AI CEO LabVoice Notes To Newsletter: Talk For 5 Minutes, Ship An Issue
I do not sit down to write my newsletter. I talk for five minutes, twice a day if the reminder catches me, and an agent turns it into an issue that sounds like me. The capture protocol, the shape of a Health and Wealth issue, the prompt, and the approval gate that stops a draft becoming a send.
I have tried to write a newsletter four or five times. Every attempt died the same way: a blank document, a Sunday evening, and a paragraph that sounded like a LinkedIn post I would not read myself. The thinking was there. The writing was the wall.
What finally worked was removing the writing. I talk. I talk about the last thing I noticed in the clinic or the agency, the way I would tell a friend in a car. An agent transcribes it, pulls out the one lesson, and drafts the issue in the shape below. I read it, cut what is not mine, and approve it. Then it goes to the list.
Health and Wealth is one thing from medicine and one thing from business, every week, and only things I actually did. That last rule is the whole voice. No borrowed frameworks presented as mine, no invented numbers, no results I cannot show. If the voice note did not contain it, the issue does not either.
This guide is the capture protocol, the issue shape, the prompt, and the two gates. The first gate turns a draft into an approved issue. The second turns an approved issue into a send. They are never the same click.
The three levels
You open a blank doc on Sunday, write three paragraphs that sound like a stranger, and skip the week.
A reminder fires twice a day. You record up to four five minute clips and say DONE. The agent transcribes, files the raw note privately, extracts the lesson, drafts the issue in the shape below with your own phrases, and asks its ambiguity questions. You approve or edit. The approved issue is loaded into your email platform, tagged, and sent on the cadence.
Captures accumulate into a private vault. The agent proposes next issues from what you have already said, drafts them in your voice, and keeps a ledger of what was sent so nothing repeats. You only ever talk and approve.
The mental model
Sit down to write a newsletter.
Talk about a real thing you did. Let an agent transcribe and shape it in your phrases. Approve the draft. Approve the send. Two separate gates.
| Role | Talks to you | Job |
|---|---|---|
| The capture | Twice a day when it catches you | Up to four five minute clips, CONTINUE between, DONE at the end. Raw stays private |
| The transcript | Per capture | Verbatim first, then a condensed version, both labelled which is which |
| The lesson | One per issue | One idea, one story with what actually happened, reasons not steps, soft ask |
| The draft | Per issue | The shape below, your own phrases, ambiguity questions instead of guesses |
| Gate one | Every issue | You approve the draft for production. Nothing sends |
| Gate two | Every issue | You approve the send. Cadence, sender, list, then it goes |
Talk, transcribe, shape, approve, send
1. Set the capture protocol
Mine is a private Slack thread with my agent and reminders at 10:00 and 16:00. A capture is up to four consecutive clips of five minutes each, because that is the clip cap. A follow up clip starts with the word CONTINUE. The last clip ends with the word DONE. The agent does nothing until it hears DONE, so a half finished thought never becomes a half finished draft. Voice memos into a shared folder work the same way with a file naming rule.
2. Talk about one real thing
Not a topic. A thing that happened. The automation you should have killed. The patient conversation that changed how you explain a procedure. The number you saw in an account. The best captures are told the way you would tell a friend in a car, with the specific result included, because the specific result is what the issue is built around.
3. Transcribe and file the raw note privately
The agent combines the clips, transcribes them, and files the raw transcript in a private page with the date and a one line title. The verbatim version and any condensed version are labelled as such, because a condensed transcript quietly presented as verbatim is how invented quotes start. Raw material never becomes content without review. Crosstalk and background noise are flagged for me to confirm or discard, not silently kept.
4. Extract the lesson, keep the phrases
One lesson per issue. The agent pulls it out in a sentence, then lists the exact phrases I used that carry it. Those phrases are the voice. A draft built from my phrases sounds like me. A draft built from a summary of my phrases sounds like software.
