Pattern: Plan First, Then Code

Have the AI write a plan before it starts coding — it prevents expensive wrong turns.

The problem with "just do it"

If you tell an agent "build this" directly, it starts acting immediately — and often makes silent assumptions that turn out wrong. You usually only notice once something's already half-built the wrong way.

The pattern

Ask the AI for a plan before any code gets written:

Describe the goal and requirements
Ask the AI to produce a step-by-step plan (no code yet)
Read the plan, correct wrong assumptions, answer open questions
Only then: "Now implement the plan"

Why this works better

A plan is cheap to read and correct — code is more expensive to read and correct. Catching a mistake in the plan ("you forgot we already have auth middleware") takes a minute. Catching the same mistake in finished code takes much longer.

When it matters most

Multi-step or unclear tasks
When multiple files or systems are involved
When mistakes would be costly (production code, database migrations)

For small, unambiguous changes ("change this color"), an explicit planning step is usually overkill.