How I Turn a Messy Voice Note Into Finished Work

I have somewhere north of four hundred voice notes on my phone. Most were recorded walking, some in a car, a few at 1am. They are the single best source of ideas I have and, for about two years, they were also the single biggest waste in my workflow — because a four-minute ramble is not a thing you can use. It's raw material that requires an hour of transcription and rewriting to become a thing you can use, and I never had the hour.

So they piled up. I'd record the idea, feel the relief of having captured it, and never open it again.

What broke the pattern wasn't better note-taking. It was realizing that the thing standing between the voice note and the finished work is a translation job — messy spoken thought into structured written thought — and translation is exactly what a model is good at. The problem is that if you just paste a transcript in and ask for "a clean version," you get something clean and dead. The rambling gets removed, and the one sentence where you accidentally said the true thing gets removed with it, because it was ungrammatical and buried in the middle of a tangent.

Everything below is aimed at that specific failure: getting the structure without losing the signal.

First pass: don't clean it, map it

The mistake I made for months was asking for a rewrite on the first pass. Rewriting is a compression, and compressing a transcript you haven't understood yet throws away whatever you didn't already know was in there.

Now the first prompt never produces prose. It produces a map.

Below is a raw transcript of me thinking out loud. It is unstructured, repetitive, and I change my mind partway through. Do not rewrite it and do not summarize it. Instead, give me: 1. Every distinct idea in here, as a separate line — even the ones I only touched for a sentence and moved on from. Include the half-formed ones. 2. For each, the exact phrase I used when I said it (quoted, verbatim, including the messy phrasing). 3. Where I contradicted myself or changed direction — what I said first, what I said later. 4. What I clearly cared about most, judged by what I circled back to, not by what I spent the most time on. Transcript: [PASTE]

The verbatim quote requirement does the same job here that it does when I'm reading a contract: it stops the model from tidying my phrasing into generic phrasing before I've had a chance to see what I actually said. Half the time my clumsy original sentence *is* the good line — it's specific in a way the polished version isn't.

Point 3 is the one that surprised me. When I talk through a problem I contradict myself constantly, and I never notice while I'm doing it. Having the contradictions listed back is usually where the real work is: I said the project needed X, and then four minutes later described a plan that quietly assumed not-X. That's not noise in the recording, that's the actual unresolved question, and it's invisible in any summary.

Point 4 exists because time-spent is a terrible proxy for importance in speech. I'll spend ninety seconds justifying something I already decided and eight seconds on the thing that's actually bothering me — but I'll come back to the eight-second thing three times. Return frequency is the better signal, so I ask for that explicitly.

Second pass: pick the shape before you fill it

Only after I've read the map do I know what the voice note was — a feature spec, a piece of writing, a decision I need to make, or three unrelated things that happened to be in my head at the same time. That determines the shape, and shape has to be chosen deliberately, because a model given a transcript and no target will default to "article with headings" every single time.

Here is a structured map of a voice note I recorded, plus the original transcript. I want to turn this into: [A ONE-PAGE SPEC / A DECISION WRITE-UP / AN OUTLINE FOR A POST / A TASK LIST]. Rules: - Use my own words and phrasings wherever they work. Do not upgrade my vocabulary. - Where I was vague, leave it marked as [VAGUE — needs a decision] rather than resolving it for me. You do not know what I meant. - Do not add ideas that are not in the source. If the structure has an obvious hole, name the hole; don't fill it. - Keep the contradictions visible. Put them at the end under "Unresolved." Map: [PASTE] Transcript: [PASTE]

"Do not upgrade my vocabulary" earns its place in that list. Left alone, a model will turn "this feels sticky" into "this presents notable friction," and something real dies in the swap. I said sticky because sticky is what it felt like. The whole reason I record voice notes instead of typing is that speech gets closer to what I actually think — laundering it into business English gives back exactly the thing I was trying to avoid.

The [VAGUE] marker is the other half. A model handed a half-formed thought will complete it, confidently and plausibly, and the completion reads so smoothly that a week later I can't tell which parts were mine. Forcing the gaps to stay visible keeps the document honest about what I've actually decided versus what got decided for me by autocomplete.

The pass I run when the note is three ideas pretending to be one

Some recordings are one thought. Plenty aren't — I start on the shot list, drift into pricing, and end up somewhere near hiring. Trying to force those into a single document produces a bad document.

This transcript contains more than one project or concern tangled together. Separate them. For each distinct thread: give it a name, list the lines that belong to it, and say what the next concrete action would be — or "not actionable yet, needs a decision about X." If two threads are actually the same underlying problem wearing different clothes, say so and explain why. Transcript: [PASTE]

That last instruction is there because it keeps catching me. Twice now I've had "we need better onboarding copy" and "nobody uses the second feature" come back as one problem, which they obviously were, and which I had been treating as two separate items on two separate lists for weeks.

What I stopped trying to automate

I don't record voice notes *for* the model. The value of talking it through is that talking it through is how I think — if I started performing into the microphone for a cleaner transcript, I'd lose the whole point and get worse ideas rendered more legibly.

I also don't let it produce the final version of anything with my name on it. It gets me from a four-minute ramble to a structured draft with the vague parts flagged, which is maybe eighty percent of the distance and one hundred percent of the tedious part. The last stretch — deciding the vague things, cutting what turned out to be nothing, writing the sentences that carry weight — is the part where the work happens, and handing it over produces something competent and anonymous.

The honest accounting: this doesn't make my ideas better. It makes the ones I already had survive contact with the following Tuesday.

Why these live in a library and not in my head

Every one of these prompts is saved, and that's not an incidental detail — it's the reason the system works at all. The moment I have a transcript in front of me is the moment I want to be done, not the moment I want to carefully compose an extraction prompt with verbatim-quote rules and vagueness markers. Left to instinct I'd type "clean this up," get the polished dead version, and quietly go back to letting the recordings pile up.

That's what I built Super Prompts for, and it's the use case I still hit most: the prompts I already know are better than what I'd write under mild time pressure, one click away, so the pressure never gets a vote.

If you've got a folder of voice notes you've never opened, try this on one of them. Map it before you clean it. You'll find at least one thing you forgot you knew.