YouTube Transcription Service

A lean Product Spec showing Coach's default section format.

Problem

YouTube researchers, students, and creators spend too much time scrubbing long videos to find the exact moment, quote, or citation they need. Video contains useful source material, but without a fast transcript search flow it behaves like locked-up evidence instead of readable text.

Hypothesis

If we give YouTube researchers an instant searchable transcript for any public YouTube video, they will use video sources more often in research, writing, and content creation because finding and copying a relevant moment becomes as easy as searching a text document.

Scope

  • In: public YouTube URL input, transcript retrieval, timestamped transcript display, transcript search, timestamp jump links, and formatted passage copy.
  • Out: private videos, bulk channel ingestion, speaker labeling, transcript editing, account libraries, collaborative notes, and paid export workflows.
  • Cut for this version: AI summaries and citation-format pickers, so the first release can prove whether fast search and copy alone solve the research pain.

User Experience

https://usecoach.ai/sample-mockup.html

Acceptance Criteria

Functional criteria

  • User can paste a public YouTube URL and receive a readable transcript with timestamps.
  • User can search within the transcript and jump to matching timestamped moments.
  • User can copy a transcript passage; the clipboard content is formatted as the quoted passage text, followed by the timestamp in [HH:MM:SS] notation, followed by the full source video URL on a new line.
  • If a search returns no results, the user sees an explicit "no matches found" message within the transcript panel; the search field and full transcript remain visible so the user can revise the query.
  • If the video has no available transcript or cannot be processed, the user sees a clear recoverable error and can try another URL.
  • The page preserves the pasted URL and any fetched transcript while the user searches, copies, or encounters a recoverable error; no fetched transcript disappears silently.

AI evals

  • Behavior contract: when AI transcription fallback is used, the transcript text must be grounded in the video audio and must not invent speakers, quotes, or timestamps.
  • Golden set: 50 public YouTube videos covering lectures, interviews, podcasts, noisy audio, long videos over 90 minutes, and videos with missing or weak captions.
  • Rubric: relevance of returned passages, grounding in transcript/audio, timestamp accuracy, copy format, and fallback behavior.
  • Ship threshold: at least 90% relevance pass rate, 100% grounding pass rate, at least 95% of timestamps opening within 5 seconds of the passage, and 0 critical failures on missing-transcript or private-video cases.
  • Failure handling: if transcript quality is too low, the video is private, or no transcript can be produced, the product shows a clear fallback instead of fabricating a passage.

Success Metrics

  • Leading indicator: at least 35% of first-time visitors who paste a valid YouTube URL copy one timestamped transcript passage in their first session.
  • Lagging confirmation: at least 20% of users who successfully generate a transcript return within 30 days and generate transcripts for three or more additional videos.
  • Check metric: transcript-related support reports for missing, wrong, or unusable output stay below 3% of successful transcript generations.