Anatomy of a good bug report for AI agents

A great report for a human is often a poor one for a machine. Agents reward precision: a selector beats a sentence, an error beats an adjective.

Three ingredients

• Target: a stable selector or XPath the agent can locate without guessing.
• Intent: one plain sentence describing the desired outcome.
• Evidence: the console error or failed request that proves the problem.

Vynix structures every note around exactly these. The result is a report your agent can act on deterministically, and a paper trail you can review later.

Pick the right output

Use Markdown for humans, the AI prompt format for coding agents, and JSON for pipelines. Same note, three shapes: choose per task from the widget toolbar.

A worked example

Compare two reports for the same problem. The human version reads "the sign-up button feels off on mobile." The agent version names the target, states the intent, and attaches the proof:

### Increase tap target on the sign-up button [bug · high]
- Selector: header nav button.signup
- Page: / (viewport 390x844)
- Computed: height 28px, font-size 13px
- Console: (none)

The second one is actionable on the first read. There is nothing to interpret and nothing to ask.

Common anti-patterns

• Describing appearance instead of the element: "the header is broken" gives the agent nowhere to start.
• A screenshot with no selector, so the agent has to reverse-engineer the target from pixels.
• Batching three unrelated problems into one note, so the fix cannot be verified cleanly.

Let the tool enforce the shape

You should not have to remember all of this. Vynix captures the target and the evidence automatically as you click, so every note arrives in the shape an agent can use. Good structure becomes the default, not a discipline.