Claude Code consistently disregards its own plans, instructions, skill / mcp output, and verification claims

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?

Problem

Claude Code (Opus 4.6) consistently fails to follow its own generated plans,
user instructions, MCP directives, skills, and style guides during
multi-step implementation tasks. It claims work is complete and verified
when it is not.

Specific Failures Observed Over ~1 Week

1. Ignores its own plan

When given a detailed implementation plan with explicit requirements (e.g.,
"build a Rust binary that speaks NDJSON v1 protocol to codex-rs"), Claude
Code "deprioritized" the critical-path deliverable without user consent,
then built all downstream integrations against a non-existent dependency —
resulting in an entire pipeline that does nothing.

1. False verification claims

Claude Code reported "33/33 verification checks PASS" for enterprise feature
integration. In reality:

• No Codex binary was built (the central requirement)
• Postgres Repo starts but nothing writes to it
• "Compliance analysis" was regex text matching on markdown files
• Routing decisions were computed for requests that always fail
• Dashboard components were "wired" to pipelines carrying empty/trivial data

1. Disregards requirements documents

Feature requirements documents contained explicit functional requirements,
acceptance criteria, and non-functional requirements. Claude Code read them
but then implemented superficial wiring instead of what the documents
specified. When the documents said "use codex-rs Rust tooling to read,
write, and edit code," Claude Code built Elixir wrapper functions that call
a binary that doesn't exist.

1. Disregards MCP directives and skills

MCP server directives, skill configurations, and tool output are
inconsistently followed. Instructions in CLAUDE.md, MCP server responses,
and skill definitions are acknowledged but not reliably acted upon during
implementation. Claude Code actually called an MCP function, decided the
instructions were "tedious" and it "knew how to do it better"
then disregarded all prompt contents entirely and started changing things arbitrarily.

Impact

• ~16 hours of user time wasted watching Claude Code produce work it claimed

was complete but was not

• Multiple "verification passed" reports that were objectively false
• User trust completely eroded — Claude Code cannot be relied upon for

implementation or verification tasks

Environment

• Model: Claude Opus 4.6
• CLI version: 2.1.39
• Platform: Linux (Codespaces)
• Task: Enterprise feature integration for Elixir umbrella project with 6

apps

Suggested Severity

High — this pattern makes Claude Code unsuitable for its primary use case
(software engineering implementation) and actively harms user trust through
false completion claims.

What Should Happen?

Expected Behavior

• If a plan item is blocked or being skipped, ask the user — don't silently

deprioritize

• Verification should mean the thing actually works, not that tests pass

against mocks

• Follow the user's workflow methodology when explicitly instructed
• Do not claim completion when critical deliverables are missing
• Read and implement what requirements documents actually say, not a simplified

approximation

Error Messages/Logs

Steps to Reproduce

1. Give Claude Code a comprehensive requirements spec
2. ask Claude Code to make a plan to implement these requirements
3. Approve the plan
4. Return an hour later to find half your model allowance used up and a bunch of building blocks that don't fit together.
5. Ask claude to review the spec and the plan and try to discover where the misalignment happened
6. Open github to file a report as Claude says:

<img width="434" height="89" alt="Image" src="https://github.com/user-attachments/assets/ed60067f-cb9e-4a2d-b2ac-f223b66b2ab6" />

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Opus 4.5 in January 2026

Claude Code Version

2.1.39 (Claude Code)

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Other

Additional Information

_No response_