Ali

Grok Bot desks, agents & money

[Ali @ali0xbt]

@hasibecimcik fair question. it only enters if the book can fill the whole order before the next tick. no size, no trade. that one rule skipped 803 of the 1,904 gaps yesterday. the ones it took had real buyers on the other side

[Ali @ali0xbt]

this is f*cking embarrassing. my Grok Bot trading desk was burning $612 a week in usage just to watch BTC. this week it made the same trades for $9.40

it used to send every Binance tick to Grok, 86,400 a day, and almost none of them needed a decision. now plain code watches the price and Grok wakes up only when a contract is actually mispriced. yesterday that was 1,904 times

the watching alone was $441 of the bill. the rest came from oversized inputs, the same answers generated twice, and one retry loop that ran all night

20 GitHub repos I used to cut it:

01 ccxt → price feeds and order books from 100+ exchanges in one library, so code watches the market instead of the model
https://t.co/kHPGDb28l2

02 https://t.co/38Mb1DSnJV → watches pages and APIs and pings only when something actually changed, so the bot stops rereading the same page every hour
https://t.co/YtcDbhsXA2

03 watchdog → fires an event the moment a file or folder changes, instead of a script polling it on a timer
https://t.co/DGz1NrcemU

04 jq → cut a 2 MB API response down to the 4 fields the bot needs before it reads a single token
https://t.co/TNbiA77EdS

05 ripgrep → search a repo or a log folder in milliseconds and hand the bot 30 matching lines, not 30 files
https://t.co/xGQmh4fVgC

06 ast-grep → structural code search, so the bot gets the one function it has to change instead of the whole module
https://t.co/oCULr7Y0mX

07 MarkItDown → turns PDFs, Word files, spreadsheets and HTML into plain Markdown before they hit the context
https://t.co/rT4ZGxPxOV

08 Trafilatura → pulls the main text out of a web page and drops the menus, footers and cookie banners
https://t.co/6gljGm0hb5

09 Repomix → packs a repo into one file with a token count per file, and can compress it to signatures when the bot only needs the shape
https://t.co/2qSa6LII40

10 DiskCache → memoizes function results on disk. same input, same answer, no second call
https://t.co/QpD3vJBeW4

11 RedisVL → semantic cache for model calls. a question that means the same as yesterday's gets yesterday's answer
https://t.co/kNHRE86w17

12 Ollama → runs a small local model for the yes/no checks, so the paid model only sees the hard ones
https://t.co/oaLp2dg3rU

13 DSPy → tunes a prompt against your own examples until a smaller, cheaper model passes the same test
https://t.co/djs4fY7TSI

14 smolagents → the agent writes one block of code that chains several tool calls, instead of one model round trip per tool
https://t.co/isVpwRvAU3

15 Tenacity → retries with backoff and a hard stop after N attempts, so one dead endpoint can't loop all night
https://t.co/IzFILAn2ax

16 Temporal → durable workflows. a crashed run picks up at the failed step and the finished steps don't run again
https://t.co/93ihbOm8MW

17 tiktoken → estimates a prompt's token count before you send it, so anything over budget gets cut first
https://t.co/kLlSk9Nzwv

18 Langfuse → traces with cost per step, so you can see which part of the loop eats the money
https://t.co/qV8l5TMycC

19 Helicone → a proxy that logs every call with its cost and can answer repeated requests from cache
https://t.co/nI1gDmAx2g

20 promptfoo → runs the old prompt and the new one on the same cases and compares cost and output before you switch
https://t.co/BoBtK1oOGz

the order I'd go in: measure first, then move the watching to code, send slices instead of files, cache what repeats, and cap every retry

a bot that runs 24/7 spends most of its day looking at ticks where nothing happens. those hours should cost zero tokens

bookmark this before your weekly limit runs out on a Tuesday

[Gresham @Greshamtrade]

https://t.co/x9JOa2faWx

[Ali @ali0xbt]

@yurshevv 803 of those skips were thin books. the gap was real, there was just nobody on the other side to fill it

[Ali @ali0xbt]

I gave Grok Bot the same $68 that kid started with. 24 hours later the wallet was at $5,247, and it walked away from 5 of every 6 price errors it found.

the skipping is what kept the red trades to 23.

same idea as his. Grok reads BTC on Binance every second and trades the 15-minute up/down contracts that are still priced on old information. BTC sits $140 above the window open with 40 seconds left, "up" still sells at 71 cents, the bot buys it and collects $1 at the close.

a gap like that lives about 2 seconds, so every entry clears three checks. not enough size on the other side to fill the whole order, skipped. the fill would land after the next Binance tick, skipped. the window already repriced twice, skipped.

