WHAT NEW IN X11.03 TUE

AI collections from X · 22 collections · 65,263 posts

personal-agentVIDEO

OpenAI Dots efficiency tips: hard-coded thresholds for your agent

The author shares how they save about 3 hours a day with OpenAI Dots by writing fixed numeric thresholds and stop conditions into briefs for inbox triage, research, and repackaging tasks.

deezzex
@deezzexArchived 2026.10.03Originally posted 2026.10.03
deezzex @deezzexOriginally posted 2026.10.03
Open original on X

Sam Altman gave me an always-on agents: Dots. ex OpenAI engineer leaked how to make it extremely efficient (+200%) for free i'm down ~3 hours a day on routine work. not because the agent is smart. because i finally wrote the rules down. here's the actual breakdown: → inbox triage: 50 min/day. not "monitor my inbox." the brief names 6 senders, a 24h window, and one rule: ping me only if there's a decision or a date in it. everything else gets a one-line summary i read once, at 6pm. it's wrong maybe twice a week and both times it was over-cautious, which is the direction i want it wrong. → research digest: 70 min/day. this used to be me with 20 tabs open. now it's a recurring brief with a fixed source list and a threshold: flag anything that changes a number i'm already tracking. no "interesting article" summaries. if it doesn't move a number, i don't want to see it. → repackaging: 60 min/day. long-form into clips, notes, and drafts. this is the one where the cloud computer matters - it's running while i'm asleep, a and i wake up to drafts instead of a blank doc. i still rewrite every one of them. that's fine. editing is 10 minutes, writing cold is 50. the part i got wrong at first: i gave it broad goals and let it decide when to surface things. that's the trap. a proactive agent with vague instructions doesn't save you time - it generates a second inbox you now have to manage. for about a day i was spending more time reviewing its judgment calls than i'd spent doing the task. the failure mode of an always-on agent isn't that it does the wrong thing. it's that it tells you about everything it did. so the fix wasn't a better prompt. it was a stop condition. the rule i use now: if you can't write the threshold as a number, the task isn't ready to delegate. source, window, number. if a brief is missing one of the three, it will interrupt you, and an agent that interrupts you costs more than it saves. three hours came from three specific tasks. not from "ai doing my job." what's the one recurring task you'd hand over first - and could you actually write its stop condition in one sentence?

This post quotes another post: x.com/yurshevv/status/2106054778815983957

x.com/deezzex/status/2106062270153339091Archived 2026.10.03 10:34