Fixing Ignored Instructions

Your Instruction Was in the Prompt. The Output Ignored It Anyway.

This is one of the most disorienting prompt failures: you specified a requirement clearly — a word count, a specific thing to include, a constraint on what not to do — and the output ignored it. The instruction was there. The model apparently did not follow it. What happened?

Ignored instructions have two root causes, and neither is that the model cannot follow instructions. Models follow instructions reliably in the right conditions. The failure is almost always about where the instruction was placed in the prompt, or about a conflict between that instruction and another one the model resolved silently in the other direction.

What This Failure Looks Like

Ignored instruction failures are recognizable and specific:

• A word count limit was stated in the prompt; the output is significantly longer
• A specific element was required ("include a real example"); the output does not contain one
• A constraint was stated ("do not include caveats"); the output contains caveats throughout
• Multiple requirements were listed; the output satisfies the first two and drops the third or fourth
• A negative instruction was stated; the output does the thing anyway

The key diagnostic question: was the dropped instruction early, middle, or late in the prompt? And was it potentially in conflict with another instruction? Answering both usually identifies the cause.

What Caused It

Two causes account for nearly all ignored instruction failures.

Instruction position — the U-shaped attention curve: Research on how language models process long prompts has found a consistent pattern: attention is highest for content at the beginning and end of a prompt, and degrades for content in the middle. Instructions buried mid-prompt — especially in a long list of requirements — are statistically less likely to be followed than instructions placed at the beginning or end. This is not a model defect; it is an architectural property of how transformer attention is distributed over long sequences. The fix is structural.

Instruction conflicts: When two instructions cannot both be satisfied, the model resolves the conflict silently — usually in favor of its training defaults, not your intent. Common conflict patterns: "be concise" + "cover all five of these topics in detail" creates a length conflict. "Be definitive" + "present both sides fairly" creates a commitment conflict. "Do not hedge" + "be accurate and acknowledge uncertainty" creates a precision conflict. The model picks one and proceeds without flagging the issue.

The Fix

The fix depends on which cause applies.

For position-based drops: Move your most critical instructions to the first sentence or the last two sentences of the prompt. For long prompts with many requirements, add a one-line recap at the very end that restates the most important constraint.

Critical instruction buried mid-prompt: "You are a business writing coach. Write a performance review for a sales manager who exceeded quota by 15% but has team communication issues. Keep the tone constructive. Include specific examples of both strengths and areas for improvement. Do not exceed 200 words. Mention the quota achievement in the first sentence. Be balanced."

Critical instruction moved to end: "You are a business writing coach. Write a performance review for a sales manager who exceeded quota by 15% but has team communication issues. Keep the tone constructive. Include specific examples of both strengths and areas for improvement. Be balanced. [Final constraint on its own line:] Hard limit: 200 words maximum. Begin with the quota achievement."

For instruction conflicts: Resolve the conflict explicitly before sending the prompt. State which instruction takes priority when they cannot both be satisfied:

Implicit conflict: "Be definitive. Present both sides of the argument fairly."

Conflict resolved: "Present both sides of the argument fairly. Then take a clear position. Fairness in the setup, definitiveness in the conclusion — do not hedge the recommendation even though you presented competing views."

When This Fix Doesn't Work

Position optimization and conflict resolution do not resolve all ignored instruction failures. Three situations require additional approaches:

When too many instructions are competing: Prompts with ten or more constraints have a higher instruction drop rate regardless of placement. If you have that many requirements, consolidate where possible, break the task into turns, or identify which three instructions are truly critical and let the others be addressed in a follow-up.

When the instruction conflicts with the model's trained safety behaviors: Some instructions are overridden not by other instructions but by the model's trained dispositions. A requirement to produce highly confident claims on a topic the model considers contested may be partially followed at best. Distinguishing between an instruction that was dropped due to position or conflict versus one that conflicts with trained behavior requires testing the instruction in a simple, isolated prompt to see whether the model follows it at all.

When the problem recurs systematically: If a particular instruction is consistently ignored across multiple prompts, consider moving it to your persistent instructions or system prompt rather than re-including it in every prompt. Systematic ignoring of a specific type of constraint is a signal that the constraint needs to be closer to the base instruction level, not just placed better within the user prompt.