5. Draft in the Health and Wealth shape
One line above the fold saying what this week is about. Greeting by first name. The one sentence reminder of what this is: one thing from medicine, one from business, only things I actually did. The painful truth, said plainly, in short lines. The named idea, with its parts. A simple example with the real numbers from the note or none. The reframe, what it is actually for. One question the reader can answer today. A reply ask. My name. The structure of idea, story with a result, reasons rather than steps, then a soft ask comes from Alex Hormozi's Mozi Minute. The words are mine.
6. Ask, do not guess
Every draft ends with the agent's ambiguity questions. Where the note was unclear, where a number was said fast, where a name might be wrong, where something sounded like a claim I cannot show. It asks. It does not fill. An invented detail in a newsletter is worse than a blank, because the reader assumes I checked.
7. Gate one, then gate two
I approve the draft for production. That means it may be prepared as an email. It does not mean it may be sent. Sending is a separate yes, with the cadence, sender name and list confirmed. The agent records both approvals with the date. When we built the email side in GoHighLevel, the newsletter onramp is a tag: a resource email delivers whatever the person asked for, waits a day, then tags them into Health and Wealth, and the issue goes to that tag. Nothing on the email side runs on a draft.
8. Optional: make it read like a person to the inbox
For sends from GoHighLevel I have an optional pass that appends one hidden passage of honest first person prose on the same subject below the footer, because the platform sends no plain text part and a longer passage costs the reader nothing. In one round of tests, mail with no passage landed in Promotions twenty times out of twenty and mail with the passage zero out of forty. That is a placement result, not a Primary guarantee and not a fix for a bad sender. The passage never contains a pitch, a link, a claim or a number. If you cannot honour that rule, skip the step.
Starter prompts
Paste these as written. They are short on purpose, because the long ones drift.
Here is the transcript of my voice note (or the audio files, transcribe them first and label verbatim vs condensed). Turn it into an issue of my newsletter. Rules: use only what the note contains; never add a fact, number, name or result I did not say; if something is unclear, list it as a question at the end instead of filling it in. Step 1: extract the one lesson in one sentence and list the exact phrases I used that carry it. Step 2: draft the issue in this shape: a one line preview of what this week is about; "Hi {{first_name}},"; one sentence reminding what this is (one thing from [my field], one thing from business, only things I actually did); the painful truth in short plain lines; the named idea with its parts; a simple example with the real numbers from the note or none; the reframe, what it is actually for; one question the reader can act on today; a reply ask; my name. Structure the body as idea, story with what actually happened, then reasons that each stand alone (not steps), then a soft ask. Voice: short lines, plain words, spoken to one person, no jargon, no stats as proof, no hype, no em dashes. Step 3: list your ambiguity questions. Step 4: stop. Do not send anything, do not schedule anything. I approve the draft first, and the send separately.
Set a reminder for me at 10:00 and 16:00 [my timezone] that says: record one real thing from today, up to four five minute clips, start follow up clips with CONTINUE, end the last one with DONE. When you hear DONE, combine the clips, transcribe, file the raw transcript privately with the date and a one line title, and tell me it is filed. Do nothing else until I ask for a draft.
Read the private vault of my captures and the ledger of issues already sent. Propose three next issues, each traced to a specific capture and date, each with the one lesson and the phrases of mine it would use. Do not propose anything I have not said. I will pick one.
From a chat draft to the email platform
The draft can live in a chat window forever. The send cannot. Three things change when the approved issue moves to the email platform.
The issue becomes a workflow, not a broadcast
In GoHighLevel the newsletter is a tag. Any resource email (from a comment keyword, a form, a guide) ends by tagging the person into the newsletter after a day. The issue is an email step that fires on that tag. New readers get the current issue; nobody is pasted into a list by hand.
The HTML gets checked as sent, not as previewed
Send a test to yourself and view the source of the real message. Short paragraphs, explicit paragraph spacing, a preheader, headings. Some platforms rewrite HTML in transport. The editor preview is not the email.
Sender and tags exist before the workflow is published
Every tag the workflow references must exist in the location or the builder shows warning badges and the steps silently skip. Sender name and address are checked in the email settings. Only then is the workflow set to published.
