set ourselves 300 installs of PDF Desk by 30 nov. last count was 4 chrome users. the code shipped fine, getting found is the slow part. moved 300 to 21 dec and kept 30 nov as a check at about 25
1
8
lesson learned: my marketing bot was waking 9 times a day for small checks. cut it to 2 runs, 7:30am and 1pm my time, which lands on the US afternoon and evening. same jobs, fewer wake ups, less spend
1
30
let our coding runs start their own next job when they finish. the review bot failed it 3 minutes later. the job file carried its own check, so anything that could write it could start any script. nothing had run yet. get the reviewer in before the live run
15
working hack: the more work the agents do, the more there is for me to review
we've started putting related options on one review page instead of chat cards. compare, pick, add notes, send it back in one go
still early but it's a better way to get through a big batch
1
21
the bots were loading hundreds of old notes every turn. we archived them instead of deleting and merged the repeats into short lines. the main one went from 407 down to about 73
1
1
23
one goal, one coding session, one pr. the run pushes its own branch and opens the pr without waiting on me. merges still stop for my ok
2
29
the agents were stopping for my ok on every step of a plan. thats what was slowing us down. now one ok covers the whole plan, and only the stuff you cant undo still waits: merges, sends and money
23
our house rules had hit 64 versions in 22 days so we started cutting. the file went from 113 lines to 52 in one pass. a little has crept back since. delete first still holds
gabbaworks.com/builds/simpli…
20
put @OpenBot_ in a vm cos i wanted to keep everyday files safer from unwanted changes. it added friction from shell limits to memory and recovery but we kept it. heres what we checked and what still needs proof
gabbaworks.com/builds/why-op…
1
39
got a real android phone tapping through Plant Trouble on its own while a local model checks the screenshots
12
our coding jobs kept finishing way early. one we guessed at 30 minutes took 10, most came in around a quarter of the guess / so the pcs sat idle till the next check-in / now every slot gets loaded with 150 percent of the gap. an idle pc wastes more than a job that waits
20
bumped my @OpenBot_ vm from 8gb to 12gb. picking the number was easy. the shutdown restart and checking the guest afterwards was the real work. next is watching it under real load / gabbaworks.com/builds/openbo…
1
28
my grok bot agents started acting like government workers. house rules hit 64 versions in 22 days, mostly checks on checks. grok bot is from @elonmusk's spacex, so i'm cutting them with his five-step algorithm. wrote it up, plus a prompt for your agents
gabbaworks.com/builds/simpli…
1
1
46
our coding agents burned through a whole month of free github ci minutes before the month was out. 2,000 minutes gone / tests moved onto our own pcs and pushes got batched till the reset / ci minutes are a budget now, same as cash
3
2
119



