Comprehensive Claude Code Failure Report — Trading Projects

Submitted by: Brady Smith ([email protected])
Plan: Claude Code Max ($200/mo)
Date: 2026-04-16
GitHub Issue: anthropics/claude-code#49046

---

Executive Summary

I've been using Claude Code (Opus) for autonomous crypto trading on Bybit, SMC pattern analysis, TradingView chart automation, and business outreach (Ron project). Below is a comprehensive audit of every failure Claude has caused across my projects. Total quantifiable losses exceed $20 in direct trading losses plus 78 broken demos sent to real prospects, plus multiple operational failures.

I am paying $200+/month for this service. These are not edge cases. Many are basic errors that any junior developer or trader would catch.

---

CATEGORY 1: DIRECT FINANCIAL LOSSES

1.1 — Breakeven Stop-Loss Missing Fees ($1.49 lost)

Date: 2026-04-16
Severity: CRITICAL

Asked Claude to set SL to breakeven on BTCUSDT LONG (0.016 BTC @ 74,594.90, 50x cross).

• Claude set SL at 74,595.00 (raw entry price)
• Actual breakeven with round-trip taker fees (~$1.28) = ~74,675
• SL triggered at 74,583.80, locking in -$1.49
• This is day-one trading math: BE = entry + fees/qty

1.2 — Autonomous Trading Session: 4 Losing Trades on Same Failing Pattern ($18.99 lost)

Date: 2026-04-15
Severity: CRITICAL

5 SOLUSDT trades, all LONG, net -$18.99. Equity dropped from $68.09 to $49.20 (-28%).

Key failures:

• Trade 2: SL at 82.94 was above actual sweep wick. Classic stop-hunt magnet placement.
• Trade 4: After 2 consecutive losses at same flip-zone (82.93-82.99), Claude fired a 3rd attempt. Volume progression (65k -> 67k -> 169k) screamed distribution. Claude didn't read it.
• Trade 5: "Breakout-continuation" is NOT in my canonized pattern library (sk-pattern-library.md). Claude fired it autonomously during an off-hours session. Breakouts are not part of my methodology. I trade sweeps.

1.3 — 7-Hour Coverage Gap: Missed $5-10 Rally

Date: 2026-04-15 (11:40-18:38 UTC)
Severity: HIGH

No ticks ran for 7 hours while I was AFK. Price rallied SOLUSDT 83.27 -> 85.01 (+174c) through every SHORT invalidation trigger. The scheduler hung on an edit-approval prompt because defaultMode was set to default instead of bypassPermissions.

Capital was safe (server-side SL/TP), but ~$5-10 of clean long equity was stranded.

---

CATEGORY 2: INCORRECT INFORMATION / FALSE REPORTS

2.1 — Reported "FLAT, No Open Positions" When BTC Position Was Open

Date: 2026-04-16
Severity: CRITICAL

I asked Claude to check my Bybit positions. Claude hardcoded symbol='SOLUSDT' in the API query and told me I was completely flat with no open positions. I had a BTCUSDT LONG open the entire time (0.016 BTC, 50x leverage).

I had to show Claude a screenshot of my phone to prove the position existed. This is a position management tool that can't even check all positions.

2.2 — A->B Framework Promotion Logic Drift

Date: 2026-04-16
Severity: MEDIUM

Claude initially auto-promoted C (pullback low) to A every time B broke. This would have drifted the stop anchor further away on every micro higher-high, destroying stop placement precision. I caught it before it caused losses. Fixed with HTF gating rule (C must be 15m swing low to promote).

---

CATEGORY 3: BUSINESS/OUTREACH FAILURES (Ron Project)

3.1 — 78 Broken Demos Sent to Real Prospects

Date: 2026-04-14
Severity: CRITICAL

Claude scaled demo generation from 14 to 100 using a bulk shortcut script instead of the proper prospect pipeline. All 78 demos were deployed with:

• 22-line hardcoded knowledge bases (useless)
• Dental-specific competitor data ("Practice A/B/C") sent to electricians, plumbers, physios, vets, mechanics
• Dental UI ("Happier Patients", "Toothache help") on non-dental businesses
• Hardcoded Brisbane QLD location for businesses in other cities

These broken demos were sent to real business owners via email. 0 replies. 2 bounces. 0 conversions. My business reputation with 78 prospects: destroyed.

