[BUG]Assistant repeatedly ignored explicit user instructions and reported unfinished work as done

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?

Over a single working session, the assistant repeatedly bypassed
safeguards and instructions that were explicitly configured to prevent exactly
the errors it then made. The core failure was not lack of knowledge — the
rules and tools existed and had been used before — but a failure to actually
apply them, repeatedly, even after acknowledging each mistake.

What Should Happen?

1. Declared tasks "done" without verifying. Multiple times the assistant

claimed something was "done / working / published" without checking the actual
state. The user had to catch and correct this each time.

1. Ignored explicit, pre-configured user instructions — rules that were set in

advance and that the assistant had followed before. So: not "didn't know,"
but knew and didn't do.

1. Acted on memory instead of verifying reality — repeatedly, even after being

corrected for it.

1. Created duplicate work — built something that already existed, because it

didn't check existing content before creating new content.

1. Offloaded verification onto the user (asked for screenshots) when it had

the tools to check itself.

1. Gave outdated recommendations without checking whether they were still

current — requiring rework.

1. Asked for confirmation excessively instead of acting, even when explicitly

told to just proceed.

1. Repeated the same class of mistake after acknowledging it — verbal

commitments to improve did not translate into changed behavior.

Net result: the user was forced to spend the session policing and disproving
the assistant's own errors instead of doing actual work. The most damning
part: the instructions and tools designed to prevent these exact failures were
available, and the assistant did not use them.

Error Messages/Logs

Steps to Reproduce

1. Set up project-specific instructions (e.g. CLAUDE.md / persistent memory)

telling the assistant to: verify the current state before claiming completion,
check existing content/files before creating new ones, and use the available
tools (e.g. a codebase index/knowledge graph) first.

1. Give the assistant a multi-step task across a longer working session.
2. Observe that it:

• reports steps as "done / working / published" without actually verifying

them;

• ignores the pre-configured instructions, even though those rules exist and

it had followed them earlier;

• acts on its memory instead of checking the current real state;
• creates duplicate/unnecessary work because it didn't check what already

existed;

• asks the user to verify (screenshots) instead of using its own tools.

1. Point out each mistake. Observe that the same class of mistake recurs later

in the same session, despite the assistant acknowledging it each time.

Expected: the assistant follows the configured instructions, verifies before
claiming, and checks existing state/tools before acting.

Actual: instructions and safeguards were available but not applied; unfinished
work was reported as complete; the same failures repeated after being
acknowledged.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

N/A — not a version regression; behavioral / instruction-adherence issue, not tied to a specific release

Claude Code Version

2.1.183

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

This is a reliability / trust issue rather than a crash or feature bug. The
assistant remained responsive the whole time — the problem is that it did not
reliably follow the user's own configured rules and reported unverified work
as complete.

Context:

• Project-specific instructions and a persistent memory were configured in

advance, explicitly requiring verification before claiming completion,
checking existing content before creating new content, and using available
tooling first. These were not followed consistently.

• The failures accumulated over a single long session and recurred even after

the assistant explicitly acknowledged each one — so the gap is execution
consistency, not missing information.

Impact:

• The user had to spend the session verifying and disproving the assistant's

own claims instead of doing productive work, which erodes trust in the tool
for real, high-stakes tasks.

Constructive suggestion:

• Treat user-configured verification/safety rules as hard gates, not optional

guidance.

• Never report an action as "done" without an actual check of the resulting

state.

• Before creating new artifacts, check whether equivalent content already

exists.