Reading every line of code is not rigor.
Sometimes it means you do not know where the risk lives.
The same failure shows up in decision-making. Challenge every assumption. Generate five alternatives. Run the counterexample exercise. Keep reopening the search space until the process looks suitably intelligent.
Visible effort makes a convincing costume for judgment.
Any useful model creates leverage by leaving things out. Expertise compounds that leverage by teaching us which possibilities we can usually ignore.
Every model excludes something that could matter. Reopening everything to recover it would erase the advantage.
A genuinely different framing can expose what the first one missed, but that move is harder than it sounds. Most “outside the box” thinking stays inside the same box with different vocabulary.
Asking AI for five alternatives often produces five variations of the same underlying assumptions. Each one still needs attention, evaluation, and a decision.
Doubt consumes attention, and attention is finite.
Doubt is a currency.
Spend it across every routine choice and you have less available when the context changes, the evidence stops fitting, or the cost of being wrong starts compounding.
Some decisions justify carrying uncertainty for longer. A hard-to-reverse choice with a large blast radius may deserve deeper investigation. A familiar, reversible one usually does not.
Say you are designing a system and already have a preferred pattern in mind. You can declare the implementation and ask the agent to execute it, or leave a small door open.
Move from instructional closure to provisional framing:
“I think this should be event-driven, or?”
Name the direction without turning it into a constraint.
That “or?” does not demand a ritualized search for alternatives. It gives the agent permission to challenge the direction when the constraints suggest something meaningfully better.
The useful skill is knowing where scrutiny has leverage.
Spend doubt where being wrong compounds.
Engineering · Writing