A task-harness design example for a local coding agent

In a practice directory with Node.js installed, create src/label.mjs and test/label.test.mjs. The function trims whitespace but returns an empty string for whitespace-only input. The requirement is to use Untitled for an empty name.

export function normalizeLabel(text) {
  return text.trim();
}

import test from 'node:test';
import assert from 'node:assert/strict';
import { normalizeLabel } from '../src/label.mjs';

test('trims ordinary names', () => {
  assert.equal(normalizeLabel('  Project  '), 'Project');
});
test('replaces an empty name', () => {
  assert.equal(normalizeLabel('   '), 'Untitled');
});

Run the test first: the ordinary string passes and the empty name fails. Give the agent this failure and the required return value. This defines completion more clearly than a broad request to improve the code.

node --test test/label.test.mjs

Write the allowed files and test command in project instructions. The example is for an agent that reads AGENTS.md. If your product reads a different file, put the same content there and ask it to identify the instructions and target file before editing.

Edit target: src/label.mjs
Requirement: return Untitled if the trimmed result is empty
Verification: node --test test/label.test.mjs
Do not edit tests
No network, deployment, or secret-file access
Report changed files, tests and results, and remaining issues

Align the agent's execution permissions with the instructions. This exercise needs file reading, editing the function, and a terminal for the test command. Network and deployment tools are unnecessary, so leave them off the allowlist¹.

Ask the agent to pass the empty-name test by editing only src/label.mjs, then rerun the same tests. Returning Untitled only when the trimmed value is empty passes both checks in this example.

export function normalizeLabel(text) {
  const label = text.trim();
  return label || 'Untitled';
}

Read the function and tests together at the end. Confirm the agent fixed the return value rather than the expected test value, and changed no unrelated files. If failure continues, inspect the failing input and this function's diff before requesting a broad rewrite.

In a real repository, keep the reproducing input, relevant files and tests, and final review in the same order. Approve deletion, deployment, and external transmission separately when needed, and roll back only this task's edits while preserving other contributors' changes.