Nigel Thorne

Resources / Working with AI

Give coding agents direction and feedback.

Fast code generation is useful. The harder job is keeping the product, architecture and evidence connected.

Design the loop around the code.

Direct

Outcome, boundaries, examples

Build

One reviewable change

Check

Evidence against the outcome

↶ Feed failures back into the next attempt. Stop when evidence meets the acceptance criteria.

Forward pressure

Make a useful next step clear before the agent acts. A scaffold, example or constraint reduces what it has to guess.

Back pressure

Test what happened. A failure must affect the next decision, otherwise it is just output in a log.

Too little direction produces plausible work on the wrong problem. Weak checks let mistakes accumulate. An endless review loop can waste as much time as no review at all.

Start where your workflow breaks.

Keep a person responsible for the outcome.

Separate discussion from permission

Exploring an idea should not trigger file changes or a deployment. State when the agent may research, edit, run checks or change an external system.

Keep changes reviewable

Agree the outcome and acceptance examples. Review the actual diff and behaviour. Use human domain expertise when the agent cannot validate an assumption.

Start with the smallest checks that catch likely defects. Add stronger checks for security, data changes or hard-to-reverse actions. Another model's agreement is not independent evidence that the product works.

Source and perspective

These guides draw on Beyond Coding's Lead Architect Reveals His Rules for Coding Agents. They adapt ideas from the conversation into practical examples, rather than treating one engineer's workflow as a universal rule.

The interview describes large delegated runs as well as the need for smaller reviewable changes. Here, the recommended scope is whatever a person can meaningfully inspect and verify. A model's confidence score is not a calibrated guarantee.