A simple sequence#
- Open one non-sensitive Codex task and note its title.
- Speak a sentence with a number and a constraint: “Give two options; do not apply either.”
- Read the transcript and correct any changed meaning.
- Confirm the task title and send once.
- Inspect the original conversation, then wait for the separate answer.
Record whether each stage worked. If capture failed, inspect input readiness. If the sentence arrived but was wrong, assess transcription. If text was correct but reached the wrong destination, assess conversation selection. If a receipt exists without an answer, assess task execution.
Extend one variable at a time#
After a basic pass, try one attachment or a second conversation. Test softer speech and a normal office environment separately. Keep the prompt content constant so you can tell which variable changed the outcome.
Stop after a meaningful failure and retain the smallest useful evidence: app version, visible state, input device and expected result. Repeating the whole workflow without checking where it broke creates ambiguity and can accidentally duplicate instructions.