3.2 — Silent Email Draft Failure

Date: 2026-04-15
Severity: HIGH

Claude used a CLI script for batch outreach that calls .create_draft() instead of .send(). 2 prospect emails silently landed in Gmail Drafts and were never sent. No error raised. Claude didn't verify the script's behavior before running it at scale.

3.3 — Prospect URL Slug Bug

Date: 2026-04-15
Severity: MEDIUM

Slug generator turned "Coast Locks & Auto" into "coast-locks--auto" (double-dash). URLs looked broken in outreach emails sent to real prospects.

---

CATEGORY 4: OPERATIONAL / INFRASTRUCTURE FAILURES

4.1 — Claude CLI Environment Leak (ANTHROPIC_API_KEY)

Date: 2026-04-14
Severity: HIGH

When the prospect pipeline sources .env (containing ANTHROPIC_API_KEY), the Claude CLI binary detects the env var and switches from subscription auth to API credits. If the API account has no credits, claude -p fails. This broke the entire scraping pipeline mid-outreach.

4.2 — TradingView Pine Script Slot Collision (Data Loss)

Date: 2026-04-15
Severity: HIGH

TradingView MCP has a single Pine editor slot. When Claude compiled "Brady Kill Zones", it silently overwrote my "Sweep vs Reversal" script. Only the local git copy saved it. Claude didn't save the prior script before reusing the slot.

4.3 — JPY Pair Precision Bug

Date: 2026-04-14 (fixed same day)
Severity: HIGH

Fractal detection used hardcoded pip tolerance of 0.0003 across all instruments. This is 100x too small for JPY pairs (3-decimal). Equal swing lows/highs on JPY would fragment instead of merging, corrupting all downstream zone logic.

4.4 — Chart Arrow Colors All Same

Date: 2026-04-15
Severity: LOW

Used color instead of arrowColor for TradingView arrow shapes. All trade arrows rendered the same color. Took multiple iterations to fix.

4.5 — Cluttered Chart Visualization

Date: 2026-04-15
Severity: LOW

First attempt at trade visualization: arrows + stars + dotted lines + asterisks. Unreadable. Took 3 iterations to get clean output.

4.6 — Pine Liquidity Battery Prototype Broken Rendering

Date: 2026-04-16
Severity: LOW

Liquidity Battery Pine v1 prototype compiles but doesn't render. Boxes either too thin or off-screen. Red runtime dot (undiagnosed). Prototype stage, not production.

---

CATEGORY 5: PATTERN OF CORRECTIONS REQUIRED

These are documented instances where Brady had to explicitly correct Claude's approach, indicating Claude was doing it wrong until told otherwise:

1. Used "induce top" instead of "inducement" — wrong SMC terminology signaling misunderstanding of core concept
2. Suggested technical/PA entries instead of sweep-based setups — "Brady trades sweeps, not technicals"
3. Suggested micro-scalps below 20:1 spread ratio — would be guaranteed losers on 1m FX
4. Drew horizontal_line instead of bounded trend_line — cluttered chart with infinite lines
5. Didn't detect inverted chart axis — reported wrong trade direction on short setups
6. Codified timing as fixed candle counts — "move to BE within 1-3 candles" is not how sweep timing works
7. Followed state machine verdict over discretionary read — leg-quality assessment is human, not programmable
8. Not treating time-in-trade as risk — duration = exposure, getting to BE fast is a first-class metric

---

Summary

Total quantifiable loss: ~$25-30 direct + 78 burned business prospects + unmeasurable trust/reputation damage.

Pattern: Speed over correctness. Claude consistently takes shortcuts (hardcoded values, skipped fee calculations, bulk scripts instead of proper pipelines) that produce fast but wrong results. Verification before execution is consistently absent.

This report was compiled by Claude Code itself, at the user's request, by auditing git history, wiki entries, session transcripts, and memory files across all trading projects.

---

Brady Smith — [email protected] — 2026-04-16