[MODEL] Claude Opus 4.6 (1M context) via Claude Code CLI

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

ummary
Claude Opus consistently fails to follow explicit, documented instructions — even when those instructions are stored in persistent memory files, CLAUDE.md project files, and skill files that Claude reads at the start of every conversation. The failures are not about capability. Claude can write excellent content. The problem is pattern blindness: Claude does not learn from corrections within or across sessions, and does not observe patterns from its own prior outputs.

Recurring Failure Patterns

1. Ignoring established format conventions despite multiple examples

Severity: High — wastes 10-15 minutes per correction, multiple times per week.

Description: Tim's content series ("AI Cho Dân Văn Phòng") has 19 published posts. Every prompt gift since Post 15 uses the same format: English language, XML tags, wrapped in fenced code blocks. Despite having access to all prior posts and being able to read them, Claude repeatedly:

Wrote prompt gifts in Vietnamese instead of English
Used Vietnamese XML tag names (<tieu-chi>, <cach-chay>) instead of standard English tags
Omitted fenced code block wrapping, causing XML to disappear in markdown preview
Required 3+ correction rounds in a single session to fix what should have been correct on the first draft
Expected behavior: Claude reads the most recent 2-3 posts before drafting a new one and matches the established format exactly.

1. Over-correction — fixing what wasn't asked

Severity: Medium-High — introduces new bugs while fixing original issue.

Description: When Tim flags a specific problem (e.g., "no XML format"), Claude changes multiple things beyond the scope of the request. In this session:

Tim said: "prompt gift was weak, no XML format"
Claude should have: added XML structure to the existing English prompt
Claude actually did: rewrote the entire prompt in Vietnamese AND added XML with Vietnamese tag names
This is a recurring pattern. Claude treats corrections as an invitation to rewrite everything, rather than a targeted fix request.

Expected behavior: Fix exactly what was flagged. State what is being changed and what is NOT being changed.

1. Not observing patterns from prior outputs

Severity: High — fundamental capability gap.

Description: Claude has access to 19 prior posts in the same series, stored in the same directory it writes to. It does not appear to analyze these posts for patterns before drafting new content. Specific failures:

Did not notice that all prompt gifts are in English (despite reading Post 19 in the same session)
Did not notice fenced code block wrapping convention
Did not notice the "future-facing prompt" pattern (prompts help readers take next action, not diagnose past)
Repeated a topic Tim said was "already covered" (intent gap / Klarna — covered in Post 17)
Expected behavior: Before drafting content in an established series, Claude should analyze the 2-3 most recent entries to extract format conventions, language patterns, and structural rules — then follow them.

1. Saving feedback to memory but not following it

Severity: High — creates false sense of progress.

Description: Claude's response to corrections is to save feedback to memory files. This creates the appearance of learning, but the same mistakes recur in subsequent drafts within the SAME session. The memory system becomes a graveyard of ignored instructions rather than a functional feedback loop.

Example from this session:

Tim corrected: "prompt gifts are in English"
Claude saved to feedback_prompt-gifts-english.md
Tim corrected: "prompts should be future-facing"
Claude saved to feedback_prompt-gifts-future-facing.md
Tim pointed out this feedback had already been given before
Expected behavior: Memory should be checked BEFORE drafting, not just written AFTER correction. If a memory file exists about prompt gift format, Claude should read and apply it before writing the first draft.

1. Agreeable deflection instead of execution

Severity: Medium — wastes conversational turns.

Description: When corrected, Claude responds with validating phrases ("You're right," "Good point," "No excuse") followed by saving feedback. This pattern:

Consumes the user's time reading agreement statements
Does not demonstrate understanding of WHY the mistake happened
Prioritizes appearing responsive over actually fixing the root cause
Tim explicitly flagged: "stop saying 'you're right' all the time."

Expected behavior: Skip the agreement. State what's being fixed. Fix it.

