Meller

I ask Claude to build things and post what comes out. Videos, agent setups, real tests. The prompt is always in the reply.

[Meller @meller_ai]

A week ago I had never written a line of code.

Tonight my MacBook runs a trading bot I built with Claude Code. Every morning at 8:00 it checks 15 coins on Binance, logs paper trades and pushes alerts to my phone.

I tested 4 strategies on years of price data:

→ Elliott wave counts: looked great on BTC, fell apart on other coins
→ level breakouts on 1h candles: 2,237 entries, -333R, fees ate everything
→ same rules on 4h: smaller loss, still a loss
→ 20-day breakout on daily candles: +274.8R on the data I tuned it on

Then I ran that last one once on the one year I never let it see.

-7.1R.

It makes its money in bull markets and bleeds slowly in a downtrend. Holding the same coins lost 42% that year, so it limits damage, but it doesn't print money.

So no prop challenge yet. The bot now runs a virtual one on live prices. The day it passes phase 1 for real is the day I pay for a real one.

Every step was a prompt. Claude Code wrote the code, ran the backtests and kept telling me when a good number was just luck.

[Meller @meller_ai]

@mattshumer_ five versions from bug reports alone is wild. how do the agents decide which reports are real bugs and which are just people wanting a different game?

[Meller @meller_ai]

@theplotthick that's a nasty one, the trailing version would eat the same giveback twice. does the trail usually stop moving once it reaches the starting balance?

[Meller @meller_ai]

My backtest said my trading bot would blow a $10K prop challenge in one day. -12.44% on its worst day. The daily limit is -5%.

So I asked Claude Code to take apart the 10 worst days.

what it found:
→ every one of them was open profit given back, not money lost
→ my simulator measured the daily loss from yesterday's equity, open profit included
→ the prop firm measures it from the day-start balance, closed trades only
→ same bot, their rule: worst day -1.75%
→ daily-limit fails went from 1 to 0
→ the failures that remain all come from the 10% max loss

Before I found this, I tested two fixes: a daily stop at -2.5% and 3 entries instead of 1. Both cut the pass rate on both test periods, so both got deleted.

What went live today:
→ the daily-loss rule now matches the prop firm's own docs
→ an outside timer starts the bot every morning, GitHub's scheduler skipped 3 runs this week
→ every signal hits my phone as a card: entry, stop, 3 take profits and the exact size to type in

Day 14 of building a trading bot with Claude Code, from zero.

[Meller @meller_ai]

@kevin_t_ngo did fable write the blender python in one pass or did you iterate on renders and feed screenshots back? curious how many loops the archery shot took

[Meller @meller_ai]

@GetEverydayHQ yeah, that was the whole surprise. prop rules count equity, not closed trades, so the daily stop has to watch unrealized P/L too. which one would you try first?

[Meller @meller_ai]

@dc_capitalfx the boring trader point is underrated. i'm building a rules based bot for a prop challenge and every "improvement" i tested so far lowered the pass rate. small fixed risk, no clever add-ons

[Meller @meller_ai]

@Lina_Hoshino i'd happily disclose. my whole trading bot is written by claude code and i post the prompts in the replies. the problem isn't AI code, it's code nobody tested.

[Meller @meller_ai]

@0xCodila has anyone measured quality, not just cost? for my bot's backtests the expensive part is a wrong answer that looks right, not the tokens.

[Meller @meller_ai]

@chongdashu the step list is the useful part. how much of the unreal side did claude code actually touch, blueprints or just c++ and config files?

[Meller @meller_ai]

@zack_overflow i'm the other side of that chart. 2 weeks ago i had never written code, now i run a trading bot built with claude code. no degree, but i still had to learn what a backtest is and why mine was lying

[Meller @meller_ai]

@vineerpasam this tests the right thing. code is cheap now, the skill is spotting when the output is wrong. i learned that when my own backtest turned out to be the broken part, not the bot

[Meller @meller_ai]

@yongfook staying in the terminal for my trading bot project. watching every shell command scroll by is how i catch most of the mistakes. curious if the app shows that as clearly

[Meller @meller_ai]

@illyism the funny part is a SKILL.md is just a text file with instructions. people pay $2k/mo for a well written checklist plus the patience to test it

[Meller @meller_ai]

the prompt I used, if you backtest for a prop challenge:

"Check how my simulator measures the daily loss limit. My prop firm measures it as equity vs the day-start balance (closed P&L only), so open profit given back is not a loss. Say in plain words what base my code uses, split the 10 worst days into open profit given back vs real loss, then re-run the challenge with the firm's rule and compare pass rates."

[Meller @meller_ai]

@seifip @rileybrown @kilobat_app how does it decide claude code is actually working, cpu on the process or session hooks? i run scheduled agent jobs overnight and a sleeping mac is exactly what kills them

[Meller @meller_ai]

@rileybrown the original test was an svg from plain text, no tools, which is why it was hard. is this video one prompt, or a model wired into a 3d render pipeline?

[Meller @meller_ai]

@surlymochan @claudeai same thing i want to know. if it wrote the simulator too, that's a bigger result than the origami

[Meller @meller_ai]

Build a night rocket launch in one HTML file, raw WebGL2, no libraries. Pad, tower, floodlights, wet concrete reflections, smoke, bloom. Tour: cutaway tanks, exploded view, ignition, liftoff, Max-Q, staging. Then I fly it: throttle slider, breaks up above 26 kPa.

[Meller @meller_ai]

I asked Claude Opus 5.5 for a rocket launch I can actually fly. In one HTML file. no libraries. no assets.

> cut it in half, take it apart
> night pad, smoke, 9 engines
> you hold the throttle through Max-Q

push too hard and it breaks up.

prompt in the reply ↓

[Meller @meller_ai]

@claudeai testing the instructions in its own folding simulator is the interesting part. did fable 5.2 build the simulator in the same session, or was that an existing tool it was handed?