How to Review PRs You Don't Have Context On
For Senior engineers and PR reviewers · Based on Priscila Andre's Comprehend-First AI Coding Skill
// TL;DR
Asked to approve a PR touching code you've never worked in? The Comprehend-First skill lets you build real context in seconds instead of rubber-stamping. Invoke Catch Me Up as a 'PR reviewer with partial context', select the Feature and Trace exploration modes, and ask the AI to walk you through the feature end-to-end and trace the data path the PR modifies. Demand a flow diagram, interrogate discrepancies, and review with genuine understanding. This turns rubber-stamp approvals into defensible reviews — without interrupting the PR author.
Why do PR reviews on unfamiliar code become rubber stamps?
Every senior engineer eventually gets asked to approve a PR that touches a part of the codebase they know exists but have never worked in directly. Without context, review degrades into a rubber stamp — you skim, trust the author, and click approve. That's how bugs slip through and how your name ends up on a change you can't actually defend.
The Comprehend-First skill fixes this by treating comprehension as the review's first and most valuable step. Since roughly 67% of real-world AI prompt usage is comprehension and only 2% is generation, using AI to understand the code under review is precisely the high-leverage move — not asking it to rewrite anything.
How do I build context for a review without bothering the author?
Invoke Catch Me Up with your reviewer role. Declare: 'I am a PR reviewer with partial context.' Then select Feature (how a specific feature is implemented end-to-end) and Trace (following the data or execution path). Ask a specific question: 'Walk me through how this feature works end-to-end and trace the data path this PR is modifying.'
Within seconds you get enough contextual depth to perform a genuine review — understanding the feature's shape and the exact path the change affects — without interrupting the PR author or waiting for a colleague across time zones. AI becomes the teammate who never tires of your questions.
How do I catch problems the author missed?
Interrogate the output. Don't accept the AI's first summary as ground truth. Ask follow-ups, request a visual — a table or flow diagram of the data path — and compare it against what the PR actually changes. When the AI's reading contradicts something in the diff or your own knowledge, probe the discrepancy. That's often exactly where bugs and misalignments live.
Add History mode when a change seems arbitrary: 'Why was this component changed and what is the current convention for its replacement?' Commit history and lint rule changes explain intent that the diff alone hides. Convention mode confirms whether the PR follows the codebase's established standards.
How does this raise the quality bar for what I approve?
The goal is to only ever approve keynote code — intentional, well-understood contributions you'd be proud to present publicly. The opposite is slop code: shipped without understanding, the primary source of AI-assisted bugs. As a reviewer you're a quality gate; if you approve code you don't understand, you're passing slop into a codebase that serves real users.
By comprehending first, you shift from consuming the PR to genuinely evaluating it. You can ask the author sharper questions, flag real risks, and sign off with confidence. This is the essence of the emerging Agent Manager role — senior work is increasingly comprehension, direction-setting, and quality control rather than raw implementation.
Next step
Before your next review of unfamiliar code, invoke Catch Me Up as a 'PR reviewer with partial context' using Feature + Trace. Ask for an end-to-end walkthrough and a traced data path as a flow diagram. Only approve once you can explain, in plain language, what the change does and why it's safe.
// FREQUENTLY ASKED QUESTIONS
Which exploration modes help most for PR reviews?
Feature and Trace are the core review modes — Feature explains how the feature works end-to-end, and Trace follows the exact data path the PR modifies. Add History to understand why a change was made from commit history, and Convention to verify the PR follows established standards. Together they turn a rubber-stamp into a defensible review.
How do I review without interrupting the PR author?
Use Catch Me Up to get the context AI can provide instantly. Ask it to walk you through the feature and trace the modified data path, then verify against the diff. You only escalate to the author with sharp, specific questions once you've built genuine understanding — rather than pinging them for basic context and waiting on time zones.
What if the AI's explanation contradicts the PR diff?
Probe the discrepancy immediately — contradictions between the AI's reading and the actual change are often where bugs live. Ask targeted follow-ups, request a traced flow diagram, and confirm which is correct before approving. Never sign off on code you can't explain in plain language just because the AI or author says it's fine.