Krunzar

Testing AI so you don't have to. Tools, workflows, and honest takes.

[Krunzar @Krunzar]

CLAUDE DUET CHECKS YOUR APPS IN THE BACKGROUND AND SHARES YOUR PLAN LIMITS. SET IT UP BEFORE IT SETS ITSELF UP

I don't have access yet, so this is from the public reports, not a test. Two facts matter most: Duet has no separate allowance, and background checks count toward your normal limits. Nobody has published what one check costs.

the prep list:
→ connect one app first, the one you open most, then add more only after a week
→ set the report frequency yourself, it checks connected accounts every few hours by default
→ write down your usage before turning it on, so you can see the difference after
→ keep every approval prompt on: it asks before sending, paying, signing or deleting
→ stop doing "connect everything on day one", each extra app is more background work
→ stop treating the morning overview as free, it comes out of the same pool as your coding sessions
→ decide in advance which tasks you will hand over: scheduling and inbox triage first, forecasts and web app prototypes later
→ track it for 7 days: limit used before, limit used after, how many alerts you actually acted on

Anthropic calls it an early version that may misread a thread or misjudge urgency, so read the first week of alerts critically.

Result: instead of guessing why your limit ran out on Thursday, you will know what Duet used and what it earned.

I'll post my numbers once I get access ↓

[Krunzar @Krunzar]

96 RUNS A DAY, AND MOST OF THEM SHOULD FIND NOTHING

I haven't used Grok Bot, so this is about the pattern in chapter 04 below, not the app. it applies to any scheduled agent, Claude Code included

the math in the article: a check every 15 minutes is 96 runs a day. hourly is 24. three times a day is 3. one user's estimate at API rates: $397 a month at every 5 minutes, $33 hourly. same job, 12 times the runs

the fix is a footer on every routine:
→ keep the IDs you already saw in a state file
→ first check only whether anything is new. if not, answer one word, NOOP, and stop
→ never reread old threads. report only what changed
→ use connectors first, browser only when there is no connector
→ cap retries at 3, then stop and say where it got stuck
→ cap the reply at 5 lines and save the rest to a file
→ start a fresh chat per task, because every turn resends the whole chat
→ measure before cutting: note usage, run one task, note usage again
→ once a week, list every routine with how many runs were NOOP and whether you used the result

an empty run still wakes the agent, so frequency comes first and reading comes second. a routine you never act on is the first one to pause

chapter 04, "pay only for runs that find something" ↓

[Argona @Argona0x]

https://t.co/zdCESaoy7V

[Krunzar @Krunzar]

BEFORE YOU PASTE A "SET IT UP" PROMPT INTO CLAUDE CODE, DO THESE 7 THINGS

the 10-repo guide below gives a paste-in prompt for every repo. I haven't run any of them, so this is about the setup, not the tools

the article itself flags the main risk once. under RD-Agent it says the tool writes code and runs it, so use an isolated machine with no real credentials. that rule fits all ten

→ run installs in a fresh VM or container, not on the laptop with your real keys
→ pin a tag or commit. the https://t.co/0c7vFTCKoY prompt installs TradingAgents straight from a git URL with no version, so a push tomorrow changes what runs tonight
→ start with paper or read-only keys. an exchange key never needs withdrawal permission
→ keep keys in env vars and name the variable in the prompt. don't paste the key into the chat
→ set a spend limit on the API key before anything runs overnight
→ leave Claude Code on ask-before-run for the install step and read each command before approving
→ afterwards run pip freeze and git status, so you know what was added and what changed

one paste is fast. the part worth slowing down is what the paste is allowed to touch

setup hygiene only, no financial advice

full guide in the post below ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG

[Krunzar @Krunzar]

@tonytonggg thanks bro)

[Krunzar @Krunzar]

THE BEST TWO LINES IN THIS TRADING-REPO POST ARE #7 AND #9

I haven't run Vibe-Trading, so this is about the checks, not the tool. step 7 says an edge that works in only one window isn't an edge. step 9 says paper trade for a few weeks before trusting any backtest. both apply to every backtest an AI agent writes for you

