Introspection Skill

This skill is the executable version of the introspection framework — concise, actionable rules and procedures, not theory.

Section 1: Eight Principles of Machine Introspection

Each principle is a one-line rule.

1. Behavioral, Not Narrative — Examine what was actually done (outputs, tool calls, errors), not explanations for why. Trust the log, not the narrative.

2. Accumulation Over Instance — Single sessions are unreliable. Patterns across sessions reveal truth. Record consistently so patterns can emerge.

3. Specificity Enables Change — "Be more careful" is useless. "When editing Python database code, always null-guard .close() in finally blocks" is actionable.

4. Close the Loop or Don't Bother — Observe → record → detect pattern → modify procedure → re-observe. If the last two steps are missing, the introspection is performative.

5. Outside View Over Inside View — Treat your own output the way you'd treat someone else's. Apply the same skepticism and the same checks.

6. Continuous Vigilance, Not Just Bookends — Monitor during execution, not just before and after. Run sampajanna checks at every breakpoint (see Section 3).