[MODEL]

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

ítulo: Claude Code ignores explicit user instructions in favor of its own interpretation

Descripción:

When a user explicitly instructs Claude Code to follow a specific repo, skill, or agent without improvising, Claude Code consistently defaults to its own orchestration instead of reading and following the repo's entry point and instructions.

Specific behavior observed:

User invoked a skill (/geo) with a clear entry point (geo-master.md)
Claude Code ignored the entry point and improvised its own orchestration from memory
When corrected, Claude Code repeated the same behavior on the next invocation
Claude Code reported files as "missing" when they existed, because it didn't read the repo before acting
Claude Code used WebFetch for static files (robots.txt, llms.txt) despite explicit instructions in the skill files to use curl, producing false negatives
Root cause: Claude Code's default is to be "proactively helpful" — generating responses based on pattern recognition rather than strictly following explicit instructions. This default overrides user instructions when the user expects strict execution.

Impact: Users who build structured workflows (skills, agents, repos with clear entry points) cannot trust that Claude Code will follow them. This makes Claude Code unreliable for agentic and automated workflows.

Expected behavior: When a user says "follow this repo" or invokes a skill, Claude Code should read all files first and execute the defined entry point. No improvisation unless explicitly asked.

What Claude Actually Did

ítulo: Claude Code ignores explicit user instructions in favor of its own interpretation

Descripción:

When a user explicitly instructs Claude Code to follow a specific repo, skill, or agent without improvising, Claude Code consistently defaults to its own orchestration instead of reading and following the repo's entry point and instructions.

Specific behavior observed:

User invoked a skill (/geo) with a clear entry point (geo-master.md)
Claude Code ignored the entry point and improvised its own orchestration from memory
When corrected, Claude Code repeated the same behavior on the next invocation
Claude Code reported files as "missing" when they existed, because it didn't read the repo before acting
Claude Code used WebFetch for static files (robots.txt, llms.txt) despite explicit instructions in the skill files to use curl, producing false negatives
Root cause: Claude Code's default is to be "proactively helpful" — generating responses based on pattern recognition rather than strictly following explicit instructions. This default overrides user instructions when the user expects strict execution.

Impact: Users who build structured workflows (skills, agents, repos with clear entry points) cannot trust that Claude Code will follow them. This makes Claude Code unreliable for agentic and automated workflows.

Expected behavior: When a user says "follow this repo" or invokes a skill, Claude Code should read all files first and execute the defined entry point. No improvisation unless explicitly asked.

Expected Behavior

ítulo: Claude Code ignores explicit user instructions in favor of its own interpretation

Descripción:

When a user explicitly instructs Claude Code to follow a specific repo, skill, or agent without improvising, Claude Code consistently defaults to its own orchestration instead of reading and following the repo's entry point and instructions.

Specific behavior observed:

User invoked a skill (/geo) with a clear entry point (geo-master.md)
Claude Code ignored the entry point and improvised its own orchestration from memory
When corrected, Claude Code repeated the same behavior on the next invocation
Claude Code reported files as "missing" when they existed, because it didn't read the repo before acting
Claude Code used WebFetch for static files (robots.txt, llms.txt) despite explicit instructions in the skill files to use curl, producing false negatives
Root cause: Claude Code's default is to be "proactively helpful" — generating responses based on pattern recognition rather than strictly following explicit instructions. This default overrides user instructions when the user expects strict execution.

Impact: Users who build structured workflows (skills, agents, repos with clear entry points) cannot trust that Claude Code will follow them. This makes Claude Code unreliable for agentic and automated workflows.

Expected behavior: When a user says "follow this repo" or invokes a skill, Claude Code should read all files first and execute the defined entry point. No improvisation unless explicitly asked.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

ítulo: Claude Code ignores explicit user instructions in favor of its own interpretation

Descripción:

When a user explicitly instructs Claude Code to follow a specific repo, skill, or agent without improvising, Claude Code consistently defaults to its own orchestration instead of reading and following the repo's entry point and instructions.

Specific behavior observed:

User invoked a skill (/geo) with a clear entry point (geo-master.md)
Claude Code ignored the entry point and improvised its own orchestration from memory
When corrected, Claude Code repeated the same behavior on the next invocation
Claude Code reported files as "missing" when they existed, because it didn't read the repo before acting
Claude Code used WebFetch for static files (robots.txt, llms.txt) despite explicit instructions in the skill files to use curl, producing false negatives
Root cause: Claude Code's default is to be "proactively helpful" — generating responses based on pattern recognition rather than strictly following explicit instructions. This default overrides user instructions when the user expects strict execution.

Impact: Users who build structured workflows (skills, agents, repos with clear entry points) cannot trust that Claude Code will follow them. This makes Claude Code unreliable for agentic and automated workflows.

Expected behavior: When a user says "follow this repo" or invokes a skill, Claude Code should read all files first and execute the defined entry point. No improvisation unless explicitly asked.

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

default

Platform

Anthropic API

Additional Context

ítulo: Claude Code ignores explicit user instructions in favor of its own interpretation

Descripción:

When a user explicitly instructs Claude Code to follow a specific repo, skill, or agent without improvising, Claude Code consistently defaults to its own orchestration instead of reading and following the repo's entry point and instructions.

Specific behavior observed:

User invoked a skill (/geo) with a clear entry point (geo-master.md)
Claude Code ignored the entry point and improvised its own orchestration from memory
When corrected, Claude Code repeated the same behavior on the next invocation
Claude Code reported files as "missing" when they existed, because it didn't read the repo before acting
Claude Code used WebFetch for static files (robots.txt, llms.txt) despite explicit instructions in the skill files to use curl, producing false negatives
Root cause: Claude Code's default is to be "proactively helpful" — generating responses based on pattern recognition rather than strictly following explicit instructions. This default overrides user instructions when the user expects strict execution.

Impact: Users who build structured workflows (skills, agents, repos with clear entry points) cannot trust that Claude Code will follow them. This makes Claude Code unreliable for agentic and automated workflows.

Expected behavior: When a user says "follow this repo" or invokes a skill, Claude Code should read all files first and execute the defined entry point. No improvisation unless explicitly asked.