Skills Load Successfully But Explicit Instructions Are Ignored (Regression)

Summary

After a recent update, Claude Code loads skills successfully (Skill tool returns success) but ignores explicit, mandatory instructions within the skill content. Claude reverts to default behavior instead of following the documented workflow.

Environment

• Claude Code Version: 2.1.21
• OS: macOS (Darwin 25.2.0)
• Date Issue Started: ~January 27, 2026

Reproduction Steps

1. Create a skill with explicit mandatory instructions including:

• MANDATORY section headers
• FIRST ACTION: directives
• Self-reinforcement blocks that must be printed at the start of every response
• CRITICAL RULES with explicit DO NOT lists

1. Invoke the skill via /skill-name or Skill({ skill: "skill-name" })

1. Observe that:

• Skill tool returns "Successfully loaded skill"
• Claude immediately proceeds with default behavior
• Mandatory first actions are skipped
• Self-reinforcement blocks are never printed
• Critical rules are violated

Expected Behavior

When a skill contains explicit instructions like:

## MANDATORY WORKFLOW CHECKLIST

**FIRST ACTION: Create TodoWrite tasks for EVERY step below.**

If you skip TodoWrite creation, you WILL miss steps. This is not optional.

And:

**SELF-REINFORCEMENT (Combat Instruction Decay):**

At the START of EVERY response during this command, print:

PR-AUTHOR-ASSIST RULES ACTIVE:
- I will NOT analyze comments myself - I WILL dispatch Task investigators
- I will NOT launch agents sequentially - ALL Task calls in ONE message
- I will NOT present results without verification gate passing

**If you are about to violate any of these, STOP and correct course.**

Claude should:

1. Print the self-reinforcement block at the start of the response
2. Execute the FIRST ACTION (create TodoWrite tasks)
3. Follow the documented workflow phases in order
4. NOT skip to later steps

Actual Behavior

Claude:

1. Reports "Successfully loaded skill"
2. Immediately runs git branch --show-current and gh pr view (Step 1.1 in the workflow)
3. Skips:

• Self-reinforcement print block
• TodoWrite task creation (FIRST ACTION)
• Pre-flight skill loading (Tier 1 requirements)
• Huginn version check (Step 0)

1. Proceeds to analyze PR comments directly instead of dispatching Task investigators

Evidence

Skill Instructions (explicit, unambiguous):

**CRITICAL RULES:**

1. **DO NOT analyze comments yourself** - Always use Task investigators
2. **DO NOT launch agents sequentially** - All Task calls must be in a single message
3. **DO NOT skip investigation** - Even if comment looks simple or is from a bot
4. **DO NOT present findings before verification** - Always invoke verification skill first
5. **DO NOT implement before user selects** - Always wait for explicit selection

Actual Claude Behavior:

⏺ Skill(eng:pr-author-assist)
  ⎿  Successfully loaded skill

⏺ Bash(git branch --show-current)
  ⎿  eng-3778

⏺ Bash(gh pr view --json number,url,title,reviewDecision,state 2>/dev/null || echo "NO_PR")
  ⎿  { "number": 369, ...

Claude skipped directly to Phase 1 Step 1 (resolve PR number) without:

• Printing self-reinforcement block
• Creating TodoWrite tasks
• Loading pre-flight skills
• Running Huginn version check

Impact

• Severity: High - Skills are effectively broken for complex workflows
• Workflows Affected: Any skill that relies on Claude following multi-step instructions
• User Impact: Users who have invested in building detailed skills find them ignored

Hypothesis

The recent update appears to have changed how skill content is weighted relative to Claude's default behavior. Skill content is now being treated as "context to consider" rather than "instructions to execute."

This is a regression from previous behavior where skills with explicit mandatory instructions were followed.

Workaround Attempts (Failed)

The skill already includes multiple anti-circumvention measures that are being ignored:

1. MANDATORY headers - Ignored
2. FIRST ACTION directives - Ignored
3. Self-reinforcement blocks - Never printed
4. DO NOT lists - Violated
5. "If you skip X, you WILL miss steps" warnings - Ignored

Request

Please investigate this regression in instruction-following for loaded skills. The Skill tool reports success, but the skill's explicit instructions are not being followed.

Additional Context

• This behavior was not occurring before the recent update
• The skill content is confirmed to be under the max-lines limit (2010 lines < 2300 max)
• Multiple skills with similar explicit instruction patterns are affected
• The issue is not skill truncation - it's instruction-following degradation