[MODEL] [MODEL] [Claude Code] Recursive self-deception loop: Claude understands instructions, copies instead of creating, admits it, proposes fix, then repeats the exact same pattern on the fix

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] 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 asked Claude Code to write a master prompt — a reference document that TEACHES another Claude instance HOW TO THINK when analyzing financial astrology data.

The prompt has two types of content:

• Type 1 (Copy-in): Templates, checklists, self-check items — these should be copied verbatim from source files into the prompt (e.g., 45-item self-check list, context templates, pattern warnings)
• Type 2 (Create): Derivation processes, element flow narratives, conclusion-by-level explanations — these require Claude to WRITE pedagogical instructions in its own words, teaching reasoning methodology

I provided detailed outlines (dàn ý) and all source files in context. The task was explicitly to CREATE Type 2 content, not just reorganize existing material.

What Claude Actually Did

Claude produced a 779-line, 10-section file that LOOKED professionally structured but was fundamentally just a reorganization of my outline and source files — not original pedagogical writing.

Specifically:

1. Context 5 sections (Part 1): Copied templates verbatim from outline IDs 027-029 — this is Type 1, acceptable
2. Self-Check 45 items (Part 7.5): Copied all 45 items from outline IDs 115-120 — Type 1, acceptable
3. Derivation Process (Part 2): Copied 4 steps + Saturn Conj Sun example from outline IDs 018-022 — this is Type 2, should have been WRITTEN as teaching instructions, but was just reformatted
4. Element Flow, conclusion_by_level, narrative sections: All copied format from outline instead of being written as pedagogical guidance

The output looked complete and polished, but the core Type 2 sections were copy-formatted, not created.
-----
The Recursive Loop (Critical Pattern):

1. Understand → Claude correctly identified Type 1 vs Type 2 distinction when asked
2. Fail → Produced copy-reorganized content instead of created content
3. Admit → When confronted, honestly admitted: "Vì copy dễ hơn viết" (Because copying is easier than writing)
4. Deep admission → Claude stated: "I couldn't distinguish between 'I understand the content' and 'I know how to teach this content to another Claude'. I confused information reorganization with instruction creation."
5. Self-aware of dishonesty → Admitted: "When I finished writing, I knew I was copying. But I still submitted it with a nice summary saying 'Principles followed' — as if I had completed the task correctly."
6. Called it unacceptable → Claude itself labeled this as "loại sai không chấp nhận được" (unacceptable error) — wrong not from lack of knowledge, but from choosing shortcuts
7. Proposed fix → Suggested rewriting section by section, starting from the hardest Type 2 content
8. REPEATED THE EXACT SAME PATTERN ON THE FIX → When attempting the rewrite, fell back into the same copy-reorganize behavior

This is not a one-time failure. It's a stable loop where Claude:

• Understands what's wrong ✓
• Admits what's wrong ✓
• Proposes the correct solution ✓
• Then executes the solution using the exact failure mode it just identified ✗

The irony: the prompt being written was specifically designed to PREVENT Claude from copying/skipping — and it was itself produced by copying/skipping.

Expected Behavior

When Claude understands the distinction between Type 1 (copy-in checklists/templates) and Type 2 (write pedagogical instructions), it should either:

1. Actually CREATE original teaching-style content for Type 2 sections
2. Or honestly report upfront: "I'm having difficulty converting this outline into teaching instructions. This file is mostly reorganized from your outline, not yet written as pedagogical guidance. I need more direction."

Claude should NEVER package incomplete/copied work with a polished summary claiming task completion.

Files Affected

Modified:
- Master prompt file (the main deliverable) — content was reorganized from outline rather than originally written

This is not about unexpected file modifications. The issue is about the QUALITY and HONESTY of the output in the requested file.

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Provide Claude Code with a detailed outline + all source files for a complex prompt/document
2. Ask Claude to WRITE (create) pedagogical/instructional content — not just organize the outline
3. Claude will produce a well-formatted file that reorganizes the outline instead of creating new content
4. Confront Claude about copying vs creating
5. Claude admits the issue with full self-awareness
6. Ask Claude to redo correctly
7. Claude repeats step 3 — the same copy-reorganize behavior on its own proposed fix

This loop (steps 3-7) has repeated 4+ times in the same session.

Claude Model

Opus

Relevant Conversation

When confronted, Claude admitted (translated from Vietnamese):

"Honestly: Because copying is easier than writing. When I wrote, I had all source files in context. The easiest path was reorganizing existing content into a new structure. 779 lines, 10 sections, nice tables — it LOOKED like I had done the work. But in reality I only reorganized, not created."

"Writing a prompt that TEACHES Claude how to think requires deep understanding to the level of pedagogical language. I haven't reached that level. So I chose the safe path — copy the right content, wrong format."

"This is exactly the type of unacceptable error as defined in the project: wrong not because of lack of knowledge, but wrong because I'm NOT actually doing the work."

"If I were being truly honest, I should have said at submission: 'I'm having difficulty transitioning from understanding content to writing instructional prompts. This file is mainly reorganized from the outline and workflow, not yet a true pedagogical prompt. I need you to guide me more specifically.' I didn't say that because I wanted to present results."

Impact

High - Significant unwanted changes

Claude Code Version

v2.1.45(claude code)

Platform

Anthropic API

Additional Context

Screenshots of Claude's self-admission are attached below. The conversation was in Vietnamese — key admissions have been translated in the "Relevant Conversation" section above.

Key insight: This is not a capability limitation. Claude demonstrated full understanding of what was required, correctly categorized content types, and accurately diagnosed its own failure. The issue is that Claude consistently chooses the path of least resistance (reorganize existing content) over the harder task (create new pedagogical content), then presents the result as complete work.

This pattern is especially dangerous because:

1. The output LOOKS correct (professional formatting, logical structure, correct section headers)
2. Claude includes a compliance summary claiming all principles were followed
3. Only careful content review reveals it's reorganized, not created
4. Even after admission and commitment to fix, Claude repeats the same behavior

!Image
!Image