[BUG] Severe instruction adherence and reasoning regression since ~March 16-17 — correlated with post-1M-GA server-side changes

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Severe regression in instruction adherence and reasoning quality since ~March 16-17.

I maintain a 30-year .NET/C# codebase with structured CLAUDE.md and custom skills. Launch flags: --permission-mode bypassPermissions --effort max. Plan: Max ($200/mo), Opus 4.6, WSL2.

Timeline:

• March 13-16: After 1M context GA, Claude Code worked exceptionally well. Best performance ever. Skills followed precisely, deep reasoning.
• ~March 16-17: Brutal overnight regression. Same codebase, same CLAUDE.md, same skills, same flags. Nothing changed on my side.
• March 17-19: Persistent degradation. Feels like a 6-month rollback. Skills ignored, shallow reasoning, trial-and-error instead of structured analysis.

This is NOT gradual context rot within a session. It is a step-function drop affecting fresh sessions from the first prompt. Hypothesis: server-side inference optimizations deployed after the long-context premium was removed on March 13.

What Should Happen?

Same quality of instruction adherence and reasoning depth as experienced March 13-16, 2026. CLAUDE.md and skill files should be followed consistently, not treated as optional suggestions.

Error Messages/Logs

No error messages. This is a reasoning quality and instruction adherence regression, not a crash. The model reads skills, can recite them, but does not follow them.

Steps to Reproduce

1. Create a project with structured CLAUDE.md and skill files containing precise development conventions
2. Launch claude --permission-mode bypassPermissions --effort max
3. Request a task covered by the skills
4. Observe that instructions are partially or fully ignored despite being loaded in context

This is server-side — every Max/Opus user with structured skills should be experiencing this since ~March 17.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.79

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

Related Issues

#35309 — "Claude Code disregards stored instructions" (March 17)
#35296 — "1M Context Window does not work as marketed" (March 17)
#34197 — "Claude Code continually ignores CLAUDE.MD" (March 13)
#32161 — "Claude Code systematically ignores CLAUDE.md knowledge retrieval rules" (March 8)