QUICK ANSWER
Record the bug from its starting state, say what should happen, and share the normal Clipy link with connected Claude Code. Ask it to restate the expected and observed behavior before changing code. The recording supplies evidence; your repository supplies the implementation. Review its proposed fix and record the result.
What should you show in the recording?
Start one step before the action that fails. Show the selected filter, the form values that are safe to share, or the navigation route needed to reach the problem. Reproduce the issue once and pause on the result. A thirty-second clear reproduction is more useful than a long tour of the app.
For example, show a filter set to Active, press Save, and show that it returns to All. Say: 'I expected Active to remain selected after Save. It changes to All.' This gives the agent a specific mismatch to investigate instead of asking it to guess your intent.
Use the bug report recorder to classify the capture before recording. Narrate while you click so the transcript and frames line up.
What should you ask Claude Code to do first?
Connect Clipy through agent setup, then wait for the recording context to finish processing. Paste the watch link in the repository you want investigated with this request:
Read this Clipy recording. Restate the steps, expected result, and observed failure. Identify the relevant code in this repository and explain your proposed fix before editing. Tell me if the recording leaves an important step unclear.
The agent should distinguish what the recording establishes from its hypothesis about the cause. Seeing a filter reset does not establish whether the cause is a reducer, query parameter, or backend response. That requires code inspection.
How do you check the proposed fix?
Ask the agent to repeat the same sequence after the change, run the relevant tests, and return a recording of the result in the PR. Use the automated screen recorder when the verification belongs in a driven browser workflow.
Compare the initial state and final result against the original recording. A successful video of one path does not establish that every related path works. Keep the test results and any remaining uncertainty in the PR description.
What if Claude Code cannot read the link?
Confirm the recording has finished processing and that the agent is connected. A private or restricted recording requires authorized access. Do not change a sensitive recording to public simply to make the handoff easier; follow the access troubleshooting guide.
Can you use this without narration?
You can share visual evidence, but a silent take does not provide a spoken transcript or state what you expected. Add the expected result in your message. For a written ticket, use the AI bug report generator with a narrated recording you own.