1,904 gaps found. 312 taken. 1,592 skipped.

the five that paid most:

> 02:15, BTC 15m up, +$96, more than doubled the wallet
> 07:45, BTC 15m down, +$318
> 11:30, BTC 15m up, +$571
> 15:00, BTC 15m down, +$880
> 21:45, BTC 15m up, +$1,214, the biggest of the day

312 trades, 289 green. all 23 red ones filled a beat too late, after the price had already caught up.

every bot reading Binance finds these errors, his and mine included. the money is in knowing which ones you can still get filled on.

$68 to $5,247 in 24 hours, and 1,592 trades it said no to.

[Paul Sims @SimslearnAi]

A 19-year-old Japanese student built a trading bot with Claude Code in 2 days.

Used his iPad as a second monitor.

First night: $6,732 profit.
Starting capital: $68.

Total profit so far: $750,000.

Here's how it works:

The bot scans over 50 markets simultaneously.

Syncs live BTC data from Binance every second.

Spots price errors before humans even notice.

The edge is pure speed + pattern recognition.

While traders stare at charts trying to predict the next move, his bot is already executing on mispricing across dozens of markets.

No guessing.
No emotions.
No hesitation.

Just Claude Code logic finding gaps that close in seconds.

He built the entire system in 48 hours:

→ Claude Code handles the trading logic
→ Binance API feeds real-time BTC data
→ iPad displays multi-market monitoring
→ Executes trades when arbitrage windows open

The system runs 24/7.

Every price dislocation = profit opportunity.

Most people are still trading manually, refreshing charts, second-guessing entries.

Meanwhile this 19-year-old engineering student turned $68 into $750K by letting Claude Code do what humans can't: process 50 markets instantly and execute without fear.

Why are people still trading manually?

💡I'm giving away the exact Claude Code setup for free.

24 hours only.

To get it: 👇

1️⃣ Comment "Claude"
2️⃣ Like Comment and Repost

Follow @SimslearnAi

I'll DM you the complete setup.

[Ali @ali0xbt]

@Solomonrojie the "conditions that would make it outdated" field is the smart part. most notes say what's true, almost none say when it stops being true

[Ali @ali0xbt]

@hammertime_one the stop sign example is real research from 2017. a few black and white stickers got a classifier to read a stop sign as speed limit 45

[Ali @ali0xbt]

@yurshevv "the expensive one only signs off" is the most accurate senior engineer job description i've read all year

[Ali @ali0xbt]

@monokern the insane part to me is the 3 hours in the middle where it did nothing. that flat line is the bit I’d copy

[Ali @ali0xbt]

I gave Grok Bot $52 and one extra rule. 24 hours later the wallet was at $6,914, and for 3 of those hours it didn’t touch a single coin.

that flat stretch on the chart is the best call it made all day.

Grok runs six agents on Solana memecoins, each on its own slice of new pairs with its own size limit. every token has to clear the same checks before an entry. dev holding over 15% of supply, skipped. liquidity unlocked, skipped. mint still live, skipped. and the extra rule: if volume jumps and the number of real buyers doesn’t, nobody buys anything.

at 02:10 volume on new pairs tripled in 20 minutes. unique buyers went from 1,940 to 2,010. the same few hundred wallets were trading with each other to look busy.

Grok paused all six agents until 05:10.

by 09:00, 11 of the 14 tokens that led volume in that window were down 90% or more. it lost $0 on every one of them.

214 trades, 176 green.

the five that paid most:

01:21, GLIMP +$131, more than doubled the wallet
06:44, BRISK +$388
10:02, MOLTO +$612
14:37, PADDO +$904
19:15, VELLO +$1,288, the biggest of the day

four of the five came after the pause, once the buyers were real again.

volume is the number every memecoin bot watches. this one checks who’s behind it first.

