Workflow

Capturing my thoughts on the go

I needed a way to capture my thoughts when out and about. This is the overview of how I did it.

Black and white close-up of radiating palm leaves
Capture first. Structure later.

I built a voice-to-notes system that takes a spoken idea on my phone or Mac, usually mixed Danish and English, and turns it into searchable Obsidian notes Claude can query across the whole vault. Capture, cleanup, and views are up and running.

System overview

  • Capture: Voice MD records, sends audio to OpenAI for transcription, and appends the text to that day's inbox file.
  • Cleanup: on the first Claude session of the day, a script finds waiting inbox material; Claude splits each recording into its own note under Notes/, fixes transcription errors without changing my words, and turns every to-do into a checkbox, with the original transcript kept at the bottom.
  • Sorting: one flat folder; saved Obsidian views group by type (idea, task, reflection, question) and by tag; one page collects every open to-do.
  • Safety net: git backs up the vault before and after every cleanup.

Capture and cleanup

Capture runs through Voice MD because it already works on Mac and iPhone, stores the API key in Obsidian's secure storage, and retries failed uploads on its own. A Shortcut stays as backup. The plugin's text clean-up stays off, and because it has no save-path setting, daily-note mode is the inbox.

Cleanup belongs to Claude, not the transcription tool, because only something that can see the whole vault can link notes, assign a type, and pick from twelve fixed tags. The script skips today's inbox file so the phone can still be writing. Rules are strict: no translation, no new ideas, wording stays mine, unclear recordings go to Not important/, and the raw transcript always remains. Claude cannot invent tags; it may suggest one only when a theme shows up in three or more notes, and only after I approve it. Because the plugin's job list syncs through iCloud and can overwrite across devices, cleanup matches backups on date and text instead of trusting that list alone.

Boundaries that keep it stable

Structure comes from type, tags, and views, not from a growing folder tree. The vault syncs through iCloud, but git history lives elsewhere so sync cannot rewrite the backup trail. The OpenAI key never enters the vault or the project files. Ideas and to-dos for the system itself go only into Linear.

Why it holds

Capture stays boring and reliable. Cleanup stays vault-aware and word-preserving. Structure stays in views and a fixed tag list. Speak once, keep the original words, and let search do the rest.

Insight: Transcription should get the words in. Judgment about structure, links, and tags belongs to a model that can see the vault, with the human still owning the taxonomy.

This sits next to how I run agentic work elsewhere: small harnesses around tools I already trust.

obsidianclaudeworkflow
TJ

TJ

I design and write for the way people think, then build the digital thing that fits.

Writing index