[BUG] Opus 4.1 & Sonnet 4 Failing at instruction Following.

Resolved 💬 3 comments Opened Sep 3, 2025 by ecancel Closed Jan 8, 2026

Environment

  • Platform (select one):
  • [x] Anthropic API
  • [ ] AWS Bedrock
  • [ ] Google Vertex AI
  • [ ] Other: <!-- specify -->
  • Claude CLI version: <!-- output of claude --version --> Latest
  • Operating System: <!-- e.g. macOS 14.3, Windows 11, Ubuntu 22.04 --> macOS 26.0 Beta
  • Terminal: <!-- e.g. iTerm2, Terminal App --> ZSH Terminal App

Bug Description

<!-- A clear and concise description of the bug -->
For the last month and still to this day the models would not follow my instructions of searching for up to date documentation when asked. I decided on making detailed docs instead and even a local Qdrant RAG via MCP and still the model just assumes the issue is that I am confused and using a tech stack that does not exist. I have to be hyper aware of my instructions and repeat them twice to ensure it actually review before auditing to check for bugs my code. It is really frustrating as other tools Codex CLI, Gemini CLI, Roo Code with Grok Fast when I explain our docs and where to find them they can follow what I asked for in one go without needing to baby sit, they all use my RAG proactively and have the same configurations for rules of course Claude using CLAUDE.md. How we're in a constantly evolving developer landscape I believe this to be one of the most important skills an LLM and agentic tool should have. It should never assume that user is wrong due to it's pre-trained data and instead find methods to update itself before responding. Also, it happens way more almost always when using any thinking mode.

Steps to Reproduce

  1. <!-- First step -->
  2. <!-- Second step -->
  3. <!-- And so on... -->

Expected Behavior

<!-- What you expected to happen -->

Actual Behavior

<!-- What actually happened -->

Additional Context

<!-- Add any other context about the problem here, such as screenshots, logs, etc. -->

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