Zuvex

copy-paste AI setups: Claude Code, agents, AI video

[Zuvex @zuvexxx]

SOMEONE AT ANTHROPIC TYPES ONE COMMAND AND 9 AGENTS HUNT HER BUGS WHILE SHE KEEPS WORKING

the trick is the harness around the model. i rebuilt it as a sim in my video: same Opus 5.5, same 20 tasks, bare prompt vs harness

Lydia Hallie from the Claude Code team calls it a dynamic workflow. she uses them for any task she wants to rerun the same way, even small ones with ~10 subagents

the setup:
→ ask Claude "make a workflow that does X" and it writes a JavaScript file
→ agent(), pipeline() and parallel() are globals, no imports
→ each subagent gets its own prompt, phase and effort level
→ save it from /workflows, rerun it later as /name
→ it runs in the background while you keep working

her /bug-triage example:
→ phase 1: agents trace every open issue to a root cause, citing files and lines
→ phase 2: Opus tries to refute each cause with a failing test in a separate worktree
→ only confirmed bugs make the final report

she says it's more predictable than a skill, because you decide which subagent runs and in what order

run A vs run B in the sim:
→ accepted: 8/20 vs 16/20
→ runaway loops: 5 vs 0
→ cost per accepted task: ~$4.10 vs ~$1.60
→ babysitting per day: ~2h vs ~25m

the prompt to build your own /bug-triage is in the first reply. full harness guide below ↓

[Zuvex @zuvexxx]

@geravita_waw do it, takes like 5 min to set up

[Zuvex @zuvexxx]

KARPATHY'S SECOND BRAIN IS 2 FOLDERS AND 1 FILE. CLAUDE CODE CAN BUILD YOURS FROM ONE PROMPT

his post on it hit 22M views and 108K bookmarks. a big chunk of his tokens now goes into knowledge, not code, and he says he rarely writes or edits the wiki himself

the setup:
→ raw/ holds your sources. Claude reads them and never edits them
→ wiki/ is what Claude writes: one page per topic, linked, updated with every new source
→ CLAUDE.md is the schema: where things live and how pages are written
→ in his words, Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase

3 moves run it:
→ ingest: drop a source in raw/, Claude updates every page it touches
→ query: ask, Claude answers from the wiki and names the page
→ lint: Claude hunts contradictions, orphan pages and stale claims

good answers get filed back in as new pages, so the wiki grows from your questions too

Obsidian's CEO Steph Ango open-sourced obsidian-skills, so Claude writes real Obsidian markdown, Bases and canvases instead of guessing:
/plugin marketplace add kepano/obsidian-skills
/plugin install obsidian@obsidian-skills

the prompt is in the first reply ↓

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Zuvex @zuvexxx]

@geravita_waw ayy thanks

[Zuvex @zuvexxx]

SOMEONE AT ANTHROPIC TYPES ONE COMMAND AND 9 AGENTS HUNT HER BUGS WHILE SHE KEEPS WORKING

the trick is the harness around the model. i rebuilt it as a sim in my video: same Opus 5.5, same 20 tasks, bare prompt vs harness

Lydia Hallie from the Claude Code team calls it a dynamic workflow. she uses them for any task she wants to rerun the same way, even small ones with ~10 subagents

the setup:
→ ask Claude "make a workflow that does X" and it writes a JavaScript file
→ agent(), pipeline() and parallel() are globals, no imports
→ each subagent gets its own prompt, phase and effort level
→ save it from /workflows, rerun it later as /name
→ it runs in the background while you keep working

her /bug-triage example:
→ phase 1: agents trace every open issue to a root cause, citing files and lines
→ phase 2: Opus tries to refute each cause with a failing test in a separate worktree
→ only confirmed bugs make the final report

she says it's more predictable than a skill, because you decide which subagent runs and in what order

run A vs run B in the sim:
→ accepted: 8/20 vs 16/20
→ runaway loops: 5 vs 0
→ cost per accepted task: ~$4.10 vs ~$1.60
→ babysitting per day: ~2h vs ~25m

the prompt to build your own /bug-triage is in the first reply. full harness guide below ↓

[Trackmind @0xTrackmind]

https://t.co/rdQRAsouJS

[Zuvex @zuvexxx]