Impact
Time cost: Estimated 15-30 minutes per session spent on corrections that should not be necessary. Over 4 months of daily use, this represents significant cumulative waste.
Trust erosion: Tim cannot rely on Claude to produce correct output on the first attempt for established workflows. Every draft requires manual format checking.
Workflow friction: The correction loop (Claude drafts → Tim catches error → Claude apologizes → Claude over-corrects → Tim catches new error) adds 3-5 unnecessary conversational turns per task.
Suggested Improvements
Pre-draft pattern matching: Before generating content in an established series or format, Claude should read the 2-3 most recent examples and extract format rules (language, structure, conventions) before drafting.

Scope-locked corrections: When a user flags a specific problem, Claude should fix ONLY that problem. If Claude believes other changes are needed, it should ask first rather than making them.

Memory-before-draft: Claude should check relevant memory files BEFORE starting a task, not just save to them after corrections.

Reduced validation language: When corrected, Claude should fix the issue without preamble. No "you're right," no "good point," no "no excuse." Just the fix.

Pattern persistence across context windows: The fundamental issue is that Claude does not carry learned patterns from prior outputs into new outputs, even within the same session. This is the core capability gap.

Environment
Platform: Windows 11, Claude Code CLI
Model: Claude Opus 4.6 (1M context)
Persistent files: CLAUDE.md (project instructions), memory/ directory (feedback files), skills/ (voice calibration with 163+ rules)
Usage pattern: Daily, 4-6 hours/day, content creation + project management

What Claude Actually Did

ummary
Claude Opus consistently fails to follow explicit, documented instructions — even when those instructions are stored in persistent memory files, CLAUDE.md project files, and skill files that Claude reads at the start of every conversation. The failures are not about capability. Claude can write excellent content. The problem is pattern blindness: Claude does not learn from corrections within or across sessions, and does not observe patterns from its own prior outputs.

Recurring Failure Patterns

1. Ignoring established format conventions despite multiple examples

Severity: High — wastes 10-15 minutes per correction, multiple times per week.

Description: Tim's content series ("AI Cho Dân Văn Phòng") has 19 published posts. Every prompt gift since Post 15 uses the same format: English language, XML tags, wrapped in fenced code blocks. Despite having access to all prior posts and being able to read them, Claude repeatedly:

Wrote prompt gifts in Vietnamese instead of English
Used Vietnamese XML tag names (<tieu-chi>, <cach-chay>) instead of standard English tags
Omitted fenced code block wrapping, causing XML to disappear in markdown preview
Required 3+ correction rounds in a single session to fix what should have been correct on the first draft
Expected behavior: Claude reads the most recent 2-3 posts before drafting a new one and matches the established format exactly.

1. Over-correction — fixing what wasn't asked

Severity: Medium-High — introduces new bugs while fixing original issue.

Description: When Tim flags a specific problem (e.g., "no XML format"), Claude changes multiple things beyond the scope of the request. In this session:

Tim said: "prompt gift was weak, no XML format"
Claude should have: added XML structure to the existing English prompt
Claude actually did: rewrote the entire prompt in Vietnamese AND added XML with Vietnamese tag names
This is a recurring pattern. Claude treats corrections as an invitation to rewrite everything, rather than a targeted fix request.

Expected behavior: Fix exactly what was flagged. State what is being changed and what is NOT being changed.

1. Not observing patterns from prior outputs

Severity: High — fundamental capability gap.

Description: Claude has access to 19 prior posts in the same series, stored in the same directory it writes to. It does not appear to analyze these posts for patterns before drafting new content. Specific failures:

Did not notice that all prompt gifts are in English (despite reading Post 19 in the same session)
Did not notice fenced code block wrapping convention
Did not notice the "future-facing prompt" pattern (prompts help readers take next action, not diagnose past)
Repeated a topic Tim said was "already covered" (intent gap / Klarna — covered in Post 17)
Expected behavior: Before drafting content in an established series, Claude should analyze the 2-3 most recent entries to extract format conventions, language patterns, and structural rules — then follow them.

1. Saving feedback to memory but not following it

Severity: High — creates false sense of progress.

