I keep having the same conversation.

Someone tells me they’ve been going back and forth with an LLM for an hour. The output is wrong. It keeps misunderstanding what they want. They’re ready to write the whole thing off as overhyped.

So I ask what they started with. And it’s always some version of “I told it to build the thing.” No requirements. No constraints. No questions asked or answered. Just “go.”

That prompt leaves every important decision unresolved. The model has to invent the scope, constraints, priorities, and definition of success. It has no secret access to the requirements sitting in your head, so it guesses. Each guess creates another round of correction.

Planning by accident⌗

You discover the requirements through rejection: wrong audience, wrong format, missing constraint, useless output. Thirty messages later, the conversation contains a decent brief buried inside a pile of discarded drafts.

Take a content strategy. “Build me a content strategy” sounds like an instruction, yet the model still needs an audience, a goal, practical constraints, and some definition of success. It will fill those gaps on its own. Then you correct the audience, so it rewrites the recommendations. You add a constraint, so it changes the channels. You explain how success will be measured, so it rebuilds the priorities again.

You are writing the brief after the draft already exists. Every late requirement sends the work backward.

There’s a saying in the Navy SEALs: “Slow is smooth, and smooth is fast.” Deliberate setup cuts the backtracking that makes rushed work slow.

Make ambiguity visible⌗

Claude and Codex both offer planning workflows. A planning phase turns assumptions into questions before the tool starts producing files, formulas, or finished prose.

Start with the outcome. Define the inputs, constraints, edge cases, and success criteria. Let the LLM challenge the request and surface decisions you have left implicit. Push on the plan until both sides share the same definition of “done.”

For research, decide which questions need answers and what evidence would change your mind. For a spreadsheet model, define the inputs, outputs, and assumptions before writing formulas. For a content strategy, settle the audience and business limits before generating recommendations.

The model begins with a map built from your decisions. Most corrections stay at the edges because the core requirements are already settled.

Use ambiguity as the threshold⌗

Small, well-defined tasks can go straight to execution. “Add a date filter to this screen” already contains a clear action and a visible result. A five-minute task deserves a five-minute plan at most.

Ambiguous work needs the conversation first. A task you could explain to a junior developer in one sentence, with clear inputs and outputs, is ready to run. A task that needs a few minutes of questions belongs in plan mode for those few minutes.

Planning should match the uncertainty in the job. The hour-long session belongs to work with enough scope and ambiguity to waste five hours when the direction is wrong.

Buy back your attention⌗

Planning delays the first visible output so the agent can work longer without dragging you back into the chat.

With a clear plan and defined success criteria, the agent can work through more of the job without constant correction. You can get lunch, switch to something else, and return to output built against decisions you already made.

A vague brief keeps you in the loop. You watch every step, correct each wrong turn, and explain the same requirement again when the next output drifts. Your time gets consumed in small interventions that should have happened once, at the beginning.

Planning feels dull because the screen stays quiet while you define the job. The back-and-forth shrinks later. So does the urge to start over.

Slow is smooth. Smooth is fast. The hour spent defining a complex job can save five hours of correcting an LLM that started work before it understood the assignment.