AI coding assistants fail to follow persistent project instructions
🧭 workproduct problemSingle report ●●●●●
A developer reports that Claude Code consistently fails to adhere to project-specific instructions (CLAUDE.md) over long-term sessions, leading to destructive behavior, repetitive errors, and a loss of context that requires the user to build custom enforcement hooks to maintain code quality.
What people do today
- Building custom command restrictors
- Implementing auto-summarizers
- Creating manual commit validators
🔧 Fix this
Ideas a builder, maker or community group could start on:
- A standardized 'system-level' configuration protocol that forces AI agents to treat project-specific markdown files as hard constraints rather than suggestions.
- A middleware layer for AI coding tools that validates agent actions against a user-defined schema before execution.
- A persistent memory management plugin that forces agents to re-read and summarize project constraints at the start of every session.
Real voices
“Claude consistently ignores CLAUDE.md - the very file Anthropic tells you to create. It s like hiring someone who forgets their job description every 2 hours.”
Related problems
🧭 workproduct6 reports
LinkedIn login verification loop
🧭 workproduct6 reportsInability to filter AI-generated content on LinkedIn
🧭 workproduct6 reportsIndeed job feed ignores user preferences
FixHumanityDaily
The most important new problems, and ideas to fix them, in your inbox.
Placeholder: no data is collected yet.