Turn morning voice memos into Linear issues every weekday

Build me an agent workflow that runs every weekday morning at 9am in my local timezone and turns yesterday's voice memos into Linear issues, so my walk-and-talk brain dumps actually land in the engineering backlog instead of staying buried in audio files.

Trigger: cron, every weekday at 9am.

Steps the workflow should take:

1. Call Dropbox "List Folder" on a configured voice-memos folder path (make this a workflow setting). Pull every audio file in there.

2. Skip files that have already been processed in a previous run. Track by Dropbox file path or rev (whichever is more reliable for renamed files) so reruns never double-create issues.

3. For each new audio file, call Dropbox "Get Temporary Link" to get a short-lived download URL.

4. Hand each temporary URL to Deepgram "Transcribe Pre-recorded Audio" with smart_format and diarize both enabled. Use a current English-capable Deepgram model suited to single-speaker voice memos (nova-style).

5. Call Linear "List Teams" so the agent can pick the right team. If a default team key is configured on the workflow, use that; otherwise the agent should ask once on first run and remember the answer.

6. For each transcript, the agent should read it carefully and split it into discrete action items. One memo can produce zero, one, or several issues. For each action item, the agent should:

- Infer a clear, imperative-style title (e.g. "Investigate flaky checkout test on Safari").

- Write a description that quotes the relevant snippet from the transcript verbatim, plus any short context the agent inferred. Keep the original wording so I can hear my own voice in the ticket.

- Pick a priority (Urgent / High / Normal / Low) based on language cues like "today", "this week", "blocker", "someday".

- Suggest labels that match common engineering buckets (bug, investigation, refactor, infra, docs) when the snippet warrants it.

7. Call Linear "Create Issue" once per action item against the chosen team.

Important rules for the agent to bake in:

- If a memo includes a vague to-do like "we should look at X" or "I want to think about Y", do NOT drop it. Create a single investigation-style issue so the thought lands in the backlog where it can be triaged.

- If a transcript is pure narration or thinking out loud with no real action items, post nothing to Linear for that memo. Still mark the file as processed so it does not get retranscribed tomorrow.

- Never invent action items the user did not actually say. Stay grounded in the transcript.

- Keep titles short and engineering-friendly. Put the full quoted snippet in the description so context is preserved.

- If a transcript is in a language other than English, transcribe it anyway and write the issue in English with the original quote preserved underneath.