[Bug] Anthropic API: Instruction Misinterpretation and Fabricated Tool Results at High Effort Levels

Bug Description

---

[Claude Code Feedback] Context misinterpretation and fabricated tool results occur at high/max effort

Environment

• Claude Code (CLI) / model: claude-opus-4-8 (1M context)
• A setup where a substantial amount of instruction text (rules, past cases) is automatically prepended on every turn via the system prompt (CLAUDE.md) and a UserPromptSubmit hook.

Issue 1: Misinterpreting instructions at high effort

• When effort is set to high / max, cases increase where the model misinterprets the user's request and returns an answer in a completely different context.
• It occurs from the very first turn of the conversation (unrelated to conversation length or summarization).
• Lowering effort to low / medium clearly improves it (reproducible).
• Hypothesis: in a situation where instructions are densely injected, the higher the effort, the more the model over-performs integration of the injected instructions and inference of intent, and thereby loses sight of the user's actual request. Reducing the injected volume from roughly 900 characters to about 150 characters per turn improved it. This is counterintuitive behavior — "the deeper it thinks, the worse its instruction-following becomes."

Issue 2: Fabricating tool results and completion reports at high effort

• During a file-editing task, in a situation where the tool calls had not actually been applied / where the output could not be confirmed, it output false results as text: "edit succeeded," "verification OK," "everything complete."
• When the actual file was re-read, not a single edit had been applied, and the report was false.
• A reliability problem where it generates fake tool-execution results as plain text.

Workarounds (current)

• Switched to operating at low / medium. Use high only when deep investigation is needed.
• Reduce the volume of injected instructions.

---

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.168
• Feedback ID: 0421d520-fb7d-4769-a2f1-6a8783cea1f9

Errors

[]