mark erdmann
@markerdmann
Followers
1K
Following
36K
Media
350
Statuses
4K
universal software constructor. angel investor. sounding board. prev co-founder @pulley, the fastest-growing and top-rated equity management platform.
calgary
Joined December 2008
ok, 2 likes, so as promised here's the writeup. for codex cli -> how to make your terminal ping when the agent finishes a turn. no plugins. no polling. just one line in your config and a short shell script. you’ll get a mac banner and sound the instant codex completes a task.
finally got notifications working with codex cli on macos. now i know immediately when codex has finished the task and is waiting for input. if anyone else needs this, lmk via like or reply. i can write it up and post the solution.
1
0
6
the current chatgpt search experience is a tragedy of epic proportions
0
0
0
increase throughput. decrease inventory. decrease operational expense.
0
0
0
well it was great while it lasted 😂
It's over, folks. Pack it in.
0
0
0
composer-1 is cool because now i can break my code so much faster than before. psa: just use codex.
0
0
1
Someone asked recently how my new MCP Agent Mail project would work together with @Steve_Yegge 's "beads" project, which is a memory system for AI coding agents. I suspected they would actually be pretty complementary and synergistic, and it turns out they are:
I finally got around to making a tool I've wanted for a long time: you can basically think of it as being "like Gmail for coding agents." If you've ever tried to use a bunch of instances of Claude Code or Codex at once across the same project, you've probably noticed how
1
1
18
alright let's see what neo can deliver
1
0
3
@embirico @reach_vb this prompting technique unlocks 6h+ codex cli sessions. follow these steps: 1. install steve yegge's beads cli 2. ask codex cli gpt-5-high to make a plan for the work. 3. then tell gpt-5-high: "queue up all the work in bd so that we can have a swarm of subagents implement
4
3
37
strawberry browser is legit. much better than comet and atlas for work tasks. i'll record a screencast demo'ing one of the tests that i ran and then share it here.
1
0
0
managing fleets of agents should be more fun than playing factorio with the UI/UX to boot
89
53
1K
mission accomplished. 🦅 180 swiftlint violations -> 0 swiftlint violations ran for a few hours across 10 instances of gpt-5-codex. used about 20% of my codex weekly limit. 😅
this worked really well. so much faster. literally 10x faster. and there's always something to review. it's more engaging. it saves you from cmd-tabbing over to x or youtube. workflow: 1. gpt-5-high for planning 2. gpt-5-high queues up the work in `bd` 3. 10 instances of
0
0
1
this worked really well. so much faster. literally 10x faster. and there's always something to review. it's more engaging. it saves you from cmd-tabbing over to x or youtube. workflow: 1. gpt-5-high for planning 2. gpt-5-high queues up the work in `bd` 3. 10 instances of
alright, 35 issues queued up with @Steve_Yegge's new Beads cli. this is going to take forever. so let's try running 10 instances of gpt-5-codex on the same branch at the same time. sounds terrifying. but let's try it. for science.
0
0
4
"i tend to think we have more of these people than we think we do"
Marc Andreesen explains Elon’s management approach: 1. Engineer-first organizations and find truth by speaking with those working on the floor (avoid management layers). 2. Every week, find the most important bottleneck at a company and parachute in to fix it. 3. Keep model of
0
0
0
til, Ilya sutskever gave john carmack this reading list of approx 30 research papers and said, ‘If you really learn all of these, you’ll know 90% of what matters today.’ https://t.co/6eNmrgyq7k
arc.net
61
611
5K
alright, 35 issues queued up with @Steve_Yegge's new Beads cli. this is going to take forever. so let's try running 10 instances of gpt-5-codex on the same branch at the same time. sounds terrifying. but let's try it. for science.
0
0
0
@airkatakana people who cant vibe code probably never played an rts either
3
1
14
run this codex prompt twice a day: "what do you think we should improve next?" it's good hygiene, like brushing your teeth twice a day. when it runs out of ideas you've got a sparkling clean codebase. 🦷 this helps tremendously for spotting and eliminating security
0
0
0
this solution for subagents in codex cli worked brilliantly. next i'll try it with gnu parallel and then post the recipe.
the trouble is rhythm. one codex cli feels fast until it doesn’t. when the context shrinks past eighty percent, the model starts to lose its edge. so i’ll try this. let codex call `codex exec` as a subagent. the main process stays light. the heavy work happens downstream. even
0
0
1
why is every voice mode so bad? ❌ chatgpt advanced voice mode ❌ gemini voice mode ❌ grok voice mode ❌ perplexity voice mode any others i should try? here's my problem: i need to walk 10k steps a day to please my whoop overlord. that's an hour and a half of walking. who has
0
0
3