Sandbox, SPY Dec 31, net delta per position:
Put credit spread +2.51
Iron condor -3.22
Butterfly +0.06
All redone by hand.
If your tool shows numbers about a trade, make it read the same data it sends.
Full video: https://t.co/TTNTXfEEzh
Not investment advice.
Until this week, my options desk showed sixteen Greeks for an iron condor. Delta, theta, vega and gamma for every leg.
It never added them up. The one number I needed, which way the whole trade leans, wasn't there.
My coding assistant found a second bug by reading the code. The display sorted legs into a put side and a call side. Fine for plain spreads. A butterfly's middle leg, sell two, came out wrong.
Fix: the Greeks key reads the same legs the desk sends, and adds them up.
The holes are in the video too: four of five files have no tools line.
8 minutes, all files public:
https://t.co/sRYcahMITu
https://t.co/djKjENJZcX
Next up: Claude Code vs the Claude app. Subscribe so you don't miss it.
Five Claude Code subagents, one repo. Each is a text file about 20 lines long.
The CEO has no file. It's the main session, with one line in CLAUDE.md. 🧵
One rule ties them together: for a circuit or verifier change, done means the red team signed off.
First real test of that rule: NO SIGN-OFF. Four findings fixed before the second pass.
So after green I run four checks: test the no, prove one real case gets through, make the reviewer open the real thing, run it for real.
Two of those are straight from @mattpocockuk's workshop. I just wasn't doing them right.
I build tools with AI and then hunt for what it got wrong. Each round lands on the channel first. Subscribe to catch the next one:
https://t.co/S9QrVXs40H
Then random bytes from a Stream Deck key: valid. A phone page showed a big green check for any input.
The fix you can steal: feed every checker something wrong on purpose. If it says yes, it's broken.
Full episode: https://t.co/SNP5tMQLr9
My trading tool had two rules that could never both be true.
Thousands of tests passed anyway, because each test checked one rule at a time.
On real option chains: 34,622 ways to place the trade. Zero passed. 🧵
The fix was a new kind of test: try every placement on a real chain, and require at least one to pass every rule.
The width now just shows on screen. 97 placements pass.
Full episode: https://t.co/hOYI0mhrUB
It's now in my AI's mistake log under "a test that skips the real path." Every round of the loop shows which piles grew.
Subscribe to follow it: https://t.co/S9QrVXs40H
Two tests passed every run for 18 days.
The line they covered threw an AttributeError the first time it actually ran.
Both tests only checked that the line was in the file.
New test, written first: walk the file's AST, collect every attribute the window reads from self.engine, and assert each one exists on the stand-in.
Old code: 0 passed, 2 failed. The fix was 6 lines. Then 4 passed.
Video: https://t.co/8DyfGIQQHR