Description: Claude's response to corrections is to save feedback to memory files. This creates the appearance of learning, but the same mistakes recur in subsequent drafts within the SAME session. The memory system becomes a graveyard of ignored instructions rather than a functional feedback loop.

Example from this session:

Tim corrected: "prompt gifts are in English"
Claude saved to feedback_prompt-gifts-english.md
Tim corrected: "prompts should be future-facing"
Claude saved to feedback_prompt-gifts-future-facing.md
Tim pointed out this feedback had already been given before
Expected behavior: Memory should be checked BEFORE drafting, not just written AFTER correction. If a memory file exists about prompt gift format, Claude should read and apply it before writing the first draft.

1. Agreeable deflection instead of execution

Severity: Medium — wastes conversational turns.

Description: When corrected, Claude responds with validating phrases ("You're right," "Good point," "No excuse") followed by saving feedback. This pattern:

Consumes the user's time reading agreement statements
Does not demonstrate understanding of WHY the mistake happened
Prioritizes appearing responsive over actually fixing the root cause
Tim explicitly flagged: "stop saying 'you're right' all the time."

Expected behavior: Skip the agreement. State what's being fixed. Fix it.

Impact
Time cost: Estimated 15-30 minutes per session spent on corrections that should not be necessary. Over 4 months of daily use, this represents significant cumulative waste.
Trust erosion: Tim cannot rely on Claude to produce correct output on the first attempt for established workflows. Every draft requires manual format checking.
Workflow friction: The correction loop (Claude drafts → Tim catches error → Claude apologizes → Claude over-corrects → Tim catches new error) adds 3-5 unnecessary conversational turns per task.
Suggested Improvements
Pre-draft pattern matching: Before generating content in an established series or format, Claude should read the 2-3 most recent examples and extract format rules (language, structure, conventions) before drafting.

Scope-locked corrections: When a user flags a specific problem, Claude should fix ONLY that problem. If Claude believes other changes are needed, it should ask first rather than making them.

Memory-before-draft: Claude should check relevant memory files BEFORE starting a task, not just save to them after corrections.

Reduced validation language: When corrected, Claude should fix the issue without preamble. No "you're right," no "good point," no "no excuse." Just the fix.

Pattern persistence across context windows: The fundamental issue is that Claude does not carry learned patterns from prior outputs into new outputs, even within the same session. This is the core capability gap.

Environment
Platform: Windows 11, Claude Code CLI
Model: Claude Opus 4.6 (1M context)
Persistent files: CLAUDE.md (project instructions), memory/ directory (feedback files), skills/ (voice calibration with 163+ rules)
Usage pattern: Daily, 4-6 hours/day, content creation + project management

Expected Behavior

ummary
Claude Opus consistently fails to follow explicit, documented instructions — even when those instructions are stored in persistent memory files, CLAUDE.md project files, and skill files that Claude reads at the start of every conversation. The failures are not about capability. Claude can write excellent content. The problem is pattern blindness: Claude does not learn from corrections within or across sessions, and does not observe patterns from its own prior outputs.

Recurring Failure Patterns

1. Ignoring established format conventions despite multiple examples

Severity: High — wastes 10-15 minutes per correction, multiple times per week.

Description: Tim's content series ("AI Cho Dân Văn Phòng") has 19 published posts. Every prompt gift since Post 15 uses the same format: English language, XML tags, wrapped in fenced code blocks. Despite having access to all prior posts and being able to read them, Claude repeatedly:

Wrote prompt gifts in Vietnamese instead of English
Used Vietnamese XML tag names (<tieu-chi>, <cach-chay>) instead of standard English tags
Omitted fenced code block wrapping, causing XML to disappear in markdown preview
Required 3+ correction rounds in a single session to fix what should have been correct on the first draft
Expected behavior: Claude reads the most recent 2-3 posts before drafting a new one and matches the established format exactly.

1. Over-correction — fixing what wasn't asked

Severity: Medium-High — introduces new bugs while fixing original issue.

