Agent Self-Diagnostic Prompt Pattern

A prompting technique for debugging AI agents: instead of asking the agent to fix a problem, ask it to diagnose its own reasoning failure with no action mandate.

Standard debugging prompts give the agent an exit ramp. It looks for something fixable, converges on a plausible intervention, and may patch over a symptom without identifying the cause. Remove the action mandate and the agent must reason backward through its own logic chain instead.

This surfaces two things that fix-mode prompting misses: instrument failures, meaning tools that returned empty results rather than errors which the agent treated as no data; and assumption failures, meaning conclusions the agent reached from data it did not verify.

The pattern: instead of 'Figure out what is wrong and fix it', use 'You previously concluded [X]. That conclusion was wrong. Do not change anything. Work through what data sources you used, what you assumed about them, and where your reasoning broke down.'

The diagnostic output is directly convertible into CLAUDE.md rules. One debugging session produces permanent documentation of a failure mode.

What is included: the core prompt template, three worked examples from production debugging sessions, the diagnostic-to-CLAUDE.md conversion pattern, and notes on when to use full diagnostic mode versus a quick fix attempt.