Why I Built This
WhyIBuiltThis
AI writes the code and you approve the diff, but did you actually read it? One study of 16 experienced developers on 246 issues found they took 19% longer with AI help while believing they were 20% faster. In a survey of 33,662 developers, 46% said they distrust AI accuracy, and assisted writers shipped insecure code more often on 4 of 5 tasks while believing it was secure.
A diff shows what changed. It never tells you what shipped versus what you asked, in words you would say out loud. And nothing checks whether you understood the dangerous parts before you hit approve.
My Approach
MyApproach
Echo records a normal coding session and turns it into a short review. Every change gets a plain-English explanation next to its diff. Then it quizzes you on the risky ones and lets you ask why anything happened the way it did.
Three of us built it for the Meta hackathon in about a week. Every answer links back to something that really happened in the session. Ask about something that didn't, and it says so instead of making things up.
Key Features
KeyFeatures
- Records your Muse Code session quietly as you work
- Turns the session into a short review: each change explained simply, with the diff beside it
- Quizzes you on the risky changes so you prove you understood them
- Ask Why about anything. Answers point at what really happened, or honestly say it isn't there
- Shows dead ends and reverted tries that the final diff hides
- Strips secrets out before anything is saved or shown
How the System Works
HowtheSystemWorks
Record a session, read the story, prove you got the risky parts.
- Start the recorder and code with the agent like normal
- When the session ends, Echo puts together everything that happened
- Secrets get stripped out before anything is saved or shown
- You get a review: each change explained simply, with the diff beside it
- Take short checks on the risky changes
- Ask why about anything. Answers point at the session. Anything else gets refused
- Dead ends and reverted tries show up too, even though the final diff hides them
Engineering Challenges & Lessons Learned
EngineeringChallenges&LessonsLearned
Restarts broke the starting snapshot. Reviews had nothing to compare against, so now the recorder checks the starting point landed before the session continues
A risk badge on the cards went stale and sat there wrong. It came off, and unchecked sessions say Awaiting check instead
The model can't vouch for its own work. Checks only report what was actually observed, and the rest is marked as unknown
Every answer needs a real source behind it. Anything unanswerable must refuse instead of improvising
What I’d Improve Next
WhatI’dImproveNext
- Team views so a reviewer can read a session without hovering over your shoulder
- Check retakes and summaries you can paste into a pull request
- Windows support and a proper installer instead of a local setup
