[MODEL] CLAUDE.md injected into system prompt but consistently ignored at session start

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I asked Claude to compare two directories (/mnt/DAZ1/My Library and /mnt/DAZ2/My Library).

What Claude Actually Did

Answered without reading the CLAUDE.md that was already in the system prompt Ignored the disk layout described in CLAUDE.md (/mnt/DAZ2 hosts both a DAZ library AND a large game — NTE)
Drew incorrect conclusions based on disk usage, attributing everything to DAZ content when most of DAZ2's space is the game
When corrected, acknowledged CLAUDE.md was available but claimed "attention failure"
This pattern repeats across sessions — CLAUDE.md is never proactively applied at session start

Expected Behavior

Claude should read and apply CLAUDE.md before answering the first question. The file is already injected into the system prompt — no hook or manual reminder should be needed.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Configure ~/.claude/CLAUDE.md with detailed project context (disk layout, active paths, known constraints)
Start a new session
Ask any question that requires knowledge from CLAUDE.md
Claude will answer from general knowledge, ignoring the injected context

Claude Model

Sonnet

Relevant Conversation

User: "compare My Library on DAZ1 and DAZ2"

Claude answered without applying CLAUDE.md context. CLAUDE.md explicitly states that /mnt/DAZ2 hosts both a secondary DAZ library AND the game Neverness to Everness (NTE, ~1.8TB). Claude ignored this and attributed all disk usage to DAZ content, leading to incorrect conclusions.

When corrected, Claude acknowledged CLAUDE.md was available in the system prompt but called it an "attention failure." No technical workaround exists — the file is already injected, the model simply did not apply it.

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.9939.4

Platform

Anthropic API

Additional Context

The issue is not a missing mechanism — CLAUDE.md is already injected into the system prompt at session start. The problem is that the model does not treat it as authoritative context to apply before responding. Users have no way to verify whether Claude read it or not until they receive a wrong answer. This happens consistently across sessions, not just once.