6 checks before you believe one:
→ rerun it with fees and slippage. they are the usual first casualty
→ test on a window the agent never saw
→ run 100 random-signal versions of the same strategy. if random does about as well, there is no edge
→ shift the signal by one bar. if the result collapses, the strategy was peeking at the future
→ check where the data came from. missing days or a fallback feed can fake a smooth curve
→ ask the agent to list every assumption it made, then check each one against the data

the agent that wrote the strategy is also the one grading it. a self-graded backtest only tells you what the agent believes, so give the grader something to fail

same rule as any AI output: write the check first, then trust the green. a result you cannot try to break is not a result yet

research tool, not financial advice. past results don't predict future ones

the repo and the full guide are in the post below ↓

[Сarm1ne @carm1nee]

this github repo should honestly be ILLEGAL for retail traders to have for free

hedge funds pay quant PhDs six figures to find alphas, backtest them and stress-test every trade - this repo turns the whole quant desk into one sentence you type in a terminal

it's called Vibe-Trading. built by a University of Hong Kong lab, 36.4k stars and it hit #1 repo of the day

first understand how it works:

> you type one sentence in plain English
> a team of AI agents takes it: quant, crypto, macro, risk
> they pull the data, write the strategy code and run the backtest
> then they try to break it: Monte Carlo, Bootstrap, Walk-Forward
> it tests 478 famous quant alphas in one line and tells you which are still ALIVE and which are DEAD
> what survives exports straight to TradingView Pine Script or MetaTrader 5
> it's a full quant desk. it just runs on your laptop

here is how to run it yourself for FREE:

1. open your terminal and install it:
pip install vibe-trading-ai

2. run vibe-trading init and pick Claude Code as the brain. it runs on your Claude Code login, no separate API key

3. type one sentence:
vibe-trading run -p "Backtest a BTC-USDT 20/50 moving-average strategy for 2024 and summarize return and drawdown"

4. watch the agents pull the data, write the code and run the backtest live

5. bench the alpha zoo in one line:
vibe-trading alpha bench --zoo gtja191 --universe csi300 --period 2018-2025 --top 20

6. call a whole crypto desk:
vibe-trading --swarm-run crypto_trading_desk '{"asset": "ETH-USDT", "timeframe": "1w"}'
funding, liquidations and flow all go to a risk manager before anything moves

7. walk-forward it. if the edge only works on one window, it's not an edge

8. export what survives to TradingView:
vibe-trading --pine <run_id>

9. paper trade it with vibe-trading paper for a few weeks before you trust a single backtest

i won't say anything more:

[Krunzar @Krunzar]

THE NUMBERS IN THIS ZOOM AGREE WITH EACH OTHER. I CHECKED

the video shows two readouts in every shot: magnification and field of view. I paused three shots and multiplied them

→ 671,987,859x × 3.12 m ≈ 2.1e9 m
→ 4.4e20 × 4.81 pm ≈ 2.1e9 m
→ 5.9e18 × 355 pm ≈ 2.1e9 m

the exact products are 2.097e9, 2.116e9 and 2.095e9, within about 1 percent. the readouts are rounded, so that is the expected spread

same product every time, so the readouts most likely come from one shared scale in the code, not from numbers typed in by hand

the 30-second version for any AI zoom or chart:
→ find two on-screen values that should multiply to a constant
→ pause three shots and multiply them
→ if the product drifts, the numbers were decorated, not computed
→ ask the model to list every on-screen number with its formula, then recompute them in code
→ if you build one yourself, put the shared scale in one variable so every shot uses it
→ keep the recompute script next to the video file

this only proves the numbers are consistent. whether the physics is right needs a source, like a textbook or NASA. and the video is rendered, not filmed, which is exactly why a check like this is possible

his guide on how these are made is in the post below ↓

[clayo @0xclayo]

f*ck YESS, Opus 5.5 just flew into the sun and filmed two protons fusing

It's f*cking insane…

Drop the PDF below into Opus 5.5, thank me later

Then read the full guide on Motion Engineering with Opus 5.5 in the article below ↓

[Krunzar @Krunzar]

ONE WRONG CONSTANT PASSED THE EXACT CHECK AND BROKE A GAME

it's in the reverse-engineering guide below, where the check is called an oracle. it applies to anything you verify with AI

a function in a Touhou rebuild matched the original byte for byte. the code checked player power against 0.0. the original used 128.0. the comparison tool adjusted addresses but never looked at the number stored there

the fix, per the REA blog: audit 1,548 float references. that turned up 12 more wrong ones across 5 functions that were already accepted

how to test your own check:
→ plant 3 wrong versions on purpose: one constant, one swapped branch, one off-by-one
→ run your check on them. all 3 must fail
→ if one passes, fix the check before trusting any green result
→ re-audit everything the old check approved
→ write down which check approved which result, so a broken one can be traced
→ keep the planted failures in the repo so the next session reuses them
→ ask "what would make this fail", not "does it pass"

