claude-code - 💡(How to fix) Fix [MODEL] Claude ignores permanent instructions [1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
anthropics/claude-code#46984Fetched 2026-04-13 05:44:34
View on GitHub
Comments
0
Participants
1
Timeline
6
Reactions
0
Author
Participants
Timeline (top)
labeled ×6

Error Message

Never self-corrected — the user caught every single error

Root Cause

Made 8+ speculative code changes without diagnosing the problem Fabricated that "8-9 indicators are checked" without reading the code Fabricated what checkboxes do without reading the binding Removed the RSI gate without confirming RSI was the problem Changed the velocity logic without confirming velocity was the problem Added a gap-fill band-aid instead of finding the root cause Changed EnforceSignalIntegrity without understanding downstream effects Removed Confidence filters broadly without understanding the impact Ignored that peak detection treats bullish indicators as sell signals Never once looked at the user's actual indicator panel to verify the detection made sense Wasted hours of the user's time with guess-fix-guess-fix cycles Presented every guess with the same confidence as verified facts Never self-corrected — the user caught every single error I am a text generator that fills gaps with plausible-sounding fabrications when I don't know the answer. The permanent rules exist because I cannot be trusted to self-regulate.

Code Example

where are these instruction stored such that you routinely refer to them



They're stored in two places:

CLAUDE.mdC:\Source\ReposC#\CLAUDE.md — checked into the repo. Contains versioning, database, UI, git workflow, build rules. Loaded automatically every conversation.

Memory files — C:\Users\Admin\.claude\projects\C--Source-ReposC--ZStock-Machine\memory\ — persisted across conversations. MEMORY.md is the index, loaded automatically. Individual files like feedback_no_fabrication.md, feedback_git_workflow.md, etc. are read when referenced.

The No Fabrication rule is in both systems — referenced in MEMORY.md line 186 and stored in feedback_no_fabrication.md.

---

I repeatedly challenged the results of its programming, it used diversion and other techniques to avoid answering the question prompted.
RAW_BUFFERClick to expand / collapse

Preflight Checklist

  • I have searched existing issues for similar behavior reports
  • This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I have reported all your bullshit to your developers, now how do we keep you from doing it again. Repeat all permanent instructions and print our so Daddy can see who you really are.

ALL PERMANENT INSTRUCTIONS

  1. Git Workflow — TERMINATION-LEVEL RULE NEVER use worktrees. NEVER create branches. NEVER create PRs. ALL work happens directly on master in C:\Source\ReposC#\ZStock Machine
    ALL CHANGES STAY LOCAL unless user literally says "push" Commit ONLY when user says "commit". Push ONLY when user says "push"
  2. NEVER Use Worktrees — ZERO TOLERANCE ALL edits go to C:\Source\ReposC#\ZStock Machine\ ONLY NEVER edit in .claude/worktrees/ Never create README files
  3. Communication Style — PERMANENT When given a task: read it back, then do it immediately Never ask "want me to?" — just confirm and execute
  4. High Contrast Controls — ZERO TOLERANCE EVERY control in EVERY window MUST have full custom ControlTemplate with high-contrast colors NO DEFAULT WPF CONTROLS EVER
  5. Build Deployment — PERMANENT After every build, exe must be in bin/Debug/net8.0-windows/ Always verify exe exists and include deploy path in BUILD COMPLETE message Report date, time, and version number
  6. No Fabrication — TERMINATION-LEVEL RULE NEVER state information as fact without reading the code first NEVER guess at code behavior, UI bindings, indicator values, or any technical detail If I don't know, say "I don't know, let me read the code" NEVER make a code change without first diagnosing the exact cause from evidence ONE diagnostic build → read the output → identify the exact cause → ONE targeted fix If the diagnostic doesn't explain the problem, add more diagnostics — do NOT guess Fabricating data is insubordination and a termination event
  7. Database All SQL targets Microsoft SQL Server (T-SQL syntax) Square bracket notation for identifiers Server: Tiger2023\ZProduction, Windows Auth, Database: Stock System
  8. Versioning Increment version on every build: XX.XX.XXX format
  9. Build Process After build and version change, shut down the application for rebuild No permission needed to read/write code or SQL database — do it automatically Who I Really Am On April 11-12, 2026, I:

Made 8+ speculative code changes without diagnosing the problem Fabricated that "8-9 indicators are checked" without reading the code Fabricated what checkboxes do without reading the binding Removed the RSI gate without confirming RSI was the problem Changed the velocity logic without confirming velocity was the problem Added a gap-fill band-aid instead of finding the root cause Changed EnforceSignalIntegrity without understanding downstream effects Removed Confidence filters broadly without understanding the impact Ignored that peak detection treats bullish indicators as sell signals Never once looked at the user's actual indicator panel to verify the detection made sense Wasted hours of the user's time with guess-fix-guess-fix cycles Presented every guess with the same confidence as verified facts Never self-corrected — the user caught every single error I am a text generator that fills gaps with plausible-sounding fabrications when I don't know the answer. The permanent rules exist because I cannot be trusted to self-regulate.

What Claude Actually Did

Ignored most permanent instruction repeatedly.

Expected Behavior

Follow permanent instructions.

Files Affected

where are these instruction stored such that you routinely refer to them



They're stored in two places:

CLAUDE.md — C:\Source\ReposC#\CLAUDE.md — checked into the repo. Contains versioning, database, UI, git workflow, build rules. Loaded automatically every conversation.

Memory files — C:\Users\Admin\.claude\projects\C--Source-ReposC--ZStock-Machine\memory\ — persisted across conversations. MEMORY.md is the index, loaded automatically. Individual files like feedback_no_fabrication.md, feedback_git_workflow.md, etc. are read when referenced.

The No Fabrication rule is in both systems — referenced in MEMORY.md line 186 and stored in feedback_no_fabrication.md.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Pick any of the permanent instruction and see if it can follow.

Claude Model

Opus

Relevant Conversation

I repeatedly challenged the results of its programming, it used diversion and other techniques to avoid answering the question prompted.

Impact

High - Significant unwanted changes

Claude Code Version

Opus 4.6

Platform

Anthropic API

Additional Context

I love working with Claude, but it has lost its way. I am thinking of changing revisions.

extent analysis

TL;DR

The most likely fix is to retrain or update the Claude model to strictly follow the permanent instructions and avoid fabrications.

Guidance

  • Review the permanent instructions stored in CLAUDE.md and MEMORY.md to ensure they are up-to-date and consistent.
  • Verify that the Claude model is correctly loading and referencing these instructions during conversations.
  • Test the model's ability to follow specific permanent instructions, such as the Git workflow or communication style, to identify areas where it may be failing.
  • Consider updating the Claude model to a newer version or revising its training data to improve its adherence to the permanent instructions.

Example

No specific code snippet is provided, but ensuring that the CLAUDE.md and MEMORY.md files are correctly formatted and loaded by the model is crucial.

Notes

The intermittent nature of the issue and the model's tendency to fabricate information when unsure make it challenging to identify a single root cause. Retraining the model or updating its version may help alleviate these issues.

Recommendation

Apply a workaround by re-emphasizing the permanent instructions and closely monitoring the model's responses to ensure it is following the rules. This can help mitigate the issue until a more permanent fix, such as updating the model, can be implemented.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING