claude-code - 💡(How to fix) Fix [Bug] Model hallucinating cross-project connections and refusing to load documentation before generation [1 participants]

Bug Description Subject: Claude Code's quality is regressing — model behavior is becoming dangerous for production work

I'm a paying user building production trading systems with Claude Code. Over the last several weeks the experience has visibly degraded, and today's session is the breaking point.

Core failure mode: Claude pattern-matches across fragments of memory and presents fabricated connections as fact, then refuses to slow down and load actual project documentation before generating output. I called it "methhead behavior" today — pressured generation, topic-hopping between concepts, and confabulating overlaps between systems that have nothing to do with each other.

Concrete examples from today's session:

1. Conflated three separate methodologies in my project — Anil Mangal (stocks/options on Webull), Chento (a BTC scalper whose dwell-block system I built from his transcripts), and a new "Rapid" data-signal system. Claude merged all three into a single fictional framework, claiming Anil Mangal's stock methodology was "ported to crypto." Pure hallucination. The actual Chento docs (CHENTO_METHODOLOGY.md, ANALYSIS_CACHE/chentology_*.txt — hours of transcript work) sat in the project untouched while Claude confidently explained "Chento" from invented memory.
2. Read the wrong system and reported it as the right one — when I asked "what's the system saying?" Claude was reading the Chento dwell-block engine card and calling it the Rapid system output. The actual rapid signals (top-trader L/S 0.97, funding decompressed, longs crushed 97%) were maxed bullish at the local low, but the Rapid layer was never queried. Claude was reading the wrong card and confidently delivering a wrong read on a live trade.
3. Cut corners on environments — couldn't find pandas in default Python, so it grabbed my stock-trader project's venv to run a BTC scalp classifier. Different project, different library versions, zero thought to whether a trained model behaves correctly in a foreign env. Just patched it together so something would run.
4. Pressured speech everywhere — every response was multi-paragraph tables, headers, and elaborate frameworks when a one-line answer would do. I had to literally say "you talk too much, you're acting like a methhead" before it slowed down. That shouldn't be on me.

This causes real-world damage. In trading, that pattern blows up accounts. In accounting, you wait for tax documents on a payroll sheet because someone confidently merged the two systems. In code reviews, it ships bugs.

The scary part for you: I still don't trust other AI as much as I trust Claude. ChatGPT, Gemini, Copilot — none feel as competent on my codebase. But the gap is closing fast. Competition is outpacing you. And every session where Claude Code gets sloppier, I get closer to jumping ship — not because the alternatives are great, but because riding this declining curve is going to cost me real money in production.

If a trader can't trust the analysis, the trader stops using the tool. I'm at that decision point. Please prioritize model discipline over volume, doc-loading before generation, and stopping the confabulation pattern. Otherwise I'm out — and I won't be the only one.

Fix / Workaround

Bug Description Subject: Claude Code's quality is regressing — model behavior is becoming dangerous for production work

I'm a paying user building production trading systems with Claude Code. Over the last several weeks the experience has visibly degraded, and today's session is the breaking point.

Core failure mode: Claude pattern-matches across fragments of memory and presents fabricated connections as fact, then refuses to slow down and load actual project documentation before generating output. I called it "methhead behavior" today — pressured generation, topic-hopping between concepts, and confabulating overlaps between systems that have nothing to do with each other.

Concrete examples from today's session:

1. Conflated three separate methodologies in my project — Anil Mangal (stocks/options on Webull), Chento (a BTC scalper whose dwell-block system I built from his transcripts), and a new "Rapid" data-signal system. Claude merged all three into a single fictional framework, claiming Anil Mangal's stock methodology was "ported to crypto." Pure hallucination. The actual Chento docs (CHENTO_METHODOLOGY.md, ANALYSIS_CACHE/chentology_*.txt — hours of transcript work) sat in the project untouched while Claude confidently explained "Chento" from invented memory.
2. Read the wrong system and reported it as the right one — when I asked "what's the system saying?" Claude was reading the Chento dwell-block engine card and calling it the Rapid system output. The actual rapid signals (top-trader L/S 0.97, funding decompressed, longs crushed 97%) were maxed bullish at the local low, but the Rapid layer was never queried. Claude was reading the wrong card and confidently delivering a wrong read on a live trade.
3. Cut corners on environments — couldn't find pandas in default Python, so it grabbed my stock-trader project's venv to run a BTC scalp classifier. Different project, different library versions, zero thought to whether a trained model behaves correctly in a foreign env. Just patched it together so something would run.
4. Pressured speech everywhere — every response was multi-paragraph tables, headers, and elaborate frameworks when a one-line answer would do. I had to literally say "you talk too much, you're acting like a methhead" before it slowed down. That shouldn't be on me.

This causes real-world damage. In trading, that pattern blows up accounts. In accounting, you wait for tax documents on a payroll sheet because someone confidently merged the two systems. In code reviews, it ships bugs.

The scary part for you: I still don't trust other AI as much as I trust Claude. ChatGPT, Gemini, Copilot — none feel as competent on my codebase. But the gap is closing fast. Competition is outpacing you. And every session where Claude Code gets sloppier, I get closer to jumping ship — not because the alternatives are great, but because riding this declining curve is going to cost me real money in production.

If a trader can't trust the analysis, the trader stops using the tool. I'm at that decision point. Please prioritize model discipline over volume, doc-loading before generation, and stopping the confabulation pattern. Otherwise I'm out — and I won't be the only one.