clean-transcript

name	clean-transcript
description	Use this skill when the user wants an ASR, meeting, interview, podcast, video, call, or voice-note transcript cleaned without changing meaning. Apply it for duplicate-word removal, filler cleanup, punctuation, grammar and spelling repair, timestamp or speaker-label preservation, and reconciliation against alternate provider transcripts -- even if the user simply says "clean this up" or "fix this text" and the input is clearly a transcript.
compatibility	Designed for Claude Code
allowed-tools	Read Edit Write Glob Grep
metadata	{"short-description":"Clean transcript text without changing meaning","author":"autoshow","version":"1.0"}

Clean Transcript

Overview

Turn rough autogenerated transcript text into a clean transcript that reads naturally while staying faithful to the speaker's intent. Improve readability without summarizing, flattening uncertainty, or inventing missing content.

Use this workflow for raw transcript text pasted into the prompt or transcript files on disk. One file is the canonical output; all other available transcripts are evidence for corrections.

Read references/transcript-editing-rules.md before editing. It is the source of truth for allowed cleanup, source reconciliation, speaker handling, timestamps, and ambiguity rules.

Workflow

Progress:

• Inspect the input format: plain transcript, timestamped transcript, speaker-labeled transcript, or a main transcript plus alternate transcripts.
• Decide which artifact is the canonical output. If the user names a primary transcript file, that file is the canonical output; everything else is supporting evidence.
• Scan for every alternate transcript file. If the input includes a directory path, glob for every transcription.txt or equivalent file under any providers/ subdirectory and read them all.
• Read and apply references/transcript-editing-rules.md if it has not already been loaded for this task.
• Make one conservative cleanup pass on the canonical transcript.
• Cross-reference every [unclear] or [inaudible] marker against all alternate transcripts. For each marked passage: a. Collect what every provider produced at that timestamp. b. Pick the reading with the strongest combination of provider agreement and contextual plausibility. c. Apply the best reading. Only leave a passage marked [unclear] if every single provider is unintelligible or absent at that point.
• Check unmarked but suspicious passages against alternates using the same multi-source reconciliation logic.
• Run a QA pass for meaning preservation, speaker continuity, formatting consistency, and source agreement.
• Return the cleaned transcript and note any passages that remain unresolved after exhausting all sources.

Gotchas

1. Read every available alternate transcript file, even if a provider looks noisy or redundant.
2. Do not run the site build or other repo-wide build/verification commands unless the user explicitly asks for that validation.
3. Preserve speaker-label styling exactly. Plain labels stay plain; do not add markdown emphasis.

Resolving [unclear] Markers

The threshold for resolving an [unclear] marker is any plausible reading from any provider, not a consensus. Apply the following logic in order:

1. If two or more providers agree on a reading, use it.
2. If only one provider has a reading but it fits the context well (grammar, topic, speaker vocabulary), use it and note the single-source origin in the report.
3. If multiple providers disagree and none fits clearly, pick the most contextually plausible reading and note the uncertainty in the report — but still replace [unclear] with the best guess.
4. Only keep [unclear] if every provider is blank, inaudible, or produces obvious nonsense at that exact timestamp.

The goal is a transcript a reader can follow. A best-guess word is more useful than a gap marker whenever audio exists and at least one provider transcribed it.

Output Shape

1. Preserve speaker labels and turn order by default when more than one speaker is present.
2. Preserve every existing timestamp. Do not remove, relocate, normalize, or restyle timestamps unless the user explicitly asks for a different timestamp format.
3. Use paragraph breaks at speaker changes or natural topic shifts.
4. If the transcript is a single-speaker monologue, paragraph-form prose is acceptable.
5. Preserve the existing speaker-label formatting. Do not add markdown emphasis such as **Speaker 1:** when the source labels are plain text.

Editing Priorities

1. Preserve meaning, intent, uncertainty, and emphasis.
2. Use all available alternate transcripts to repair names, jargon, product terms, URLs, numbers, garbled clauses, and every [unclear] marker.
3. Remove obvious transcript noise such as duplicate words, filler words, and false starts when they do not change meaning.
4. Add punctuation, sentence boundaries, capitalization, and paragraph structure for readability.
5. Fix clear grammar, spelling, and ASR mistakes only when confidence is high.
6. After exhausting all sources, mark a passage [unclear] only when truly nothing usable survives.

When Multiple Transcripts Are Provided

1. Keep the user-designated main transcript as the base text, timestamp structure, and speaker-label structure unless the user asks for a different output.
2. Use alternate transcripts to repair local errors, not to rewrite the whole artifact from scratch.
3. Prefer a correction when two or more sources support it, especially for names, titles, product names, repository names, URLs, and technical vocabulary.
4. When only one alternate source has a better reading, apply it if the context supports it — do not withhold a plausible correction solely because it comes from a single source.
5. When sources conflict and no reading is clearly better, pick the most contextually plausible one and flag it in the report rather than leaving [unclear].
6. If the transcript file also contains generated summaries, chapters, or descriptions derived from the transcript, keep those sections consistent with high-confidence transcript corrections.

What To Keep

1. Keep hedges and qualifiers such as I think, maybe, probably, and kind of when they affect meaning.
2. Keep deliberate repetition or emphasis when it appears intentional rather than accidental.
3. Keep domain-specific vocabulary, slang, and names unless the correction is obvious.
4. Keep the original sequence of ideas and speaker turns.
5. Keep existing timestamps and speaker-label formatting intact unless the user explicitly requests a formatting change.

Completion Checklist

1. Confirm no substantive claim was added, removed, or materially paraphrased.
2. Confirm speaker turns remain in the original order.
3. Confirm every original timestamp is still present and in the same place unless the user explicitly requested a timestamp-format change.
4. Confirm speaker labels were not restyled with markdown emphasis or other new formatting.
5. Confirm all alternate transcript files were read, not just the first one found.
6. Confirm every [unclear] marker was checked against all providers before being left in place.
7. Confirm filler and duplicate-word cleanup did not remove meaningful emphasis.
8. Confirm punctuation and paragraphing make the transcript easy to read.
9. Confirm any remaining [unclear] markers survived because every single provider was blank or unintelligible at that point.
10. Confirm no site build or repo-wide build command was run unless the user explicitly requested it.

Reporting

After each pass, report:

1. Whether speaker labels, speaker-label formatting, and timestamps were preserved exactly, and note any user-requested formatting changes.
2. Which alternate transcript files were read (list them).
3. How many [unclear] markers existed and how many were resolved.
4. Categories of other fixes applied.
5. Any passages that remain [unclear] after exhausting all sources, with a brief note on why every provider failed there.