Use Codex without giving up engineering judgment

Give a bounded task

Name the user outcome, relevant files, constraints, and acceptance checks. Ask for the smallest coherent change rather than an open-ended rebuild.

Request evidence

Have the tool run tests, formatters, and a focused demonstration. A plausible diff is not proof that the feature works.

Review risky boundaries

Inspect secrets, authentication, money, destructive file operations, database migrations, and external calls carefully. Add explicit confirmation or dry-run behavior where appropriate.

Keep changes reversible

Use small commits, preserve existing interfaces, and avoid mixing unrelated cleanup with the feature. Reversibility makes experimentation inexpensive.

Working example

# Example task brief for an AI coding agent
TASK = {
    "outcome": "Reject duplicate customer emails during CSV import",
    "scope": ["src/importer.py", "tests/test_importer.py"],
    "constraints": [
        "preserve the existing public function signature",
        "do not discard invalid rows silently",
        "return row numbers for every duplicate",
    ],
    "verification": [
        "existing tests pass",
        "new duplicate-email tests pass",
        "sample import produces the documented report",
    ],
}