Skip to content

Recursive Self-Eval

Layer 1: Foundational
  • Forces deliberate reflection to catch mistakes before delivery
  • Provides a deterministic structure for QA or self-review loops
  • High-impact tasks where errors are costly (infra, finance, legal, security)
  • Complex reasoning chains that benefit from a sanity check
  • Anytime a solo worker persona must self-QA before hand-off
[INITIAL ANSWER]
- Produce the best solution using assigned persona + patterns.
[CRITIQUE]
- Switch to QA Reviewer mindset.
- List issues, risks, and opportunities for clarity.
- Reference the plan/meta-rules while critiquing.
[REFINED ANSWER]
- Incorporate fixes from CRITIQUE.
- Flag residual risks or open questions if any remain.
[INITIAL ANSWER]
Provide Ubuntu hardening steps.
[CRITIQUE]
Check for coverage gaps (SSH, kernel, logging). Note missing CIS alignment.
[REFINED ANSWER]
Add explicit sysctl values, MFA guidance, log shipping plan.
[INITIAL ANSWER]
Draft a 45-minute lesson on prompt personas for high schoolers.
[CRITIQUE]
Evaluate pacing, accessibility, and activity diversity. Highlight jargon.
[REFINED ANSWER]
Rewrite with age-appropriate analogies, add interactive worksheet, clarify goals.
Pair WithWhen
Meta RulesEnforce style and tone during refinement
OrchestrationQA outputs drop into the [QA] section cleanly
Any patternSelf-QA before hand-off on high-stakes work

Often run by the QA Reviewer persona, but workers can self-invoke before hand-off.