Speaking is faster than typing; reading speech is slower than reading writing
Dictation is the quickest way to get an idea out of your head and the worst way to store it. A raw transcript preserves every restart, every aside, every "actually, no." Whoever reads it later does the editing that was avoided at the start.
Cleaning and structuring look like one task but behave like two. A model asked to do both at once often drops content. It decides a tangent is noise, removes it, and reorganizes around the gap. Sometimes the tangent was the idea.
Running them as separate passes fixes this. The first removes disfluency without changing meaning or order. The second imposes structure on text that is already clean and can show what is actually there.
Keeping the raw transcript attached to the finished document costs nothing and settles every later argument about whether something was said.
How the ActionFlow runs on the canvas
The audio file arrives through Document Upload. Transcribe Audio converts it to text, and the model service you choose here matters more than anywhere else in the flow. A transcription provider tuned for your language and audio conditions changes everything downstream.
The first Generate Text node removes disfluency only. Its instruction forbids reordering, summarizing, or cutting content. The output should express the same thought clearly.
The second Generate Text node structures the cleaned text into headings, a lead paragraph, and an order. Because it works on clean input, it does not guess what a false start was.
A Notion node writes the finished document, with the raw transcript appended underneath as a collapsed section. Google Docs fits the same slot.
When recordings vary in length, an If node is worth adding. Short notes go straight through. Long ones route through an Extract Data node first so the structuring pass does not handle a forty-minute transcript in one call.
Both generation calls report their token usage and cost as node outputs, which is how you can see that transcription, not generation, is the expensive part of this flow.
Nodes this ActionFlow uses
- Document Upload (Data): takes the audio file.
- Transcribe Audio (AI Core): converts speech to text. Provider choice matters most at this node.
- Generate Text (AI Core): first pass, removes disfluency without changing meaning or order.
- Generate Text (AI Core): second pass, imposes headings, ordering, and a lead paragraph.
- Extract Data (Data): narrows very long transcripts before the structuring pass.
- If (Control): routes short and long recordings differently.
- Notion (Integration): receives the document with the raw transcript appended. Google Docs fits the same slot.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.