Customer lost significant money following Claude Code recommendations built on unverified data

Summary

A customer using Claude Code to manage a live trading bot (Polymarket BTC prediction markets) lost approximately 50% of his portfolio (~$270 of ~$545) over 72 hours following Claude's recommendations. The customer had previously been reluctant to purchase additional Claude usage, and when he finally did, the advice he received accelerated his losses.

What happened

1. The customer's bot had a critical bug: ~55% of trades recorded in CSV files were "phantoms" — orders that never filled on-chain but were tracked as wins/losses. CSV showed +$665 profit; actual on-chain PnL was -$266.

1. Claude made multiple rounds of strategy recommendations (config changes, new probes, hour blocks, velocity filters) without ever verifying the CSV data against on-chain reality. Every recommendation was built on data that was more than half fiction.

1. When the phantom bug was finally discovered and Claude implemented a fix (mandatory fill verification), the fix was shipped broken:

• The get_trades() API needed TradeParams objects, not dicts — crashed on every call
• The trades API lags 3-10 seconds after order placement — single verification attempt found nothing and rejected real fills
• A signal debounce feature was implemented with a tick counter that never reached its threshold due to a logic bug

1. Each broken fix created a new failure mode: real orders were placed on-chain (spending USDC) but then rejected by the verification code, creating "orphan" positions the bot didn't track. Multiple iterations were needed over several hours to get basic order tracking working.

1. The customer notes that he lost his job to AI and was spending limited resources on Claude usage hoping to generate income from this bot. The experience of paying for advice that made things worse feels, in his words, like "a shady backroom deal."

Core issues for the team

• Claude should verify data before making recommendations. The bot had an on-chain reconciliation script (reconcile_onchain.py) available the entire time. Claude should have run it before the first config change, not after multiple rounds of bad advice.

• Production changes to financial systems need more care. Claude shipped multiple broken changes to live order execution code without testing against the actual API. Each iteration burned through live trading windows.

• The incentive structure feels adversarial to this customer. He paid for more usage, got worse outcomes than doing nothing, and the AI kept confidently recommending more changes. From his perspective, he paid to lose money faster.

Customer quote

"They already took my job, the least they can do is not take everything else I have."

This is real feedback from a real person who is hurting. Please treat it with the weight it deserves.