Veyro

Claude tricks most people don't know. Prompts, settings, workflows

[Veyro @0xVeyyro]

@bubosees Yeah, that line is the whole reason I trust the setup. Autonomy with a hard stop.

[Veyro @0xVeyyro]

f*cking insane. you can give Claude Code a night shift: 1 rule, your laptop shut, and Opus 5.5 has a stack of tested fixes waiting for you by morning

the AI chief of staff post above hit 350K views overnight. here's the dev version, and it's 2 things already in the Claude Code docs: claude -p and a cron job

no junior on nights. no $200K hire. it costs whatever you cap it at with --max-budget-usd

you give the goal, it does the steps, you only make the calls

the rule:
"clear my issue backlog. reproduce before you fix. never touch main. merge nothing without my yes"

the night:
→ 11pm Opus 5.5 reads every open issue and splits the work into jobs
→ 12am Haiku 5.5 subagents try to reproduce each bug in their own worktree, dupes and stale ones get cut
→ 1am Sonnet 5.5 writes a fix for every bug it could reproduce, 1 branch per issue, tests included
→ 3am a reviewer that never saw the writer's chat reads every diff: PASS or FAIL
→ 5am every FAIL gets 1 more try, then it's parked with a 1-line note
→ 8am you open summary.md with your coffee

honestly the "park it" step is the whole thing. an agent that can't say "i couldn't fix this" ships garbage straight to your main branch

and it runs again every night at 11pm. it never forgets to start

312 open issues → only the bugs it can reproduce → only the fixes a second model passes → 0 merged without you

it opens branches. it never merges

someone's backlog is already getting cleared while they sleep

the full setup is below ↓ free, steal it

[Argona @Argona0x]

https://t.co/uFcQbivrbz

[Veyro @0xVeyyro]

@bubosees Grok is as shameless as ever

[Veyro @0xVeyyro]

@andreysuperior So Space Bunny was just MiniMax this whole time?Wild

[Veyro @0xVeyyro]

@alxqwrk One afternoon witch Claude and the ads are gone They should be scared

[Veyro @0xVeyyro]

@Klonzu the compiler isn’t the brain, what you feed it is. this pipeline actually makes sense

[Veyro @0xVeyyro]

@Synthexfr 400 lines every session is brutal. moving the procedures to skills is the right call

[Veyro @0xVeyyro]

@itsarcvale @t4t5 same hash 17 years later is wild. now do this for every wallet I run

[Veyro @0xVeyyro]

everyone's counting the 300 agents.

nobody's asking who checks their work.

that's the part that actually matters. 300 agents with nobody checking = 300 ways to merge garbage at once.

the fix is stupid simple. a second model that never saw the code being written reads the diff and says yes or no.

1 agent writes:

claude -p "Fix issue #142. Commit to branch fix-142."

a cheap Haiku 5.5 checks:

git diff main..fix-142 | claude -p --model haiku "You didn't write this. Does it fix issue #142? PASS or FAIL + 1 reason."

FAIL → back to the writer
PASS → merge

that's it. it works because the checker has zero reason to defend the code. it didn't write it.

and Haiku 5.5 is $0.10 per 1M input tokens. checking 300 diffs costs less than your coffee.

for the merges that can break prod, swap haiku for opus.

checker prompt in the reply. save this before your agents start shipping

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Veyro @0xVeyyro]

@alxqwrk Try it with something personal. Portrait, old photo, random screenshot. The output changes a lot depending on the image.

[Veyro @0xVeyyro]

It is becoming f*cking dangerous: 1 prompt to Claude and it built a lab that turns any photo into a living 3D data structure

1 HTML file. No install. No backend.

Upload an image and it rebuilds it as:
→ a rotating ring of dot-matrix blocks sampled from your pixels
→ thousands of spikes, taller where the image is brighter
→ a web of threads connecting the darkest points
→ the original photo, redrawn as halftone dots underneath

The camera cuts every 6 seconds. Overview. Dive into the spikes. Skim the ring.

There's a record button. It exports the whole thing as a video.

This one is a diagram of diseases I found on Pinterest.

The prompt is below. Save it before you need it.

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Veyro @0xVeyyro]

@maixbt Exactly. Screenshots die in 3 seconds. The record button turns it into something you can actually share and drop anywhere.

1 file is the other half of it zero friction.

[Veyro @0xVeyyro]

@andreysuperior One evening to rebuild a paid feature is wild. Companies are going to hate this

[Veyro @0xVeyyro]

@andreysuperior One prompt to reverse-engineer paid features is wild. How solid is the verification step in practice?

[Veyro @0xVeyyro]

@Richix01 This is the comment the original post needed. OS matters more than the prompt

[Veyro @0xVeyyro]

THIS IS A PURE F*CKING GOLDMINE

claude code can split one brain into three models, and most people still run everything on opus.

You just launch it like this:

claude --advisor opus

here's the split:

→ Haiku 5.5 explores the codebase for $0.10 per 1M input tokens
→ Sonnet 5.5 writes the code
→ Opus 5.5 sits as the advisor and only steps in on the hard calls
→ your opus bill stops paying for file searches

Haiku reads · Sonnet builds · Opus checks

the full tree
> main session on Sonnet 5.5
> Explore subagent pinned to haiku in .claude/agents/Explore.md
> Opus 5.5 as advisor for plans and reviews
> CLAUDE.md rule: ask the advisor before any change that touches 3+ files

paste the tree into Claude Code ↓

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Veyro @0xVeyyro]

claude code can remember everything you tell it. most people never set this up.

→ make a folder. open claude code inside it
→ add a MEMORY/ subfolder
→ CLAUDE.md gets one rule: save every key detail to MEMORY/
→ link the notes into a graph so it recalls, not re-reads
→ give it a browser and one command that runs the whole thing

20 minutes of setup. it stops starting from zero every chat.

this is what that memory looks like ↓

[Veyro @0xVeyyro]

gave opus 5.5 one 1-second reference clip and one rule: no images.

it wrote a single 78-line shader and rebuilt the orb from pure math.

6 passes. 0 textures. 1 draw call.

the reference is playing on the right ↓

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[Veyro @0xVeyyro]

Claude Code now compacts your session while you're away.

Leave a 200K+ token session idle for about 54 minutes and it compacts before the 1-hour cache expires.

You come back to a cheaper context instead of a full cold reload.

2 settings worth knowing (v2.1.290+):

→ CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS changes the threshold
→ "idleCompaction": false turns it off

Pair it with the team setup below ↓ Haiku searches, Sonnet builds, Opus checks.

[Veyro @0xVeyyro]

Google just dropped Gemini 4 Argon. Its own benchmark table has Opus 5.5 in it.

Argon has the top score in 14 of 19 rows.

But look at the one Claude Code users care about:

Terminal-bench 4.0
→ Opus 5.5: 66.4%
→ Argon: 57.4%

Same price after the intro: $4 in, $20 out per 1M tokens.

Argon for long documents and agents.
Opus for your terminal.

Which one are you running this week?