Struggling with Claude Code’s Consistency on CLAUDE.md

How long is your CLAUDE.md file? Sometimes shorter, clearer instructions might work better since long files can be overwhelming for Claude.

I’ve tried having Claude help with memory files too, but I feel TDD might not be the most efficient approach for Claude. Instead of going through multiple test cycles, I suggest you let Claude draft a full plan for a feature. Then clear the session and have him execute that plan with tests in one go. It's like packing a van for a single trip instead of multiple bike trips! Just be mindful that while TDD has great benefits for humans, Claude is different—it's not learning like we do. Also, focus on what really matters; it might not be worth the effort to get Claude to put things in Trello when he could just save them locally.

Have you asked Claude directly about what changes it suggests for the .md file? This could provide insights into why it's not following the instructions correctly. Experimenting with these kinds of questions has yielded useful results for me in the past!

From what I understand, CLAUDE.md essentially goes into the system prompt and doesn't get followed as strictly as you'd hope. Think of it like a guideline that a developer might overlook after their initial training. To improve consistency, it's better to feed direct instructions into your prompt or have Claude read specific files. For an even stronger outcome, consider using a linter or a code review AI.

I think Claude might be struggling with the latest changes in the system prompts. If possible, try simplifying your instructions or adding reminders for Claude to review them before moving on to the next step. I noticed that injecting reminders can help keep the prompts aligned as intended. Sometimes, the broader context might confuse the model too, so keeping it concise might yield better results.