A voice note recorded without any structure comes back as a wall of text. Technically searchable, practically unreadable. The fix is not a syntax to memorize. It is saying out loud the same structure you already have in your head.
This post covers two things: how to speak that structure, and what the resulting file actually looks like in your folder or vault. The second half matters more than it sounds, because the file is the note. There is no database copy.
Speak the labels
Transcription is literal. Whatever you say lands in the text, punctuation inferred from your cadence. So say the labels.
"Title. Project Alpha ideas."
"List. Three priorities." "Fix the signup flow." "Speed up search." "Add dark mode."
"Header. Risks we need to address."
"Vendor API might hit rate limits." "Team bandwidth is tight this quarter."
"Summary. Focus on signup and search first. Dark mode can wait until Q2."
That gives you a transcript whose sections are already named. You can scan it, and if you later run the optional rewrite step, the labels are unambiguous instructions for it to work from.
Three cues carry almost all of the value:
- "Title" followed by what the note is about. This one earns its keep twice, because the title becomes the filename.
- "List" before you start listing.
- "Header" when you change topic.
What lands on disk
A note dictated into a Markdown folder or an Obsidian vault becomes one file:
Project Alpha ideas.md
Open it in any text editor and this is the whole file. No frontmatter, no metadata block, no invisible header:
Title. Project Alpha ideas.
List. Three priorities. Fix the signup flow. Speed up search. Add dark mode.
Header. Risks we need to address. Vendor API might hit rate limits. Team bandwidth is tight this quarter.
Summary. Focus on signup and search first. Dark mode can wait until Q2.
Two things to notice.
The filename is the title, not a timestamp. Brain Dump derives the file's name from the note's title and sanitizes it for the filesystem: / \ : * ? | become -, " becomes ', < and > become parentheses, whitespace collapses, leading dots are stripped. A collision appends 2 before the .md. A note with no title at all gets a placeholder name you can change. Rename the note and the file is renamed with it; edit the body and it is not.
There is no frontmatter. This surprises people who expect a voice app to stamp its own metadata into every note. Writing created_at and updated_at is a setting, and it ships off. With it off, those keys are not written. If a note already has frontmatter you wrote, or a plugin wrote, it is preserved verbatim, keys and order intact. A title: key is updated only if it is already there; the writer never adds one.
Turning the transcript into Markdown
The transcript above is structured English, not Markdown syntax. Converting it is a separate, optional step that runs after capture, not during it. It ships off, and capture never waits on it.
Run it and you get the shape the labels described:
# Project Alpha ideas
- Fix the signup flow
- Speed up search
- Add dark mode
## Risks we need to address
- Vendor API might hit rate limits
- Team bandwidth is tight this quarter
Summary: Focus on signup and search first. Dark mode can wait until Q2.
Keeping the two apart is the point. Capture stays fast and local. Cleanup happens when you ask for it. See privacy and on-device AI for what runs where.
You can also just edit it. The editor handles headings, lists, tasks, callouts, highlights, wikilinks and Mermaid blocks, and it writes them back as plain Markdown.
Mistakes worth skipping
Saying punctuation. "Comma" and "period" end up in the transcript as the words "comma" and "period." Pause instead, and let the transcription engine infer it.
Leaving notes untitled. An untitled note gets a placeholder filename, and a folder of placeholders is a folder you have to open one file at a time. Say a title first. It costs three seconds.
Correcting yourself mid-sentence. "No wait, scratch that" becomes part of the text. Finish the thought and fix it later; editing text is cheap, untangling a self-interrupting transcript is not.
Structure that other tools can read
Because the output is a plain .md file with real Markdown, the structure is not just for you.
- [ ]tasks are parsed as tasks, and can sync to Reminders and Calendar.[[Wikilinks]]are typed with autocomplete over your vault's titles, filename stems, aliases and paths, and[[Note#Heading]]resolves for real.- Frontmatter
tags:andaliases:are read, written, and editable in the properties panel. Inline#tagsin the body are not indexed, so put tags in frontmatter if you want to find things by them. - Attachments are embedded in Obsidian's wikilink form,
![[folder/image.png]], not.
One extension to know about: task reminders use an @date autocomplete that is Brain Dump's own, not standard Obsidian task syntax. It is readable text in either app, but it is not a shared convention.
FAQ
Do I have to say punctuation marks? No. Brief pauses are enough.
Can I create deeper heading levels? Say "Subheader," or edit the file afterwards. The editor supports the full heading range.
What about checkboxes?
Type - [ ] in the editor, or convert a list after capture. Spoken lists come out as lists, not as checkboxes.
What if I mess up mid-sentence? Keep going. Edit the file after.
Where does the file live?
Wherever you pointed the app: an iCloud Drive folder, an on-device folder, or an Obsidian vault. Inside a vault, Brain Dump also adds one hidden .braindump directory for its own sidecars and audio, and touches nothing else of yours. See what Brain Dump reads and writes in a vault.
Related: Custom dictionary for technical terms to improve accuracy on jargon and names, and meeting notes template for a structured approach to work notes.