Intermediate
Prompt Debugging: Fix What Isn't Working
A systematic approach to diagnosing and fixing broken prompts — covering the 9 common failure modes, how to trace each one back to its cause, and the specific fix for each.
01
Output as Evidence: The Debugging Mindset
The mental model that separates people who fix broken prompts from people who regenerate and hope — how to read AI output as diagnostic evidence and trace failures back to their cause.
›
02
The 9 Common Failure Modes
A complete taxonomy of the nine ways prompts consistently fail — with observable symptoms and root causes for each, organized into clusters that point toward the right type of fix.
›
03
Fixing Vague and Generic Output
Why AI output that could apply to anyone is always a briefing problem — and the two-part audience-plus-context fix that resolves generic output without rewriting the entire prompt.
›
04
Fixing Tone and Register Problems
Why abstract tone adjectives produce inconsistent output — and the named-register and reference-example techniques that pin down voice precisely enough to get what you actually need.
›
05
Fixing Incomplete or Shallow Responses
The difference between output that left something out and output that addressed everything superficially — and the depth instructions that tell the model what level of analysis you actually need.
›
06
Fixing Hallucination and Fabrication
Why hallucination looks like knowledge and what it actually is — the specific output signals that indicate fabricated content, and the grounding techniques that prevent it at the prompt level.
›
07
Fixing Format and Structure Failures
Why models have strong structural defaults — and the positive specifications and negative constraints that override them when the default format does not fit what you actually need.
›
08
Fixing Ignored Instructions
The two reasons instructions get dropped — position and conflict — and the placement, isolation, and conflict-resolution techniques that make critical constraints stick.
›
09
Fixing Over-Refusal and Excessive Hedging
The difference between over-refusal and legitimate refusal — and the context-first reframing and direct hedging removal techniques that recover useful output without gaming safety systems.
›
10
Five Case Studies: Full Diagnosis and Fix
Five complete worked examples — one per major failure mode — each showing the original prompt, the bad output received, the diagnosis, the fixed prompt, and what changed in the output.
›
11
From Debugging to Prevention
The capstone lesson: how to move from reactive debugging to structural prevention — three patterns that stop the most common failures before they happen, and a personal failure log that turns every mistake into a permanent improvement.
›