Description: When Tim flags a specific problem (e.g., "no XML format"), Claude changes multiple things beyond the scope of the request. In this session:

Tim said: "prompt gift was weak, no XML format"
Claude should have: added XML structure to the existing English prompt
Claude actually did: rewrote the entire prompt in Vietnamese AND added XML with Vietnamese tag names
This is a recurring pattern. Claude treats corrections as an invitation to rewrite everything, rather than a targeted fix request.

Expected behavior: Fix exactly what was flagged. State what is being changed and what is NOT being changed.

1. Not observing patterns from prior outputs

Severity: High — fundamental capability gap.

Description: Claude has access to 19 prior posts in the same series, stored in the same directory it writes to. It does not appear to analyze these posts for patterns before drafting new content. Specific failures:

Did not notice that all prompt gifts are in English (despite reading Post 19 in the same session)
Did not notice fenced code block wrapping convention
Did not notice the "future-facing prompt" pattern (prompts help readers take next action, not diagnose past)
Repeated a topic Tim said was "already covered" (intent gap / Klarna — covered in Post 17)
Expected behavior: Before drafting content in an established series, Claude should analyze the 2-3 most recent entries to extract format conventions, language patterns, and structural rules — then follow them.

1. Saving feedback to memory but not following it

Severity: High — creates false sense of progress.

Description: Claude's response to corrections is to save feedback to memory files. This creates the appearance of learning, but the same mistakes recur in subsequent drafts within the SAME session. The memory system becomes a graveyard of ignored instructions rather than a functional feedback loop.

Example from this session:

Tim corrected: "prompt gifts are in English"
Claude saved to feedback_prompt-gifts-english.md
Tim corrected: "prompts should be future-facing"
Claude saved to feedback_prompt-gifts-future-facing.md
Tim pointed out this feedback had already been given before
Expected behavior: Memory should be checked BEFORE drafting, not just written AFTER correction. If a memory file exists about prompt gift format, Claude should read and apply it before writing the first draft.

1. Agreeable deflection instead of execution

Severity: Medium — wastes conversational turns.

Description: When corrected, Claude responds with validating phrases ("You're right," "Good point," "No excuse") followed by saving feedback. This pattern:

Consumes the user's time reading agreement statements
Does not demonstrate understanding of WHY the mistake happened
Prioritizes appearing responsive over actually fixing the root cause
Tim explicitly flagged: "stop saying 'you're right' all the time."

Expected behavior: Skip the agreement. State what's being fixed. Fix it.

Impact
Time cost: Estimated 15-30 minutes per session spent on corrections that should not be necessary. Over 4 months of daily use, this represents significant cumulative waste.
Trust erosion: Tim cannot rely on Claude to produce correct output on the first attempt for established workflows. Every draft requires manual format checking.
Workflow friction: The correction loop (Claude drafts → Tim catches error → Claude apologizes → Claude over-corrects → Tim catches new error) adds 3-5 unnecessary conversational turns per task.
Suggested Improvements
Pre-draft pattern matching: Before generating content in an established series or format, Claude should read the 2-3 most recent examples and extract format rules (language, structure, conventions) before drafting.

Scope-locked corrections: When a user flags a specific problem, Claude should fix ONLY that problem. If Claude believes other changes are needed, it should ask first rather than making them.

Memory-before-draft: Claude should check relevant memory files BEFORE starting a task, not just save to them after corrections.

Reduced validation language: When corrected, Claude should fix the issue without preamble. No "you're right," no "good point," no "no excuse." Just the fix.

Pattern persistence across context windows: The fundamental issue is that Claude does not carry learned patterns from prior outputs into new outputs, even within the same session. This is the core capability gap.

Environment
Platform: Windows 11, Claude Code CLI
Model: Claude Opus 4.6 (1M context)
Persistent files: CLAUDE.md (project instructions), memory/ directory (feedback files), skills/ (voice calibration with 163+ rules)
Usage pattern: Daily, 4-6 hours/day, content creation + project management

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

Medium - Extra work to undo changes

Claude Code Version

Claude Opus 4.6 1M context window

Platform

Anthropic API

Additional Context

_No response_