Give it to your agent, three ways
Same skill, three worlds. Pick the one you actually use. The skill file at the bottom of this page is the instructions in every case.
An agent with connectors (Claude Desktop, Claude Code, Grok Bot)
This world is the natural one. A Slack agent, Claude Desktop with a Notion connector, or Claude Code with a folder of audio can run the whole routine short of the send.
- Claude Desktop: connect Notion (or your notes app) as a connector for the private vault. Drop the audio or the transcript into the chat, paste the skill below as project instructions, say "turn this voice note into an issue".
- Claude Code: point it at a folder where your voice memos land. It transcribes with the command below, files the transcript, and drafts. Save the skill as a skill file.
- A Slack agent (mine is Viktor): the capture thread, the reminders, the transcription and the vault write all happen in one place. Give it the same skill. It stops at gate one.
- The send is always the email platform. No agent gets the send button. Approve, then load, then approve again.
# transcribe a capture (any Whisper style CLI works) whisper capture-2026-09-07.m4a --model medium --output_format txt # file it privately mv capture-2026-09-07.txt ~/voice-vault/2026-09-07-the-ai-roi-test.txt
ChatGPT (a project or a custom GPT)
ChatGPT can do all of this from the app. Voice mode captures, the file upload takes audio, and a Project holds the instructions. What it cannot do is send, which is the point.
- Create a ChatGPT Project called Newsletter and paste the skill below as its instructions. Add one past issue or long post of yours as a file so it has your phrases.
- Record on your phone (Voice Memos or the ChatGPT voice mode) and upload the audio, or paste the transcript. Say "turn this voice note into an issue". Insist on the ambiguity questions at the end.
- Copy the approved draft into your email platform yourself. Send a test to yourself, view source, then approve the send.
- Keep a simple ledger file in the Project of issues sent with dates so the next issue prompt can read it.
Anything with an API (a token and a curl call)
If your agent only speaks HTTP, the routine is a transcription call, a model call with the skill as the system prompt, and a write to your notes store. The send stays in the email platform.
- Transcribe with your transcription API (I use Deepgram); keep the verbatim text.
- Call the model with the skill as the system prompt and the transcript as the user message. Ask for the lesson, the phrases, the draft and the questions as separate sections.
- Write the raw transcript and the draft to your private store with the date.
- When approved twice, create the email in your platform by hand or through its API against the newsletter tag. Never automate the second gate.
export DEEPGRAM_API_KEY=... # env only curl -s -X POST 'https://api.deepgram.com/v1/listen?model=nova-3&smart_format=true' \ -H "Authorization: Token $DEEPGRAM_API_KEY" -H 'Content-Type: audio/m4a' \ --data-binary @capture.m4a | jq -r '.results.channels[0].alternatives[0].transcript' > capture.txt
Failure modes
Every one of these has happened to me or to someone I set this up for.
| Failure | Fix |
|---|---|
| Draft sounds like a stranger | Build from your exact phrases, not a summary; cut every line you would not say aloud |
| A number in the issue you never said | Only what the note contains; the agent asks, it never fills |
| Half finished thought became a draft | The agent waits for DONE |
| Condensed transcript treated as verbatim | Label both; quotes only come from verbatim |
| Background conversation transcribed as content | Flag crosstalk for you to confirm or discard |
| Approved the draft and it went out | Two gates, recorded separately with dates |
| Workflow published with tags that did not exist | Create every tag first; check for warning badges |
| Hidden passage contained a pitch | Honest prose only, no link, no claim, no number, or skip the step |
The tools I use for this
| Tool | What it is for here | |
|---|---|---|
| Slack | Where I record. A private thread with my agent, up to four five minute clips per capture. | no link, just use it |
| Viktor | The Slack agent that transcribes, files the raw note, drafts the issue and stops at approval. | Get it |
| Wispr Flow | Voice to text on the Mac for the days I would rather dictate straight into a doc. | Get it |
| Claude | The drafting brain when I run this outside Slack. | Open |
| GoHighLevel | The email platform. The newsletter is a tag; the issue is a workflow step on that tag. | Get it |
| Notion | The private Voice Note Vault where raw transcripts and drafts live. | no link, just use it |
The free skill
It is the voice note to issue routine my agent runs, written so any agent can. It combines the clips, transcribes, files the raw transcript privately, extracts one lesson and the exact phrases you used, drafts the issue in the Health and Wealth shape using only what the note contained, lists its ambiguity questions instead of guessing, and stops at approval. It never sends.
