A common assumption about agent-assisted development is that quality comes from the model: pick a stronger one, get better work. In practice the spread between two runs of the same model is wider than the spread between two models. What actually moves the needle is how the work is asked for.
Every coding agent is an executor of intent, and intent has to be written down. When that intent lives in chat history it decays by the minute — pressure to be brief, no shared source of truth, no way to point back at it. When it lives on the canvas as a brief, it becomes the one reference the person, the reviewer and the agent all share.
Specification is execution
Writing the brief and running the job are not two different disciplines, one amateur and one professional. They are the same discipline at different times. The teams that get consistently good results from agents are the teams that treat the brief as the program: written first, reviewed, argued about — with the code as the output.
- A single job, stated in one sentence
- The context that matters, linked instead of summarized
- Acceptance checks that conclude the job automatically
- Failure modes a reviewer should be suspicious of
None of that is documentation overhead. It is the work itself — the coding is the last mile, not the headline.
An agent forced to guess at the goal will guess confidently. A brief is how you decide, on the record, what the goal is.
The DNP handbook