AddThisFeature

Voice Transcription

Turn recorded or uploaded speech into text people can edit inside the app.

involved AI Content

What it adds

A transcription pipeline that takes recorded or uploaded audio and returns editable, timestamped text attached to a record.

What your agent is told to do

5
  1. 1

    Identify where audio already reaches the app — attachments, uploads, call or meeting recordings — and add in-app recording only where a microphone genuinely fits the task.

  2. 2

    Reuse the app's existing background-job system for the transcription itself. A request that holds a web connection open for the duration of the audio will time out on anything longer than a short clip.

  3. 3

    Return the result as a draft the user reviews and corrects before it is saved to the record. Do not write transcribed text straight into a customer-visible field.

  4. 4

    Put ceilings on duration, file size, and transcribed minutes per account per period, and state the limit before the user records or uploads rather than after the file has been sent.

  5. 5

    Do not retain the source audio by default. Decide a retention period explicitly, delete temporary files on failure as well as success, and make sure a deleted recording takes its intermediate copies with it.

Edge cases it handles

8
  • A browser can deny microphone permission outright, or the user can revoke it mid-recording. Explain what was blocked and offer file upload as the alternative rather than leaving a dead record button.
  • An upload interrupted by a dropped connection must be resumable or must fail cleanly with the partial data discarded, never leaving a half-length recording that looks complete.
  • Long audio has to be split to be processed, and the pieces must be reassembled in the original order with the boundaries stitched so a word cut in half is not duplicated or lost.
  • Where the provider reports low confidence, mark those words visibly and keep per-segment timestamps so the reviewer can jump to the audio and check rather than guessing.
  • Do not attribute lines to named speakers unless speaker separation is genuinely available and reliable. A confidently wrong attribution in a meeting transcript is worse than an undifferentiated block of text.
  • Temporary audio, chunk files, and provider-side copies must all be removed when retention ends. Deleting only the row in the database leaves the recording sitting in object storage.
  • Silence, music, background noise, or a language the configured capability does not support must produce a clear explanation, not an empty transcript presented as a finished result.
  • When the transcription capability is unavailable, keep the recording and queue the job. The user's audio must survive the outage even though the text does not arrive immediately.

Definition of done

9
  • Transcription runs as a background job and never blocks or times out a web request.
  • The transcript is presented as an editable draft and is only saved to the record once a person accepts it.
  • Audio longer than a single processing window is transcribed in order with no lost or repeated words at the joins.
  • Low-confidence words are marked and every segment carries a timestamp that seeks the audio.
  • Duration, size, and per-account usage limits are enforced and shown before the upload starts.
  • Source and intermediate audio are deleted at the end of the stated retention period, including from external storage.
  • A denied microphone permission, an interrupted upload, and an unavailable transcription capability each produce a distinct, actionable message.
  • The feature matches the existing design system.
  • No existing functionality is broken.

Related features

How it works

  1. 1

    Copy the link

    Grab the Markdown instruction URL for this feature.

  2. 2

    Give it to your AI

    Paste it into Claude Code, Cursor, v0, Lovable — whatever you build with.

  3. 3

    It inspects, then implements

    Your agent reads your existing app first, then adds the feature to fit it.

Works with your stack

These instructions are written to adapt. They tell the agent to detect your framework, match your existing design system, and reuse what you already have — rather than assuming a particular stack.

Need it tighter than that? Customize the feature and tell it exactly what you're running.