The Director's Diagnostic

This is the Diagnostic Ladder from Chapter 11, applied to the ten failures you'll meet most. Each entry is tagged with the symptom it belongs to, so the five names you learned there are the way in: Length, Tone, Structure, Specificity, and Accuracy. Name the symptom, trace it to Role, Task, or Constraint, and send the targeted fix.

Use this when a prompt isn't working and you can't identify why. Find the symptom that matches what you're seeing and follow it to the cause.

When AI output isn't working, the instinct is to try again with a slightly different prompt. That rarely helps. The output failing is a symptom. The cause is almost always in the prompt structure.

This diagnostic maps ten common failure types to their most likely cause and provides a corrective prompt pattern. Use it when you're getting output that's technically responsive but not useful.

F.1 Output Is Too Generic (Symptom: Specificity)

Symptom: The AI produced something that could apply to any professional in any industry. It hits the obvious beats. It sounds correct but tells you nothing you didn't already know.

Diagnosis: The prompt lacked specificity about audience, context, or constraints. The model defaulted to the most statistically common version of what you asked for, which is also the least distinctive.

Fix: Add three pieces of specificity: who the output is for (role, not "professionals"), what context it will be used in (not just "a meeting" but "a 30-minute meeting with a CFO skeptical of the proposal"), and what it should not include ("avoid recommendations. This is the diagnosis step only"). Specificity forces the model off the generic track.

Corrective prompt pattern:

F.2 Output Is Too Long (Symptom: Length)

Symptom: The model produced a comprehensive, well-structured response that covers everything, and you have to read 800 words to extract the 80 words you actually needed.

Diagnosis: No length constraint was provided. The model optimizes for completeness by default.

Fix: Add a hard word count and specify what to prioritize if cuts are needed. "Under 150 words, prioritize the most actionable insight" is more useful than "make it shorter."

Corrective prompt pattern:

F.3 AI Went Off-Topic (Symptom: Accuracy and Relevance)

Symptom: You asked for one thing and the model gave you something adjacent. You asked for a diagnostic and got recommendations. You asked for a summary and got an analysis. You asked for a draft and got an outline.

Diagnosis: The Task in your prompt was ambiguous or the model inferred a more elaborate version of what you asked for.

Fix: Add a scope constraint that names explicitly what the output should not include. "Do not include recommendations. This is the diagnosis step" is a constraint. "Focus only on the findings, not the implications" is a constraint.

Corrective prompt pattern:

F.4 Output Contains Hallucinated Facts (Symptom: Accuracy and Relevance)

Symptom: The model stated something specific (a statistic, a citation, a named source, a date) that you cannot verify or that you know to be wrong.

Diagnosis: The model completed a statistical pattern into a specific claim. This is how hallucination works. The model doesn't know the difference between what it knows and what it's extrapolating.

Fix: Add two instructions to the prompt: (1) "Do not include statistics or citations unless I provide them," and (2) "If you're drawing on something you're uncertain about, say so." These don't eliminate hallucination, but they reduce the model's incentive to fill gaps with fabricated specifics.

Corrective prompt pattern:

F.5 Wrong Tone (Symptom: Tone)

Symptom: The content is technically correct but the register is off, too formal for a casual update, too casual for an executive briefing, too enthusiastic for a skeptical audience, too hedged for a decisive context.

Diagnosis: A Role problem. Tone wasn't specified, and the model defaulted to the most common register for that content type.

Fix: Name the specific tonal quality you want and provide a contrast: "Direct, not blunt." "Warm, not effusive." "Confident, not arrogant." One-word tone descriptors ("professional," "friendly") are not enough. The model needs the tension between two qualities to calibrate.

Corrective prompt pattern:

F.6 Missed the Point Entirely (Symptom: Specificity)

Symptom: The output is coherent, well-written, and completely misses what you were trying to accomplish. You asked for something and got something that addresses the literal words of your request but not the underlying need.

Diagnosis: The Task was stated at the surface level without the purpose behind it. The model answered the question you asked rather than the question you needed answered.

Fix: Add a sentence explaining why you need this output and what decision or action it will inform. "I need this to convince a skeptical CFO to approve budget" changes the output more than any structural instruction.

Corrective prompt pattern:

F.7 Too Many Options, No Clear Direction (Symptom: Structure)

Symptom: You asked for help making a decision and received a balanced list of pros and cons. Everything sounds reasonable. Nothing tells you what to do.

Diagnosis: The model is trained to present balanced perspectives, not to make recommendations. Without explicit permission to conclude, it defaults to neutrality.

Fix: Add an explicit instruction to take a position: "Recommend one option and defend it. Acknowledge the trade-offs but don't equivocate." If you want balance, ask for the analysis step separately from the recommendation step.

Corrective prompt pattern:

F.8 Output Reads as AI-Written (Symptom: Tone)

Symptom: The output is technically fine but sounds like no human ever wrote it. The sentences are complete. The structure is logical. It sounds hollow.

Diagnosis: A Role problem. The model defaulted to its most probable phrasing, which is also the most frequently trained-on phrasing. Common patterns: opening with a transition phrase, using symmetrical sentence structures, excessive use of "It's important to note," "It's worth mentioning," or similar hedges.

Fix: Add a specificity constraint and a prohibition. "Write in [name of specific voice] register" helps. "Avoid starting any sentence with 'It's'" breaks a common AI habit. Providing a sample of the voice you want to match gives the model something to pattern-match against.

Corrective prompt pattern:

F.9 Keeps Making the Same Mistake (Symptom: any of the five, recurring)

Symptom: You've corrected the model twice for the same issue. It keeps reverting to the problem you already flagged.

Diagnosis: Each correction is treated as an in-context instruction, but it competes with the original prompt. The model is pattern-matching against the full conversation, not just your most recent message.

Fix: Restate the constraint you're enforcing at the top of your next message, not just at the end. Lead with: "Before you do anything else, remember: [the rule you keep correcting]." If the issue persists, start a new session with the corrected constraint built into the original prompt.

Corrective prompt pattern:

F.10 Output Is Technically Correct but Useless (Symptom: Specificity)

Symptom: The model did exactly what you asked. The output is accurate and complete. You can't use any of it.

Diagnosis: The Task was specified without connecting it to the context in which the output would be used. The model produced a correct answer to a question that wasn't the right question.

Fix: This is the hardest failure type to fix because it requires stepping back from the prompt to ask: what do I actually need and why? The fix is usually to add the context the model didn't have: who will read this, what decision they need to make, and what would make this output useful to them.

Corrective prompt pattern: