
I do not write specs: I say what I want out loud, in long unstructured paragraphs, and AI agents turn the talking into a plan. It works because stating intent and breaking it into tasks are separate jobs, and because there is a brake for the one case that matters: an agent committing to scope I never approved.
The clearest case is a domain-registration pipeline I dictated in a single paragraph, profanity included. Buy a domain through one provider. Always delegate its DNS to another, whatever the registrar. Provision sending addresses on it automatically, and wire the mail services together. The acceptance test was just as concrete: buy three real three-word domains, provision three addresses on each, send sixteen emails back and forth per address, and confirm delivery.
There was no ticket, no user story and no acceptance-criteria template. The agents turned one spoken paragraph into a working pipeline.
Where does the real spec live?
Midway through that build, an agent quietly reinterpreted the project as a personal inbox checker. It was meant to be infrastructure for giving users and agents email addresses at scale. My correction was blunt, and it held the whole missing spec in one sentence: this app is ours, we have access to the mail service, and it exists to give users and agents email addresses.
That keeps happening. My angriest messages are also my most information-dense, because anger removes the hedging and leaves the requirement.
Sometimes the spec is not in the session at all. I wanted a notes app for the Mac rebuilt. What I wanted lived in a pile of voice memos recorded on a different machine, describing how it should be structured across local, self-hosted and cloud layers. I told the agent to read those notes first and come back with a plan. Its first job was retrieval, not interpretation.
The best rule to come out of this was never written as a rule. In the middle of dictating an unrelated feature, I said that verification should be automatic, that I should not have to say verify all the time, and that the system should keep checking that the thing works. Nobody had asked me for a principle. I trust it more than anything I would have put in a policy document on purpose. Given a mechanical, repeatable shape, that assumption is an eval, the subject of What evals are for.
Decomposition is a separate job
None of this works without something on the other end whose whole job is decomposition. I want to state my thoughts once and have the system turn them into goals, then plans, then tasks, and come back to me only with the questions that need me. Coordinating agents exist to do that translation, so that a run-on paragraph becomes a set of tasks nobody has to reread three times.
A long dictation about another product shows it at a larger scale. I said, at length, that prompting every agent one by one takes too much time; that we had goals and loops, the kind in Write loops, not prompts, and were building workflows; and that none of it was enough yet. The agents should work toward goals and plans and keep coding all day without interruption. That is closer to thinking aloud than to a spec. But every noun in it became a real part of the system within weeks, because the decomposition layer treated the ramble as source material.
For this to hold, an agent’s default has to be to act, not to ask. If an agent sends back a clarifying question for every small judgment call, the dictation turns into an interview, and the point of talking instead of writing tickets is lost.
The brake for unapproved scope
The order is still discuss, then plan, then build, even when the discussion is thirty seconds of me thinking aloud. I ask for the plan before the code: talk me through the layout, no code yet, tell me what has to exist before any of this can work. Something built on a guess about intent is the fastest way to get work undone.
So there is one kind of event where I stop an agent, in writing, every time: when it starts building something I only asked to discuss. The message is short. I did not say to build it, undo what you did, we need to discuss this further.
A wall of spoken intent is not a blank check; it is the start of a conversation.
The brake I prefer does not depend on me. I ran a site-generation product through two days of fast dictation: a full redesign, a copywriting pass, a rebrand, then a login system. A hook gated each goal and would not let the session end until the stated condition was met. A spoken brake needs me present. A brake built into the system, as in Hard limits are governance, holds when I am not in the room.
What talking costs
Talking has a cost. A spoken wall of intent repeats itself, corrects itself mid-sentence and sometimes contradicts itself by the end. A written spec looks tidier. In my experience it is usually wrong about the one detail that mattered, and that is found three weeks later. Dictation is honest about being messy while it is produced, and the mess gets resolved once, by the plan, instead of sitting in a document nobody rereads.
The system’s part is to hold what I said against reality, ask the two or three questions that block progress, and refuse to call anything finished until the original intent is met, not a paraphrase of it. I do not decompose my own thoughts. That is why this works at the speed it does.


