5,410 dictations in the last few months, 1,616 of them straight into a terminal. This is not a post about a tool. It is a guide to the rail you build after it.
What is here:
- Three real examples of a spoken sentence turning into work
- Ten minutes of settings inside Wispr Flow, MCP included
- The rail in three connections, and a one-week task list
- Four open-source Skills, two lines to install
Part 1: what it sounds like in practice
The rule I learned fast: one spoken sentence should be one complete task. Not "do something", but what, for whom, and what comes out.
Example 1: code
Open a PR with this description: we added a test for the event arriving twice, run the tests and attach the result.
The agent opens a branch, writes the test, runs it, and returns a link.
Example 2: a client
Summarize the last call with the client into three points, draft a follow-up message in Hebrew, and do not send it.
The output: a draft in the client folder. I read, fix a sentence, send.
Example 3: knowledge
What did we settle in that meeting about the timeline?
The answer arrives with a source: which meeting, which sentence. No recording to open.
The three parts that keep repeating: a clear verb, a clear source, a clear output. That is the whole difference between dictation that stays text and dictation that becomes work.
Part 2: ten minutes of settings inside Wispr Flow
The shortcut
The default is push-to-talk: Fn on Mac, Ctrl+Win on Windows. A quick double tap locks hands-free mode, no holding. Change it under General, then Shortcuts.
The personal dictionary
It learns on its own, but do not wait for it: add product names, internal terms, and the English words that keep showing up in your Hebrew. This is also what feeds speaker names in the Notetaker.
Snippets
A signature, a booking link, a fixed opening line. Say the trigger word, get the paragraph, formatting, links and lists included.
Style per app
WhatsApp gets short sentences, email gets a formal tone, the terminal gets exact text with no over-cleaning. One voice, three tones, set once.
Command Mode
Not only dictate, also edit: select text, hold Fn+Ctrl, say "shorten to three lines and drop the opener", release. The text is edited in place. Enabled under Settings, Experimental; needs a paid plan. Double tap locks, Esc cancels.
Notetaker
Install, connect a calendar (so every meeting knows who is in it and the names in the summary are right by themselves), grant microphone and system audio. Works in in-person meetings through the laptop mic too. "What did I miss?" returns the last few minutes in the chat without leaving the meeting.
MCP: your meetings, readable by any agent
This is the part most people miss.
- On Mac: Wispr Flow settings, then MCP (after Connectors).
- Claude or ChatGPT: one click on Connect, approve in the browser, done.
- Any other tool (Cursor, VS Code, Claude Code): the URL under All other apps, as a new MCP server.
- What the agent sees: meeting summaries and transcripts, notes, tasks, calendar. Read-only, the agent deletes and edits nothing.
From that moment, "what did we settle about the timeline?" becomes a question the agent answers with a source. Notetaker beta, Mac only for now.
Plans
Free up to 2,000 words a week, Pro unlimited. The free tier would not have covered my 5,410. No Wispr Flow yet? Signing up through this link gives you a free month of Pro (my referral link): https://wisprflow.ai/r?YONATAN1
Part 3: the rail, in a version anyone can build
Three connections. Each works alone; together they are the system.
1. Voice to work
Wispr Flow dictates into any text field, the terminal included. People who work with a coding agent (Claude Code, Cursor, Codex) get the most out of it: the long prompts are exactly where fingers get lazy and the mouth does not.
2. Conversation to knowledge
The Notetaker summarizes meetings with real names, and there is a "what did I miss?" button for whoever joins late. At our place the summaries flow automatically into the client folder and into a knowledge base you can ask. The simple version: summaries into a shared folder, then into the tool you already work in. The advanced version: a small script that reads the summaries and feeds a knowledge base.
3. One rule for numbers
A quote, a price or a commitment reaches a document only after a look at the full transcript.
Not because the summary is bad, but because a summary is interpretation and a transcript is fact. With real names in the summary, that check takes seconds.
Part 4: the task list, one week
- Day 1: install and the shortcut. Ten minutes of dictation instead of typing, only for the long prompts. Do not try to switch everything in a day.
- Day 2: personal dictionary. The internal terms, product names and the English words that recur in your Hebrew. This is what turns recognition from "good" to "mine".
- Day 3: meetings. Turn on meeting summaries, and make sure every client meeting sits on a calendar invite. That is how the names in the summary come out right by themselves.
- Day 4: the connection. One place where summaries land, and make it the place where you actually work. Not a folder that opens once a month.
- Day 5: the rule. Write it somewhere visible: "numbers move only with a transcript." Then work by it.
Week 2, for whoever wants more: dictate straight to the agent. Every dictation that ends triggers a task. At our place this runs through an automatic connection between Wispr and our server, which recognizes the intent and hands it to whoever executes.
Part 5: the Skills kit
A Skill is a fixed instruction with one input and one output. These are the four that run most around calls at our place, and they are open source now: https://github.com/yonatangross/rail-kit
- post-call. Input: a transcript. Output: a short follow-up message in the client's language, and a one-page scope. Nothing is sent on its own.
- client-context. Input: a client name. Output: everything known, from every call, on one page, conflicts between sources included.
- prep-call. Input: a client name and a date. Output: a seven-section cheatsheet in the right register.
- sync-call-state. Input: an approved call summary. Output: a proposed record update that waits for a yes before it is written.
In Claude Code it is two lines:
/plugin marketplace add yonatangross/rail-kit
/plugin install rail-kit@yonyonIn any other agent, copy the folder. All of it, with the PDF guide and the 30-second check, at https://yonyon.ai/rail with no form.
Closing
The tool is Wispr Flow. The system is the rail. The four Skills, the one-week list, the MCP one-pager and the 30-second voice-rail check wait on one page, no form: https://yonyon.ai/rail
And if you want the rail running on its own for you, summaries that land in a knowledge base by themselves, dictation that triggers an agent, Skills tuned to your business, that is what I help businesses build at Yonyon AI. A 20-minute call is enough to know if it fits. The link is at the bottom of that page.