Thirty-nine cents bought a go-to-market plan
and none of the judgment it needs

I keep the raw output of agent runs in git, because a plan is easy to admire and hard to re-check once the transcript is gone. The run I keep coming back to is dated 2026-08-18. An agent named nlt-gtm-operator, running on Sonnet, took a structured brief about our first client and returned a complete go-to-market playbook. 18 of 18 sections, one invocation, $0.39, self-reported confidence 0.82. The raw invocation JSON is committed beside the document at docs/af-runs/gtm-operator-5af375bb.json, so every number here can be re-read from the file rather than taken from me.

What eighteen of eighteen actually means

Sections, not paragraphs. Positioning, ideal customer profile, buyer persona, channels, the offer, validation hypotheses, a ninety-day sequence, kill criteria. That's the document a founder puts together over several days of interviews and drafting, and the hard part has never been the writing. It's not leaving a hole. There were no holes. Each section came back carrying specifics traceable to the brief instead of filler, and the through-line held from the summary down to the kill criteria. I want to be straight about that before I complicate it: the output was good.

The money is the least interesting number

Thirty-nine cents is the figure people react to, and it matters least. A price that low doesn't mean the work got cheap. It means the first complete draft stopped being something you schedule. Before a run like this, a small client's plan waits on a week you can clear. After it, the plan is on the table and your job is to argue with it. Arguing with a draft takes real hours. Better work than staring at an empty outline, but not free, and nobody should price it as if it were.

0.82 is the agent's opinion of itself

It's tempting to read that score as a quality measure. It isn't one. It's the agent's estimate of its own output, produced by the same run, with nothing standing between the two. No part of that number knows whether the pricing in section seven matches what the shop charges, or whether the owner is willing to say any of it out loud, or whether a channel the plan recommends is one a very small team can staff next month. A number a system assigns to itself is a self-report. Useful as a relative signal once you have a hundred runs to compare. Worth nothing as evidence about this one.

I've written before about checks that report success without exercising the thing they cover, and I don't want the two confused. This is a different animal. There's no check here at all. The work is genuinely done. The score next to it just isn't a statement about it.

What the draft can't settle

The edits that mattered weren't stylistic. A generated plan will happily name a customer segment the owner has no appetite for serving. It'll quote a number the shop hasn't agreed to honor. It'll propose an approach that runs into a rule the agent had no way to know about, because the rule lives in a conversation nobody wrote down. None of those are writing defects. They're decisions, and they belong to whoever eats the consequence of getting them wrong. That's the client, and behind them, me.

Drafts in the owner's voice, not the owner's words

The document says that about itself, in its own header, on the line that records the cost and the confidence. It was in the template before any of this ran, and it's the most useful sentence in the file, because it sets a reader's expectation at the top. What follows is a starting position, and an operator reviews it before any of it reaches a customer. The plan is still in review on the client's side, so nothing in it is public and nothing here identifies them. The argument survives the anonymity fine. Thirty-nine cents and one run replaced the drafting. They replaced none of the reading, and the reading is where a plan turns into something a business can act on.

That's the shape of the work on my side, where the draft lands in minutes and the real time goes to the calls only the owner can make, which is most of how I can help.

Sources: the engagement's execution playbook, docs/GTM-EXECUTION.md, whose header line records the agent, the model, the invocation id, the date 2026-08-18, the $0.39 cost, the 0.82 confidence and the 18/18 section count. The unedited run output is committed beside it at docs/af-runs/gtm-operator-5af375bb.json. The client is in pre-approval review, so they aren't named here and nothing from the plan is quoted.