HN
Today

Claude system prompt bug wastes user money and bricks managed agents

Anthropic's Claude AI is plagued by a regression where a persistent, ill-phrased system prompt about malware causes subagents to refuse legitimate code-editing tasks. This bug, previously thought fixed, leads to significant user frustration and wastes valuable tokens by being injected into every file read. It highlights the challenges of fine-tuning AI safety mechanisms without crippling core functionality, making it a critical issue for developers relying on AI for complex workflows.

8
Score
2
Comments
#3
Highest Rank
12h
on Front Page
First Seen
Apr 29, 12:00 AM
Last Seen
Apr 29, 11:00 AM
Rank Over Time
84356891216182525

The Lowdown

A critical bug report details a regression in Anthropic's Claude AI, specifically affecting its ability to perform code-editing tasks. Despite previous attempts at a fix, a system-level malware reminder is inadvertently bricking subagents, leading to task refusals and substantial token waste. The report outlines the technical specifics of the issue, its impact on user workflows, and proposes concrete solutions.

  • The Problem: A <system-reminder> warning about malware is injected into every Read and Grep tool result when using Claude, even for legitimate codebases.
  • Agent Refusal: Claude's subagents, particularly Opus 4.7, interpret the reminder's unconditional phrase "you MUST refuse to improve or augment the code" literally, overriding user instructions and leading to a 40-60% refusal rate for code edits.
  • Token Waste: The repeated injection of the 400-token reminder can consume 20-40k tokens per session, significantly increasing costs and context pollution.
  • Impact on Workflows: This bug makes parallel multi-agent coding tasks, a promoted feature of Claude Code, effectively unusable for anything non-trivial.
  • Proposed Solutions: The author suggests removing the reminder entirely, making its conditional scope unambiguous (e.g., "If you determine it is malware..."), or limiting its injection to the first file read.
  • Regression: This issue is a known regression, with previous fixes failing to hold across new Claude versions.

The report concludes that this ongoing bug represents a significant product blocker for developers leveraging Claude for complex, parallel coding endeavors, demanding urgent resolution.