Tickets were written for people who knew what you meant

“As a user, I want to export my timesheet, so that I can use it in Excel.”

For a team of people, that is a good ticket. The engineer who picks it up knows which timesheet, has opinions about CSV versus XLSX, and will ask in the stand-up whether managers should export their team’s hours. The story was never the whole specification. It was a reminder of a conversation that would happen anyway.

Give the same ticket to a coding agent and the conversation doesn’t happen. The agent reads what is written, and the ticket says very little. So it fills the gaps with the most plausible answers: every report, as XLSX, for everyone. It writes tests for its own guesses and opens a clean pull request. Nothing about it looks wrong until finance opens the file.

The failure isn’t the code

Agents are good at code. What they are bad at is knowing what you meant when you didn’t say it. A human builder notices a gap and asks; an agent notices a gap and decides. The more capable the agent, the more convincing the decision.

That changes what a ticket is for. It is no longer a reminder for someone who shares your context. It is the context.

What a ticket has to say now

A ticket an agent can build from says four things that a reminder for a colleague can leave out:

1. An intent. Who wants this, and why. When the ticket doesn’t cover a decision, the intent decides.
2. Context. The repository, the file, the screen, what exists already. “Context is a link, not a memory.”
3. Acceptance checks. Observable statements, not “should be fast”. Each one is ticked later with the evidence that proves it.
4. No open questions. Every unknown is either answered or written down as a question, and the ticket isn’t handed over while one is open.

We wrote that down as the Brief Standard, an open Markdown format anyone can use, and built NEWWORK Projects around it.

Ask, don’t guess

The most important part is the fourth. AI can write a convincing brief from three words of notes, and that is exactly the problem: a convincing brief full of invented numbers is worse than a vague one, because nobody questions it.

So when Write with AI fills something in that your notes don’t say, it lists it as an assumption you keep or strike. When it can’t know something (a name, a limit, the exact wording of a button) it asks. The ticket shows its readiness, and it isn’t Ready for agent until someone answers.

We don’t guess. We ask. It turns out to be the whole product.

Try it on one of yours

Paste a real ticket into the Brief Checker. The score is computed in your browser by the same rules as the app, and the rewrite shows the brief it could have been, with the questions it would have asked you.