Use visual targeting, not guesswork
A screenshot tells Codex what the page looked like. An annotated screenshot tells Codex what to care about. That difference matters because frontend issues often have many plausible targets on one screen.
How to show Codex exactly what broke
Codex is much more useful when you stop asking it to infer the target from a raw screenshot. Show the exact broken item, the expected behavior, and the surrounding state.
Pair the image with a short note
The strongest handoff is usually a screenshot, a box or arrow, and one short sentence. Say what is wrong and what should happen instead. That usually beats a long technical explanation.
Include context only when it helps
Codex does not need every possible signal. It needs the signals that reduce ambiguity. LLM Clip keeps the image first, then adds lightweight context when the issue needs it.