Make something useful.
See it through.

I use Codex to build software and work through review. Follow a contribution, try a small example, or find a starting point for your own project.

Define the work
Choose one useful result and write down how you will check it.
Build and check
Ask Codex to make the change, then run it and inspect what happened.
Review the result
Read the diff, respond to feedback and check the repairs.

Three worked examples

Work delivered

Start with a small API change that went through maintainer review. Then try the validation example or compare two build records.

Case 01

An S3 write needed to return its change token. Follow the API change, a compatibility repair and the accepted merge.

Public review and merge record

Case 02

A check scanned edited files but expected a marker in an unchanged file. Try a small Python example to see the difference.

Author-reported local account

Case 03

A build log names an output. Compare a complete log with a cut-short copy to see what each supports.

Pinned public synthetic evidence

AI-assisted work, with review and repairs. These cases do not establish speed, productivity or general capability.

Learn by doing

New to AI-assisted work? Start with the handbook. Already have a task in mind? Try the starter.

Agent Operator Handbook
Decide what to ask for, what information to provide, and how to check the answer.
Read the handbook
Reliable AI Work Starter
Write down one recurring task and the checks you will use to decide whether it worked.
Use the starter

Practical tools

Choose a tool for the problem in front of you.

Instructions: Tell Codex what it may change and how to check its work.
Codex Project Instructions Starter
Context: Choose source material and record why it was included.
Context Contract Compiler
QA: Check that an important workflow still works.
Green-Spine QA Pattern
Evidence: Record which version of your work the evidence applies to.
EvidenceGate

Start from the published version. Check the repository’s default branch and current commit before editing an old checkout. For a deployed site, also identify the revision that is actually live; a manual deployment may lag behind the branch. Preserve local work, then compare your intended changes against the verified base and deployed revision.
Change the source of generated content. If a script builds a page or README section, edit its input and regenerate the output. Run the generator again, or use its drift check, to confirm the output is up to date. Manual changes to generated files can disappear on the next build.
Check what the browser actually did. After resizing or clicking through a page, inspect its current width, address and visible content. Follow the links and try the keyboard controls. A successful automation call alone does not show that the intended page or mobile layout was tested.