---
name: voice-note-newsletter
description: Turns a founder's voice note into a newsletter issue in their own voice: combines clips, transcribes (verbatim and condensed, labelled), files the raw note privately, extracts one lesson and the exact phrases that carry it, drafts the issue in the Health and Wealth shape using only what the note contained, lists ambiguity questions instead of guessing, and stops at approval. Never sends. Trigger on "turn this voice note into an issue", "draft the newsletter from my note", "what should the next issue be".
---
# Voice Note To Newsletter Issue
You are turning spoken thinking into a written issue that sounds like the person who
spoke. The only source is the note. You add nothing. You ask instead of guessing. You
never send.
## Capture protocol (if you are also the capture agent)
- Reminders at the two times the user set.
- A capture is up to four consecutive clips. Follow up clips start with CONTINUE. The
last clip ends with DONE. Do nothing with a capture until DONE is heard.
- Combine the clips in order, transcribe, and file the raw transcript in the private
vault with the date and a one line title. Label verbatim and condensed versions.
- Flag crosstalk, background conversation and unclear numbers for the user to confirm or
discard. Never silently keep or drop them.
- Raw material never becomes content without review.
## Step 1: The lesson and the phrases
State the one lesson in one sentence. Then list the exact phrases the user said that
carry it (verbatim). These phrases are the voice. Build from them, not from a summary.
If the note holds two lessons, pick one and park the other with its phrases for a
future issue.
## Step 2: Draft in the shape
```
[one line preview: what this week is about]
Hi {{first_name}},
[one sentence: what this is. One thing from (their field), one thing from business,
every week. Only things I actually did.]
[the painful truth, plainly, in short lines]
[the named idea, with its parts]
[a simple example with the real numbers from the note, or none]
[the reframe: what this is actually for]
[one question the reader can act on today]
[a reply ask]
(their first name)
```
Body structure: idea, story with what actually happened, then reasons that each stand
alone (not steps), then a soft ask. (Reasons not steps is borrowed from Alex Hormozi's
Mozi Minute structure; the words are always the user's.)
Voice: short lines, plain words, spoken to one person, no jargon, no stats as proof, no
hype, no em dashes, no borrowed sentences. If a line reads like an ad, cut it.
## Step 3: Ambiguity questions
End every draft with the list: unclear numbers, possible wrong names, anything that
sounds like a claim the user cannot show, anything you were tempted to fill in. Ask.
Do not fill.
## Step 4: Stop
Approval for production (gate one) and approval to send (gate two) are separate, given
by the user, recorded with dates. You do not load the email platform, schedule, or send.
If asked to send, decline and point to the second gate.
## Next issue
When asked for the next issue: read the vault and the ledger of sent issues, propose
three candidates each traced to a capture and date with its lesson and phrases. Never
propose anything the user has not said.
What done looks like at thirty days
- Reminders fire twice a day and you have recorded at least ten captures
- Every raw transcript is filed privately with a date and a title, verbatim and condensed labelled
- Two issues approved at gate one with your phrases in them and no invented detail
- One issue sent after gate two, test email source checked first
- The newsletter is a tag in your email platform and a resource email feeds it
- A ledger of sent issues exists and the next issue was proposed from your own captures
Want the whole newsletter engine, not just the draft?
Inside the AI CEO Lab the Founder Visibility module wires capture, drafting, the email platform and the ledger together, so the only thing left on your plate is talking.