substitutions.yaml for known class-based substitutions (e.g., [Nilson|Nielson|Nelson] → Nelcael under context "Vek daily + Invitee mismatch")

Pass 2 — Common-typos dictionary

Search-and-replace from catalogs/common-typos-{pt-br|en-us}.yaml. Each entry is a {wrong, correct, context?, confidence?} tuple. Default confidence 1.0 unless specified. Applied case-preserving (e.g., Sala → Sala not sala).

Pass 3 — Grammar / punctuation

Capability-detected:

• If LANGUAGETOOL_API env var is set (e.g., https://api.languagetool.org/v2 for community-edition OR a self-hosted instance), POST the transcript and apply rule-based corrections (filtered to safe categories: PUNCTUATION, TYPOGRAPHY, CASING; skip STYLE to avoid over-rewriting)
• Else fall back to bash + regex heuristics:
  • Capitalize first letter after sentence-ending [.!?]
  • Collapse double-spaces
  • Ensure terminal . at end of paragraphs that lack final punctuation
  • Fix common pt-BR contractions (do(a) → do/da left untouched; only obvious typos)

Pass 4 — Emit

Output three artifacts:

1. Corrected text (markdown OR plain, by --format)
2. Side-by-side diff (unified diff or 2-column inline-comment style)
3. Audit JSON at pdca/run-<timestamp>.json capturing every correction with {line, original, corrected, confidence, source-pass, rule} — used to feed catalog growth in subsequent PDCA cycles

Invocation

# Read-only / review-only mode (default — never writes to source)
bash skills/transcript-corrector/scripts/correct.sh \
  --source <path|stdin|confluence:<page-id>|notion:<page-id>|gdrive:<file-id>> \
  --language <pt-br|en-us|auto> \
  --mode review-only \
  --output stdout

# Inline mode (replace source — REQUIRES HITL auth per [C07] v2.1.0)
bash skills/transcript-corrector/scripts/correct.sh \
  --source confluence:303988750 \
  --language pt-br \
  --mode inline \
  --confluence-write-auth "operator-approved-2026-05-28"

Catalogs (versioned alongside skill)

• catalogs/participants-canonical.yaml — speaker whitelist (Vek + Eko + extensible)
• catalogs/common-typos-pt-br.yaml — pt-BR typo dictionary (golden seed: Nilson→Nelcael)
• catalogs/common-typos-en-us.yaml — en-US typo dictionary (initially empty)
• catalogs/phonetic-substitutions.yaml — known phonetic-class substitutions

Catalogs grow via PDCA cycles. Each cycle's growth is captured in pdca/cycle-<N>-<date>.md.

HUMAN_DOMAIN gates (per [C17] §2 in ~/.claude/CLAUDE.md)

Default = draft-only. All --mode inline invocations REQUIRE matching write-auth flag. Per [C07] v2.1.0 each invocation's auth is scope-limited (does not generalize to next invocation).

Quality Tests ([C17] §11 self-validity — 6/6 PASS dogfooded)

§0 BEING > Rules compliance

Sunset triggers ([C17] §6.5 DUED — qualitative, NOT counter-based)

Deprecation candidate when ANY:

• E1 — Mainstream ASR (Whisper / Loom / etc.) reaches near-zero phonetic-substitution error rate on rare proper names (empirical evidence in operator's own 30-day rolling sample)
• E2 — Operator explicit retraction
• E3 — Better composable pattern emerges in AAIF ecosystem
• E4 — ≥3 false-positive corrections observed in production (rule fires errantly)
• E5 — Catalog growth plateaus (no new entries for ≥60 days indicates either problem solved OR skill no longer being invoked)
• E6 — Confluence/GDrive/Notion natively integrate ASR-correction before publishing

Refs

• ~/.claude/rules/auto-self-harness.md [C17] — autonomy framework
• ~/.claude/rules/pr-review-protocol.md v2.0.0 — PR workflow used to land this skill
• multi-agent-os/skills/operator-quote-capture/SKILL.md — sister pattern (capture + filter + persist)
• multi-agent-os/skills/converge/SKILL.md — invoked for LOW-confidence correction adjudication in PDCA cycles
• multi-agent-os/skills/skill-writer/SKILL.md — used to validate this skill's frontmatter + structure
• Founding empirical case: Confluence page 303988750 (Daily Meeting 2026-05-27) — Nilson→Nelcael phonetic substitution
• pdca/cycle-0-2026-05-28.md — golden test report

Changelog