MasuSenseiBotV3 Agent

Overview

MasuSenseiBotV3 Agent is the third iteration of an AI mentorship pipeline I've been refining for PR review workflows. The earlier versions tried to do too much. This one focuses on a single thing - giving educational, contextual feedback on pull requests, the kind a senior engineer would give if they had time to write thoughtful reviews on every change.

How it works

The pipeline ingests PR diffs, computes context around the changes, and runs them through an LLM with prompts shaped specifically for educational tone. The output is a structured review comment that explains what the change does, what's strong about it, and what could be improved - ideally in a way that helps the contributor learn, not just patch the code.

Why educational tone matters

Most AI code review tools optimize for catching bugs. That's useful but it misses the part that actually grows engineers. I wanted this bot to feel like a mentor who explains the why behind each suggestion, references patterns the team uses, and treats the contributor as someone learning rather than someone making mistakes. Tuning the prompts for that tone took longer than the technical pipeline.

What I learned

The biggest lesson was that LLM systems for human-facing workflows need a lot of guardrails around tone. A small prompt change could shift the bot from helpful to condescending in ways that were hard to spot until you read enough sample reviews. I ended up writing tone-checking prompts that ran on the bot's own output before posting, which sounded silly at first but actually worked.