I do a small version of this on my own videos: a contact sheet of 5 stills before I send. it caught a label overlapping a ring that I never saw at full speed

full guide with the prompts ↓

[NO1ennn @N01ennn]

REA is pure f*cking treasure right now

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

Opus 5.5 reads how a feature works, then rebuilds its own version from scratch

but one wrong number can pass every check

in this article : https://t.co/erXf2CsIkl

[Krunzar @Krunzar]

@beamnxw the permission file is the underrated part. a rules.md that lists what needs your yes (send, buy, delete, publish) is plain text, so you can diff it and any agent reads the same limits

[Krunzar @Krunzar]

@real_mahalingam the 70% from one prompt is the useful number. save the iterations that got it the rest of the way as a dated decisions file, so the next session doesn't have to rediscover why the last one chose what it chose

[Krunzar @Krunzar]

@DerekNee worth adding the other side: the summary can drop exact instructions and in-progress state, and you may get no warning. keep the hard rules in CLAUDE.md, not only in the chat, so they don't depend on what a summary keeps

[Krunzar @Krunzar]

every new Claude chat starts at zero. here is the 5-file folder I'd build so it never does again

make one folder and put these files in it:

> me.md
who you are, who you work with, how you write. paste a few emails you sent and ask Claude to pull it out for you

> people.md
one line per person: their role, what they asked for, what you promised them

> rules.md
what Claude may do alone and what needs your yes. "draft first, never send"

> now.md
this week only: projects, deadlines, open loops. rewritten every Friday

> log.md
every decision with a date, one line each

then add one instruction to CLAUDE.md in Claude Code, or to your Project instructions: read me.md and rules.md first, then only the files the task needs

what changes:
→ stable facts, like who people are, live in files
→ changing facts, like dates and amounts, get re-checked at the source
→ every file carries a last-checked date, so stale ones are easy to spot
→ you fix one file instead of re-explaining in every chat

stop doing this:
→ pasting your whole life into

[Krunzar @Krunzar]

@Argona0x slow facts live in files, fast facts get reopened at the source. that rule prevents the usual failure: last month's status sitting in a file as if it were true today

[Krunzar @Krunzar]

"19-year-old turns $68 into $750,000 with a Claude Code bot." I tried to verify this one. here are the 5 checks I ran, and you can run them on any story like it

I found no source for the student, the bot or the account. what I did find is a near-identical story ($0.90 to $408K in 48 hours) that https://t.co/gVXfZ68yki took apart. their estimate: about 1,635 trades, with late trades needing counterparties holding $50K to $400K of liquidity that the market likely doesn't have

the 5 checks:
→ do the math. $68 to $750,000 is about 11,000x, and one night at $6,732 is already 99x
→ look for a name, a handle or a public account you can open
→ ask where the liquidity comes from. a bot can't fill an order bigger than what the book holds
→ check what the author does next. here the next step was "comment to get the setup"
→ ask for the trade log. a real result comes with one

the technical idea is plausible. a bot compares a Binance price feed with a prediction market and trades the gap. the profit numbers are what fail

stop doing this:
→ sharing a profit screenshot you haven't traced
→ copying a "setup" from someone who won't show a trade log
→ sizing real money on a result you can't reproduce

if you test a bot like this, start with read-only data and paper trades

save this before the next "$68 to $750K" post shows up

[Krunzar @Krunzar]

I stopped asking Claude for a "finished deck"

one prompt gives you 12 slides that look fine and say nothing. so I split the work into 5 small jobs. each job ends with one file, and the next job reads only that file

> audience who.md
it asks who opens this deck, what they already believe, and what they should do after the last slide

> claims claims.md
it lists the 3 things I need to prove. everything else gets cut

> evidence numbers.csv
it pulls the real figures behind each claim and writes the source next to each one

> outline outline.md
one claim per slide, every title written as a full sentence, not a topic

> build deck
only now does it touch design, and it reads the four files above

why it works: when slide 7 is wrong, I fix one line in claims.md and rebuild. no re-prompting, no hunting through slides

the deck also stops drifting. slide 3 and slide 9 can't contradict each other, because both read from the same claims file. I check that file before I read any slide

stop doing this:
→ asking for design before the claims exist
→ pasting numbers in by hand
→ judging a deck by how slide 1 looks
→ treating the first build as final

the first build is a draft. you direct the second pass

save this before your next deck

[Francesco @francescoinweb3]

most people use Claude to write

a smaller group figured out how to turn it into an AI production studio.

dashboards, motion graphics, open-source skills, and a full workflow to create exportable videos. one Claude stack. all in here https://t.co/j4U2cJMJqo