The job
You need to compare two related GitHub pull requests before approving either one. One PR changes the backend contract, the other wires it into the frontend, and your review has to catch mismatched field names, duplicated logic, and files both branches touch.
Why this is hard without Sephir
GitHub keeps each pull request in its own tab, while ChatGPT only sees what you paste. You copy diffs, lose file comments, paste the second PR, then scroll between tabs to remember which change came from which branch. The breaking change hides in the gap between them.
How Sephir does it
- Pin the two GitHub PR tabs you want the agent to read.
- Open the Sephir sidepanel with
Cmd+Shift+S. - Ask it to compare the pull requests — or invoke
/pr-compareif you saved the skill. - Watch the trace:
extractPageText(tab 1)→extractPageText(tab 2)→ comparison. - Review the breakdown of shared files, contract mismatches, and follow-up questions.
- Save the trace as a skill for the next split-feature review.
The skill behind it
This skill reads both visible PR pages, keeps the tab sources separate, and returns a review you can paste into GitHub without pretending it inspected your whole repository.
What it costs
Sephir runs this recipe on your own ChatGPT Plus / API key. Typical token usage: ~3,000-8,000 input + ~500-1,500 output per run on a frontier model (Claude Opus 4, GPT-5, or Gemini 3 Pro). See for how Sephir charges for the extension, not your model tokens.
Related
- Read how Sephir works
- Compare this workflow against