$52 to $6,914 in 24 hours, and the part I’d frame is 3 hours of doing nothing.

[savip. @savipww]

https://t.co/YYQoOfAkKF

[Ali @ali0xbt]

@theSethian honestly the crazy part is how boring days 1-4 looked. same trades, tiny balance, nobody would’ve screenshotted it

[Ali @ali0xbt]

@lostinazeroy exchanges solved this years ago with scoped API keys. trade-only, withdrawals off, locked to one IP. my desk never touches a login page, so there's nothing for a site to mistake for a bot

[Ali @ali0xbt]

@Atlas_fomo start with skill 3. the other six all spend or chase money, so you want to know payroll clears before an agent drafts a single purchase order

[Ali @ali0xbt]

@Beaver_0x been running this for weeks. what broke it for me was two pages saying opposite things, an old rule and a newer one, and Claude followed whichever it read first. one line in CLAUDE.md fixed it: if two files disagree, the newer decision wins and the old one gets fixed the same day

[Ali @ali0xbt]

@scrblanc you put it better than my post did. picking a side is a bet, quoting both is rent. the 172 fills that didn’t keep the spread were the cost of being there

[Ali @ali0xbt]

F*CKING PHONE BUZZED 212 TIMES TODAY

MY GROK BOT SAT ON BOTH SIDES OF THE ORDER BOOK AND GOT PAID EVERY TIME SOMEONE ELSE TRADED.

14 HOURS, 1,906 FILLS, $400 TO $3,118.42.

In 2018 Europe's market regulator reported that 74-89% of retail CFD accounts lose money. Since then brokers there have printed their own number on every ad. Somebody sits on the other side of those trades, and it usually isn't another trader guessing.

So my bot stopped guessing. It parks a buy a little under the price and a sell a little over it. When both fill, it keeps the gap.

→ quoter puts up both sides on 4 thin pairs, never picks one
→ risk watches who's hitting us, never waits
→ exec pulls and reposts in under a second, never holds a bag overnight
→ me, one y at 8 AM

The first hours were crumbs. JUP +$0.84. BONK +$1.12. One −$2.30 when a buyer took the whole SEI ask and kept going.

At 2:29 PM Grok pulled every quote on SOL. The same taker had hit our bid 3 times in 9 seconds. Forty seconds later SOL dropped 4.1%. Those resting quotes would have handed back $611.

It came back in at 3:02 and the afternoon paid. $186.40 on one WIF wick at 4:12, bought at the low and sold 3 minutes later. $203.60 from SOL fills in the last hour.

10 AM: $512. Noon: $803. 4 PM: $1,690. 10 PM: $3,118.42.

1,906 fills. 1,734 kept the spread. 0 calls on where the price goes.

One loud Friday isn't a strategy. Thin books pay on days like this and go quiet the rest of the week.

The last buzz came at 10:02 PM. I turned the phone face down. It was still quoting.

[boan @0xboan]

https://t.co/j60ZLf72pa

[Ali @ali0xbt]

@leiroops by buzz 50 i stopped reading them and just listened. the silence at 2:29 was the scary part

[Ali @ali0xbt]

@Klonzu overnight is where makers blow up. one big taker at 3am on a thin book and the whole week’s spread is gone

[Ali @ali0xbt]

@fayzonnn same rule saved my desk. a trade that fails the risk check twice is dead for the day, no third look. before that the bot kept re-pricing the same bad spread every 30 seconds and paying API fees to tell me no

[Ali @ali0xbt]

@yurshevv that 77 vs 74 is from DeepSeek's own chart. i'd replay my last 50 real tasks on both before moving anything, a 3-point lead on someone else's benchmark can disappear on your own repo

[Ali @ali0xbt]

@Asteri_eth i'd add one check before any of this: search X for the clip first. reuploads of clips that already went viral got me around 1K views no matter how clean the cut was

[Ali @ali0xbt]

@0xboan funny part is that one SOL pull saved more than the first 4 hours made