Deterministic, auditable transcript-to-tutorial workflow for Zoom class captions with optional resource enrichment. Use when asked to 'process this transcript', 'convert lecture to notes', 'run transcript pipeline', 'generate class tutorial from Zoom captions', 'validate transcript coverage', or 'enrich class resources' (Notion/Canva/Drive links) for bootcamp notes. Combines deterministic scripts (ingest, validate, publish) with chat-stage prompts (refine, synthesize, enhance) and keeps traceability in .pipeline/* artifacts.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
transcript-pipeline
description
Deterministic, auditable transcript-to-tutorial workflow for Zoom class captions with optional resource enrichment. Use when asked to 'process this transcript', 'convert lecture to notes', 'run transcript pipeline', 'generate class tutorial from Zoom captions', 'validate transcript coverage', or 'enrich class resources' (Notion/Canva/Drive links) for bootcamp notes. Combines deterministic scripts (ingest, validate, publish) with chat-stage prompts (refine, synthesize, enhance) and keeps traceability in .pipeline/* artifacts.
Transcript Pipeline Skill
Audience: Users converting raw class captions into publish-ready study notes.
Goal: Run a deterministic, auditable transcript-to-tutorial workflow with optional resource enrichment.
Purpose
Use this skill to convert raw class captions into high-quality study notes while preserving accountability through ledger + validation artifacts.
Use scripts for deterministic work. Use chat/stage prompts for language-heavy transformation.
Keep resource extraction status explicit (success/fallback/blocked).
Anti-Patterns
NEVER let Claude self-certify coverage — always run validate_coverage.py. Self-certification drifts 5–10% per stage and compounds; by Stage 3 the reported coverage is meaningfully wrong.
NEVER skip segment-ledger creation (Stage 0). The ledger is the only source of truth for what was in the raw input vs what survived synthesis — without it, you cannot prove no-orphan-claims later.
NEVER merge chunks (merge_chunks.py) without first aligning their topic inventories. Unaligned merges create duplicate topics, drop unique ones at the seam, and silently degrade coverage.
NEVER override --deep-pass failures by hand-editing final_notes.md. If deep-pass flags a missing section, re-run the stage that should have produced it — hand-edits detach the output from the audit trail.
NEVER run the Colab-notebook pipeline on a non-ML session. The prompts in run_colab_notebook_pipeline.py are tuned for code commentary and produce bizarre output on generic lecture content.
NEVER publish (Stage 5) before validation (Stage 4) has returned PASS. The publish step sanitizes for learner view, which masks validation failures you would have caught before distribution.
NEVER treat exceptions.json as "interesting notes" — it's a fail-report. If it exists after validation, the pipeline failed; re-run from the earliest referenced stage.
NEVER modify .pipeline/segment_ledger.jsonl by hand, even to fix a typo. The ledger is hashed downstream; any edit invalidates every later artifact.
