TL;DR workflow

Explore the problem. Then build the solution.

A two-stage mental model for using AI to write better software with more clarity and less thrashing.

Read full article
The mental model

Two different jobs, two different modes.

Stage 1Explore

Reason about the space, compare approaches, ask questions, and find the shape of the problem.

Stage 2Build

Give a coding agent a tighter brief and let it focus on implementation, validation, and iteration.

Stage one

Use conversation for breadth before you ask for code.

1

The goal is not a patch yet. It is a clearer understanding of the problem.

1
Map the problemDescribe the situation and surface the constraints.
2
Compare optionsTest possible strategies and expose their trade-offs.
3
Choose a directionTurn the exploration into a decision you can hand off.
Stage two

Once the direction is clear, switch from ambiguity to execution.

Give the coding agent the conclusion, the relevant context, and the constraints. Its job is now to build—not to decide what the problem should have been.

The handoff

A good handoff carries the thinking forward.

ContextWhat is happening

The symptoms, affected files, users, and constraints.

DecisionWhat should change

The chosen approach and why it is preferable to the alternatives.

ProofHow to validate it

The checks, exports, tests, or evidence that will show whether it worked.

A concrete example

Canonical URLs: understand the shape before touching the site.

1

One canonical URL pattern becomes the source of truth.

1
ExploreReason about duplicate paths and the canonical URL strategy.
2
BuildNormalise sitemap entries and canonical tags in the post-render flow.
3
ValidateCompare the implementation with the Search Console evidence and refine it.
Final takeaway

Get clarity before you ask for code.

One model can explore the space; another can build once the direction is clear. Separating the phases makes the work more disciplined and easier to improve.

Read the full article