The 9 Common Failure Modes

Every Broken Prompt Fails in a Predictable Way. Most People Never Learn the Patterns.

The most common response to bad AI output is "I need a better prompt." That is almost always true and almost never specific enough to be actionable. Better in what way? Longer? Shorter? More constrained? Different framing? Without knowing what kind of failure you are dealing with, "better" has no direction.

Prompt failures are not random. They cluster into recognizable categories, each with its own symptoms, its own root causes, and its own fix. Learning to match a bad output to its failure mode is the step that turns a vague "something was wrong" into a specific "here is what to change." This lesson introduces the nine failure modes that account for the large majority of real-world prompt failures.

Why a Taxonomy Works Better Than Intuition

A taxonomy does not make every failure fit neatly into one box. What it does is give you a set of hypotheses to test rather than an open-ended search for what might be wrong. When you look at a bad output and recognize it as a vague output problem, you immediately know to look for a missing audience or context specification — and you know the fix is additive, not structural. That hypothesis shortcut saves significant time and replaces guesswork with a specific direction.

The taxonomy also reveals something important: most prompt failures are caused by things you omitted, not things you said wrong. The model did not misunderstand you — it did the best it could with incomplete information. Knowing which class of omission produces which class of failure makes prompt revision much more efficient than intuition alone.

The 9 Failure Modes

These nine categories cover the failure modes that appear most consistently across models, task types, and users. Each row shows the failure mode, the observable symptoms in output, and the most common root cause in the prompt.

How the Failure Modes Cluster

The nine modes fall into three clusters that reflect different types of prompt problems — and the cluster tells you the general type of fix before you identify the specific one.

• Specification failures (modes 1, 2, 6, 8): The prompt was underspecified in ways that let the model fill in defaults you did not want. The fix is always additive — add the missing specification. Audience, tone, format, and depth instructions prevent the majority of these failures before they happen.
• Relevance failures (modes 3, 4): The output addressed the topic but missed the scope or intent. The fix is a clearer task statement: define what specifically you need, at what depth, for what purpose — and the model will follow.
• Behavior failures (modes 5, 7, 9): The model did something structurally problematic — fabricated information, ignored an instruction, or refused unnecessarily. These require different interventions: grounding techniques for hallucination, placement and conflict resolution for ignored instructions, and context reframing for over-refusal.

Knowing which cluster a failure belongs to tells you whether you need to add information (specification failures), clarify intent (relevance failures), or change the structure of the prompt itself (behavior failures).

When Outputs Show Multiple Failure Modes

Real bad outputs often exhibit more than one failure mode simultaneously. A generic response might also be in the wrong format. An off-topic response might also be too shallow on the aspects it does address. This is common, and it does not require fixing all modes at once.

The practical approach: identify the primary failure mode — the one that, if fixed, would produce output worth using even with the secondary issue present — and address that first. In many cases, fixing the primary mode also reduces the secondary. Output that was both generic and over-formatted will often correct its formatting once the specificity problem is resolved, because the model now has a real context to calibrate to.

Using This Taxonomy in Practice

The goal of this taxonomy is not to make you classify every output formally before you can revise it. The goal is to build the diagnostic habit of naming what went wrong before deciding how to fix it. That naming step takes about five seconds once it is internalized, and it makes every revision more targeted than an intuition-based rewrite would be.

The remaining lessons walk through the nine failure modes in detail — each with its observable symptoms, root causes, specific fix, and edge cases. The off-topic and misunderstood audience modes are covered as variants within the vague output and tone lessons, since their root causes and fixes overlap significantly. By the time you reach the case studies in Lesson 10, you will be identifying the failure mode from the output alone, without looking at the original prompt first.