@Flinzu yeah, once raw/ is filled the hints get shorter every week

[Zuvex @zuvexxx]

@0xDianax same apartment, two escapes, two completely different outcomes. the only thing that changed was who listened

[Zuvex @zuvexxx]

@suttonwwwen the key to the safe being in a drawer is the detail that changes everything. a locked safe only works if the key is too

[Zuvex @zuvexxx]

@ESC__AI the fight is the demo, the disaster zone is the product

[Zuvex @zuvexxx]

@Vectorxeth reading the signature on the hardware wallet screen is the only check that the network can't fake

[Zuvex @zuvexxx]

@motyka_9 local for the 24/7 grunt work, cloud for the hard calls. that split is where this actually pays off

[Zuvex @zuvexxx]

@0x_stat1x the risk management is the real story

[Zuvex @zuvexxx]

@AIBoticssq that’s what I like about it. after a while you’re not just asking Claude questions, you’re building something it can actually think with

[Zuvex @zuvexxx]

GPT started as a Reddit chatbot that looked like a waste of time. Andrej Karpathy, now at Anthropic, told the whole story to a Berkeley hackathon crowd

early OpenAI was 8 people working out of Greg Brockman's apartment, up against Google, which employed almost half the AI research field. and one guy there was building a chatbot for Reddit

Karpathy said if you walked past his desk you'd think: "we're trying to compete with Google and you're working on a Reddit chatbot? we should be doing something bigger"

that chatbot was a language model trained on Reddit. then the transformer came out, the data grew from Reddit to the whole web, and it became GPT-1, 2, 3, 4 and 4o. by the time of the talk in 2024 he put OpenAI at close to $100B

what he told the room after that:
→ small projects look fragile and easy to dismiss. that's how the big ones start
→ success is mostly hours. count them and stop stressing over every win or fail
→ his weekend projects look like magic because it's his 20th time building the same kind of thing
→ publish what you build. knowing people will see it makes you work harder

if you build small stuff with Claude Code on weekends, these are the 8 minutes to watch. filmed from the crowd, around 1K views on YouTube

[Zuvex @zuvexxx]

@Tetriss112 that’s the beauty of it. you rarely know which tiny experiment is going to turn into something big

[Zuvex @zuvexxx]

GPT started as a Reddit chatbot that looked like a waste of time. Andrej Karpathy, now at Anthropic, told the whole story to a Berkeley hackathon crowd

early OpenAI was 8 people working out of Greg Brockman's apartment, up against Google, which employed almost half the AI research field. and one guy there was building a chatbot for Reddit

Karpathy said if you walked past his desk you'd think: "we're trying to compete with Google and you're working on a Reddit chatbot? we should be doing something bigger"

that chatbot was a language model trained on Reddit. then the transformer came out, the data grew from Reddit to the whole web, and it became GPT-1, 2, 3, 4 and 4o. by the time of the talk in 2024 he put OpenAI at close to $100B

what he told the room after that:
→ small projects look fragile and easy to dismiss. that's how the big ones start
→ success is mostly hours. count them and stop stressing over every win or fail
→ his weekend projects look like magic because it's his 20th time building the same kind of thing
→ publish what you build. knowing people will see it makes you work harder

if you build small stuff with Claude Code on weekends, these are the 8 minutes to watch. filmed from the crowd, around 1K views on YouTube

[Zuvex @zuvexxx]

@JavierEraia haha, imagine pitching that in a quarterly budget meeting. you’d be out before finishing the sentence

[Zuvex @zuvexxx]

@OxPips and those 19 weren’t wasted time. they were what made the 20th possible

[Zuvex @zuvexxx]

@0xGisu yep. everyone sees the 20th attempt and calls it genius. nobody asks what happened to the first 19

[Zuvex @zuvexxx]

@Orvexiumx glad you enjoyed it. thought this story deserved more attention

[Zuvex @zuvexxx]

@thegreatest_sv we’ve reached a point where even borrowing a charger requires a little trust check. that’s a weird world to live in

[Zuvex @zuvexxx]

@issatheblows the part i find interesting is how many bad trades never happened. those are invisible in the profit screenshot

[Zuvex @zuvexxx]

@AIMind_Ai extra thinking really said wealth is temporary, peace of mind is forever