Verification Loops

Let it see the result

Agents do much better when they can run something that tells them if the work is right: tests, a linter, a type checker, a screenshot. Tell the agent the verification command in CLAUDE.md, and ask it to run it before saying it is done.

Build, measure, improve

You cannot improve what you cannot measure. Evals, checks and budgets turn a demo into a dependable system.

A verification script

One command that fails loudly gives the agent a clear signal. set -e stops at the first failing step.

#!/usr/bin/env bash
set -e
npm run lint
npm run typecheck
npm test

Do not trust "done"

An agent can report success without checking. Read the test output yourself for important changes, and be suspicious of tests that were weakened to make them pass.

Quick check: What most improves an agent's reliability on a coding task?

• A longer prompt with more adjectives
• A command it can run to verify its work
• Hiding the test results
• Fewer files in the repo