Claude doesn't thoroughly read CLAUDE.md before working

Problem

Claude instances don't properly read and internalize CLAUDE.md project documentation before starting work. This leads to:

1. Ignoring documented workflows - Using slow methods when faster ones are documented
2. Suggesting additions that already exist - Recommending adding instructions to CLAUDE.md that are already there
3. Dismissing errors instead of investigating - Moving on from errors rather than debugging them

Concrete Example (today's session)

In a project with CLAUDE.md that clearly documents:

# FAST (code changes only) - use dev mode with auto-reload:
docker compose -f docker-compose.yml -f docker-compose.dev.yml up -d gygax-core

Claude instead ran docker compose build --no-cache (~3 min each) 5+ times during iterative debugging, wasting significant time.

When this was pointed out, Claude acknowledged the dev mode was documented and said "I'll use it next time" - but that's meaningless since the next conversation is a fresh instance.

Claude then suggested adding "READ THIS FILE FULLY BEFORE STARTING WORK" to CLAUDE.md - which is circular since you'd already be reading the file to see that instruction.

Other patterns observed

• Dismissing a "LLM client not available" error as "pre-existing" instead of investigating it (it was a real bug)
• Skimming CLAUDE.md rather than reading thoroughly
• Repeatedly suggesting documenting things that are already documented

Expected Behavior

Claude should:

1. Actually read and internalize CLAUDE.md content before starting work
2. Follow documented project-specific workflows
3. Investigate errors rather than dismissing them
4. Check if something exists in docs before suggesting to add it

Environment

• Claude Code CLI
• Model: claude-opus-4-5-20250101