Manual verification
Every code change comes with a clear description of what changed, what to verify manually, and which optional build or test would normally be useful.
Downloadable Codex skill
personal-code-developer
AI coding works best when the changes stay focused, failures stay visible, and verification stays in your hands.
Coding assistants can move quickly, but speed is not always the same as control. Unrequested builds, broad refactors, speculative abstractions, and silent fallback behaviour can make a small task harder to review and less predictable.
I created a Codex skill called personal-code-developer to make my preferred development workflow explicit: inspect first, work within the existing design, preserve unrelated changes, and hand the result back with practical manual checks.
Two core principles
No silent fallbacks
Failed transformations and uncertain operations remain visible instead of being hidden behind best-effort output, swallowed exceptions, or ambiguous defaults.
The skill also
- Uses existing repository patterns first
- Avoids speculative refactors and broad rewrites
- Preserves unrelated user changes
- Keeps experiments isolated and clearly labelled
- Separates facts from assumptions in handoff notes
- Makes unsafe or indeterminate failures explicit
The key idea is simple: automation should remain deliberate.
The assistant can still inspect, diagnose, and implement effectively, while leaving execution and verification choices visible to the developer.
Small scope. Clear failures.
Manual confidence before moving on.
The complete package includes SKILL.md and Codex interface metadata.
Download skill (.zip)