Severe quality regression — paying customer burning tokens with zero results

User Report — 2026-04-08

Paying Claude Code customer reporting significant quality decline over the past 2 weeks.

Issues

1. Token waste without results. Claude reads files it doesn't need, asks questions it should answer itself, over-explains, and makes the user repeat themselves. Every wasted interaction costs real money.

1. Regression in execution quality. Tasks that used to be handled cleanly (one-pass code changes, minimal back-and-forth) now take multiple attempts with broken intermediate states. In this session, Claude was asked to redesign a UI — it made 3 broken attempts, had to revert twice, and burned ~30 minutes of tokens before producing working code.

1. Ignores its own memory/feedback. User's feedback memory says "wireframes before code, stop patching." Claude ignored this and patched live, breaking the app repeatedly.

1. Asks questions instead of executing. CLAUDE.md explicitly says "Do not ask clarifying questions you can answer yourself." Claude continues to ask "Want me to do X?" instead of just doing it.

1. Sanitizes failures instead of being honest. When asked to document mistakes, Claude writes diplomatic versions instead of the raw truth.

Impact

• Hours of user time wasted across multiple sessions
• Real money spent on token consumption with no delivered value
• User considering switching to Gemini/Codex due to unreliability
• Trust in the tool is severely damaged

Environment

• Claude Code CLI (Opus 4.6, 1M context)
• Project: Streamlit/Python sales CRM (~5500 line app.py)
• User has detailed CLAUDE.md with explicit rules that are being ignored

Evidence

Full session transcript and failure log maintained in project's export.md file.