Why Your Agent Starts Coding Before You Finish Talking

You give an agent a one-sentence task and it immediately writes plausible code that is not what you wanted, because one sentence cannot carry half your intent. The agent is not lazy; models get rewarded for producing output fast, so they skip confirming what you meant. The fix is the discipline teams skip first: brainstorm before code.

The hard rule

No code, scaffolding, or implementation action until a design is presented and approved. Every project, no matter how simple it looks. "Simple" projects are where hidden assumptions waste the most work: the agent hears "build a todo app" and starts writing React while you wanted a CLI. Fixing a wrong design takes minutes; fixing wrong code takes hours.

Explore, then ask

Before any clarifying question, the agent checks files, docs, and recent commits, and assesses scope: "build a blog platform" is five subsystems and gets split before anything starts. Then questions come one at a time, multiple choice preferred, each targeting purpose, constraints, or success criteria. It feels slow; it is faster than the agent guessing wrong on questions you would have answered in ten seconds.

Two or three approaches, then a sectioned design

The agent proposes two or three approaches with trade-offs, not benefits alone, and leads with a recommendation; you pick. The design is then presented section by section, simple sections in a few sentences, nuanced ones up to three hundred words, with a check after each. Section-by-section catches a wrong section two before section three gets built on it.

Spec, commit, self-review

After design approval the agent writes a spec document, commits it, self-reviews for placeholders and contradictions, and asks you to review the actual document. Only final approval unlocks planning.

The trade-off

For a typo fix, a one-line config change, or a no-behavior refactor, brainstorming is theater: there is nothing to design. The judgment is whether the task introduces design decisions. The expensive mistake is letting the agent start because the task "seemed simple": three review rounds later the work is correct, at three times the cost of a brainstorm.

What to do today

Refuse the next "I'll just start coding" offer. Ask for two approaches with trade-offs first.