meta says ai is allowed in their coding round but the policy does not mean what most writeups say it means, here is what actually works in practice
saturday morning. coffee shop on fillmore. i finally pulled up my own meta debrief notes on my phone and saw where the recruiter quoted my ide prompts back to me. like the actual sentences i typed into the ai panel during the coding round. word for word. on the debrief doc. i had to put my cup down lol. i got the offer two weeks ago and it was that detail that finally got the new ai-enabled loop to make sense to me, because every writeup online about this round is wrong in some specific way. so i pieced together a writeup from my own notes plus what a recruiter contact filled in over dms a few days later
short version. yeah meta does let you use ai during the coding round. recruiter says it on the prep call. there is an ide model you prompt during the actual coding portion. that part of the writeups is correct. the writeups stop being correct around the next bit
the in-ide ai is round specific. coding round only. they pull the panel the second they pivot to system design or to behavioral. if your prep was leaning on the model your brain has not whiteboarded unaided in weeks and now you are bare. i watched two other people on my loop go silent during that pivot. the kind of silent where you can hear the macbook fan from the other side of the call. one of them never recovered the round, you could see it on her face
bigger thing. every prompt you type into the ide ai is logged with your name attached. recruiter pulls it up next to your code in the debrief. autocomplete acceptances also count as prompts apparently. that is how i got my typing read back to me at the end. recruiter literally quoted them, the exact phrasings i had typed. the part i kept asking the recruiter contact about because i could not believe it but yeah. all of it. on the doc
worth flagging. browser tabs and chrome extensions are visible during the screenshare. one of the people on my loop had a tab open for some web based helper, the proctoring scan picked it up, the offer got pulled mid week. anything web or extension based is dead the second the screenshare is on. one whiff of it and youre out
so what worked. i tested a couple setups in the weeks before. the in-ide model is fine for the coding portion if you keep your prompts clean and remember they are being read. you do not lean on it because the second they pull it you are alone. for the rounds where it gets pulled you need something completely outside the meta sandbox. something that does not show up because it never touched their environment in the first place
ok so here is what i actually did. macbook is the audio capture point. macbook listens, transcribes the call, gets the answer back from the model and pushes it to the ipad. the ipad is just a screen at the end of all that. nothing actually happens on the ipad. no mic, no model running, the ipad does not even know what the question was, it just shows whatever the macbook decided to push to it through firestore. only reason for the ipad in this setup is so the answer text is not sitting on top of my zoom window. ipad on the desk at like a 30 degree angle, looks like a notepad, that is it. when the panel pivoted to system design the macbook was still doing its thing in the background. behavioral round, same. nothing meta could see because the macbook process is hidden the right way at the OS level. excluded from screen capture, removed from window enumeration, hidden from the dock and cmd+tab. the call is happening on the macbook and the helper is sitting next to it, but the proctoring scans never find the helper because architecturally it isnt visible to them. zero confirmed detections in any of the rounds i used it for
so the gap is the side helper, not the ide one. meta means it when they say ai is allowed but they only mean their ai for their portion of the loop. the rest of the rounds your brain has to carry, and your brain has been on autopilot for a month if you prepped wrong. having a setup that carries the design and behavioral rounds is what got me the offer. one of the other people i talked to in the candidate group chat was at the same level role with only the ide ai, got rejected after the design round, two weeks later. cleanest a/b test you are going to see in this kind of loop honestly
so question for anyone else who has done the new format. what setup did you run, did the side helper carry the design and behavioral rounds or did you only use the in-ide one. interested because the writeups online keep skipping this whole layer