Skip to content
← Insights
Nov 15, 20264 min read

Turn a voice note into a GitHub issue without losing its uncertainty

Preserve the transcript, resolve the repository explicitly and give each voice capture an identity that survives a failed request.

A voice recorder with an abstract waveform connected to an issue card, beside a review lens and a small locked credential box.

A voice note is useful when it reaches a place where you will act on it. For a developer working in GitHub, that place can be an issue. The automation needs to preserve the idea while making its destination obvious before publication.

I would build the capture path around a short review screen. Transcription and formatting can happen automatically. The repository and the final issue text should still be visible at the point where the note becomes shared work.

Preserve the transcript behind the summary

Begin with an audio recording or an existing transcript. If you use a cloud transcription service, the audio leaves the device. Select that path deliberately, especially when a note concerns a customer or a private project.

Keep the original transcript separate from the formatted summary. A model may convert "maybe investigate" into a confident task, or lose a qualification while shortening a sentence. The reviewer needs a way to compare the two.

Ask the formatting step for a small result: a title and an issue body. Optional metadata should remain optional when the source is unclear. An uncertain assignee is better shown for review than invented from a name in the recording.

Quoted source words should stay recognisable as source words. A proposed interpretation can be useful, but label it separately. Otherwise a future reader may mistake the formatter's suggestion for the speaker's original requirement.

Resolve the destination before sending

A spoken repository name can be misheard. Match it against an approved list, then show the resolved owner and repository. Do the same for labels and assignees.

GitHub's create-issue API documents the supported request fields and permissions. Build against those fields. A due date spoken in a recording needs a defined destination in your workflow, such as reviewed body text or an explicitly supported project field.

Relative dates need an interpretation too. Resolve "Friday" using the recording's date and intended timezone, then display the resulting date. A delayed upload should not silently reinterpret the note relative to the day the automation eventually runs.

Check whether the destination is public. A correctly transcribed note can still disclose private information when posted to the wrong repository. The review screen should make that visibility decision easy to notice.

The issue API also documents conditions under which optional assignments or labels can be ignored. Read the created issue back. A successful creation response is insufficient evidence that every optional field was applied as intended.

Treat similar titles as a hint

GitHub's issue-search API can help retrieve possible duplicates. A match on a few title words should open a comparison, not automatically discard the new note.

Two bugs can share a symptom and have different causes. One idea can also appear under very different wording. Show likely matches with enough context for the user to choose.

Search is also a weak retry guard. If the issue creation response is lost, the new issue may not appear immediately in search. Generating a new title and searching again can create duplicates.

Give each capture a stable local identifier before the first create attempt. Preserve the returned issue identity when available. After an ambiguous result, reconcile against the repository before issuing another create request.

A relay can keep the capture state through a mobile connection failure. It still needs a defined way to associate a returned issue with the same capture. Model-generated title similarity cannot provide that identity.

Keep credentials out of shared shortcuts

GitHub's token guidance recommends narrow permissions and describes GitHub Apps for long-lived integrations. Use the least access the capture path needs, with a revocation plan.

A secret embedded in a shortcut can be copied with the shortcut. Apple's sharing documentation makes distribution convenient; inspect every embedded value before distributing an automation.

Raw audio also needs a retention decision. Committing it to a repository creates another durable copy. Keep it only when the later workflow needs it, in a location with suitable access.

Test the path with a disposable repository. Include an unclear label and a failed connection after creation. Confirm that the review preserves uncertainty and the retry finds the existing issue. Measure the actual elapsed time on that setup; a promised twenty-second capture says little about a different network or transcription service.

AutomationGitHubProductivity