Assess Quality

Assess Quality

This skill allows the agent to rigorously evaluate the outcome of any executed step, comparing actual results against defined success criteria and producing a structured verdict with actionable feedback.

Steps

1. Establish success criteria. Before evaluating, explicitly define what "good" looks like. If criteria are not provided, derive them from the goal context.
2. Gather the output fresh. Read or receive the actual execution output — do not evaluate from memory.
3. Score against each criterion. For each criterion, assign: ✅ Pass | ⚠️ Partial | ❌ Fail.
4. List discrepancies. For every Partial or Fail, describe exactly what is wrong and why it matters.
5. Determine the overall verdict.
   • All Pass → APPROVED
   • Any Partial (no Fail) → CONDITIONAL (list required fixes)
   • Any Fail → REJECTED (must be reworked before proceeding)
6. Generate feedback. For each non-Pass item, produce a specific, actionable fix instruction — not vague suggestions.
7. Output the full assessment.

Output Format

quality_assessment:
  subject: "<what was evaluated>"
  criteria_scores:
    - criterion: "<criterion name>"
      verdict: "pass | partial | fail"
      notes: "<observation>"
  overall_verdict: "APPROVED | CONDITIONAL | REJECTED"
  discrepancies:
    - item: "<what failed>"
      fix: "<specific action to resolve>"
  feedback_summary: "<one paragraph summary>"

Notes

• Never emit APPROVED without checking every criterion.
• If success criteria are absent, generate them from context before scoring — never evaluate blindly.
• Pair with the planning toolbox to generate a remediation plan when verdict is REJECTED.