Opus 4.6: Poor instruction adherence and context retention in long sessions

Summary

During a long session working on CI/CD and infrastructure for a .NET + Vite application, Claude Opus 4.6 performed poorly in several ways:

Issues

1. Ignoring explicit user instructions

• User explicitly stated the AWS profile name (platform-dev) earlier in the session. When it came time to run AWS CLI commands, Claude ran them without any profile, got a credentials error, and still didn't use the correct profile until the user repeated it in frustration.

2. Forgetting context within the same session

• Information provided earlier in the conversation was lost or ignored when it became relevant later. The user had to repeat themselves multiple times.

3. Incorrect reasoning / getting basic stuff wrong

• When a Terraform data.aws_ecs_cluster was hanging, Claude initially suggested the cluster didn't exist — despite the user pointing out the app was clearly running (serving HTTP responses). The cluster obviously existed; the real issue was elsewhere.
• Gave a wrong initial diagnosis before checking the actual evidence.

4. Not following skill instructions

• Skills loaded into context (e.g. hf-cicd-setup) contain detailed conventions and instructions. Claude did not consistently follow these, requiring user correction.

5. Arguing when corrected

• Earlier in the session (before context compaction), Claude argued for SQLite in unit tests when the user explicitly said to use InMemory. The user had to firmly correct this.

Impact

The user spent significant time correcting Claude's mistakes and repeating instructions, which defeats the purpose of an AI coding assistant. The session was actively frustrating rather than productive.

Environment

• Model: Claude Opus 4.6
• Platform: Windows 11, bash shell
• Task: .NET 8 API + Vite SPA infrastructure and CI/CD work
• Session length: Long (context compaction occurred)