What I Stopped Doing #
Almost everything I now hand over to a generative model bypasses the keyboard. I say it out loud, and over the past eighteen months the share of my work done this way has grown without any deliberate decision on my part.
The biggest increase has been in project scoping, the stage when I have to describe what is needed, who it is for, and under what constraints. By "spec," I mean the document that turns a still-vague need into a workable scope.
I judged the change by the output, not by how it felt. This is not a controlled test, but the effect has repeated often enough to change how I work. When I tried to formulate the same request in writing, I tended to compress the context and got thinner responses. Speaking lets me include repetitions, second thoughts, and exceptions that only occur to me late.
If value comes from the context you provide, this should be good news. It took me a while to understand why it is not.
The Bottleneck Wasn't the Keyboard #
Writing a spec is slow, and the slowness was not a flaw in the medium.
To put something in writing, you have to decide what order it should appear in, what goes in and what stays out, and which condition outweighs another. Those are not steps you add to the thinking; they are the thinking. The keyboard did not impose them. I can type a disordered stream every bit as easily as I can speak one. The constraint was the document itself, which cannot be delivered until those decisions have been made.
Writing did not necessarily preserve the rationale. Two years later, I might still have forgotten why I ruled out an option. But it forced me to formulate that rationale at least once. Dictation does not necessarily erase it; it allows me not to formulate it at all.
A well-crafted client proposal is already half the work, because it defines what the product does, who it is for, and within what limits. That half exists in the hour someone spends writing it, not only in the file that reaches the client.
When I speak, that hour does not disappear. It simply stops being mine. I hand over the material in the order it comes to me, repetitions, jumps and all, and the system turns it into a deliverable document.
Who Writes the Record #
Every formal record involves two people.
One tells the story. The other decides what goes on the record. The first speaks for twenty minutes; the second produces two pages. Months later, when someone picks the matter up again, it returns in the form the second person gave it. The story came from the first. The document belongs to the second.
In my case, the second is not a person, or even a single model. It is the pipeline that takes the dictation, transcribes it, interprets it, applies instructions and context, and then generates the document. The point is unchanged. The story is mine; the form that survives is produced elsewhere.
On August 24, 2026, Thibault Sottiaux, who leads OpenAI's core products, including ChatGPT and Codex, described a similar habit in an interview: he often dictates into his phone, including first thing in the morning, because speaking a request is more efficient for him than typing it.
That same conversation returns to a broader product principle: technology should adapt to the person, not the other way around.
It is an impeccable design principle. It also tells us, without saying so, who holds the pen.
The Cleaned-Up Version #
There is a way to read these eighteen months that overturns the entire argument, and it needs to be taken seriously.
Perhaps writing was better at compressing context than at surfacing it. A written spec is a cleaned-up reconstruction after the fact, one that makes the problem look more orderly than it was in the mind of the person grappling with it. Speech carries the actual sequence in which ideas occurred to you, including the "no, wait, not like that" that gets erased from the page before anyone reads it.
Speech is not the original thought either. Even when speaking, I select, compress, and forget. But the filter is different and, in my case, less restrictive.
What I call a loss of work may actually be the loss of a professional fiction. My experience points in exactly that direction: more gets through when I speak.
The File I Keep Beside It #
The concession has to be accepted in full. The point still stands, and quality is not the issue.
The structure is produced either way. Someone, or something, decides what matters, what to discard, and how to order the constraints. The better output shows that the work has been done. It is good evidence that it has been done well, not a guarantee that no important constraint was lost.
For the past few months, I have adopted a countermeasure. Alongside the final document, I keep the raw transcript of the dictation. It is not the original thought, and it does not necessarily preserve tone, hesitation, or emphasis. But it is the closest record I have of the starting point.
You could argue that this is enough: put the two files side by side, and the gap becomes visible.
Try it. You can see that a constraint has disappeared; you cannot see whether it was dropped because it was irrelevant, because it conflicted with another, or because it was not understood. The comparison shows what transformations occurred. By itself, it does not reveal the rationale behind them.
This is the same failure mode found in corporate archives: they preserve the delivered version, not the judgment behind the version that was discarded. Without that judgment, the material preserves the outcome but teaches far less about why, whether to a person or a machine.
Of course, I do not always approve the first version. I edit it, revisit decisions, ask for alternatives, and restore elements I recognize as missing. But I am working on a structure that has already been proposed. The most dangerous omissions are precisely those I cannot challenge, because they no longer appear in the text I am reviewing.
Then there is the signature. When I approve the final document, I perform the act that closes every formal record: read, approved, and signed. The signature validates the outcome, and the approval process can still correct it, but it does not by itself reconstruct the selection that produced it. I attest that the text represents me today, not necessarily that tomorrow I will be able to explain every exclusion it contains.
The Third File #
Writing did not preserve the rationale, but it forced me to formulate it. That is the difference a before-and-after comparison cannot show.
We need a third artifact: a transformation log. For each element in the source material, it should state what was retained, what was merged, what was discarded, which conflicts were resolved, and which assumptions were introduced. It should distinguish between what I confirmed and what the system inferred.
An explanation generated after the fact is not enough. It can sound plausible without corresponding to the process that produced the document. The log should not pretend to expose the model's internal reasoning. It should document explicit transformations, be created alongside the document, remain linked to the relevant source passages, and undergo the same review as the final text.
It would not make the system fully transparent. But it would make the selection visible, contestable, and reusable. It would turn a signature into a review of the rationale.
It Worked Better #
For months, I argued that context is an asset, and that companies make a mistake when they fail to notice where it is accumulating. I fell into the same trap, with one aggravating factor that does not apply to companies in quite the same way.
The aggravating factor is that it worked. When a method delivers better results, it is hard to give it up in order to preserve full traceability of your own judgment. I did not go back, and I do not think you will either.
The problem, then, is not that a model writes the spec. It is that the most convenient workflow produces a finished document without, by default, leaving a verifiable record of the decisions that generated it.
The question is not whether you will stop talking to machines. It is whether, two years from now, when someone asks how you arrived at a decision, you will have not only a before and an after, but also the rationale that connects them.