Evaluation Framework

Structured, rubric-based evaluation for any output before it's used, shared, or published. Produces reproducible quality assessments with specific evidence and actionable improvements.

When to Use This Skill

Apply evaluation when:

• Any document or output is about to be shared externally
• Comparing two versions or approaches
• Reviewing AI-generated content for quality before use
• Doing a systematic pre-launch check
• Auditing existing documents for gaps

Core Evaluation Dimensions

Universal Rubric (applies to any output type)

Dimension	Weight	Criteria
Accuracy	25%	Is the content factually correct? No hallucinations, outdated data, or false claims?
Completeness	20%	Does it cover all necessary areas? What's missing?
Usefulness	25%	Does it achieve its stated purpose? Better than alternatives?
Clarity	15%	Is it readable, well-structured, unambiguous?
Freshness	15%	Is the information current? Any stale references or outdated framing?

Scoring Scale

Score	Label	Criteria
5	Exceptional	Exceeds expectations; better than hand-crafted alternatives
4	Good	Minor gaps; clearly useful and fit-for-purpose
3	Adequate	Functional but notable room for improvement
2	Below average	Notable gaps or errors that undermine usefulness
1	Poor	Significant issues; needs substantial rework

Evaluation Process

Step 1: Understand the Purpose

Before scoring, answer:

• What is this output trying to achieve?
• Who is the audience?
• What's the benchmark for "good"?

Step 2: Score Each Dimension

For each dimension, provide:

• A score (1-5)
• Specific evidence (quote the exact issue or strength)
• An actionable improvement (if score < 4)

#### [Output Name] — Evaluation

| Dimension | Score | Evidence | Improvement |
|-----------|-------|----------|-------------|
| Accuracy | [1-5] | "[exact quote or specific example]" | [what to fix] |
| Completeness | [1-5] | "[what's missing or present]" | [what to add] |
| Usefulness | [1-5] | "[does it achieve its goal?]" | [how to improve] |
| Clarity | [1-5] | "[structural or language issues]" | [how to clarify] |
| Freshness | [1-5] | "[stale elements found]" | [what to update] |
| **Weighted Total** | **[X.X/5]** | | |

Step 3: Calculate Weighted Score

Weighted = (Accuracy × 0.25) + (Completeness × 0.20) + (Usefulness × 0.25) + (Clarity × 0.15) + (Freshness × 0.15)

Step 4: Classify and Recommend

Score	Grade	Recommendation
4.5–5.0	A	Share/publish as-is
3.5–4.4	B	Minor polish recommended
2.5–3.4	C	Notable improvements needed before sharing
1.5–2.4	D	Significant rework required
< 1.5	F	Start over or fundamentally rethink

Step 5: Prioritized Improvement List

Always produce a ranked list of improvements:

### Improvements (Priority Order)

**MUST FIX** (blocks sharing):
1. [Specific issue] — [why it matters] — [how to fix]

**SHOULD FIX** (significant quality gain):
2. [Specific issue] — [why it matters] — [how to fix]

**NICE TO HAVE** (minor polish):
3. [Specific issue] — [why it matters] — [how to fix]

Output-Type Specific Extensions

For Strategic Documents (plans, analyses, reviews)

Add assessment of:

• Actionability: Are next steps specific enough to act on?
• Honesty: Does it acknowledge risks and failure modes?
• Evidence: Are claims supported by data or reasoning?

For Technical Documentation

Add assessment of:

• Reproducibility: Can someone follow this without prior knowledge?
• Accuracy of commands: Are code examples tested and correct?
• Coverage of edge cases: What happens when things go wrong?

For Research Summaries

Add assessment of:

• Source quality: Are the underlying sources credible?
• Synthesis quality: Are patterns and insights genuine vs surface-level?
• Bias check: Does the summary fairly represent conflicting evidence?

Quality Criteria

A good evaluation MUST:

• Score every dimension (never skip or bundle)
• Provide specific evidence for each score (no vague "quality is low")
• Make every improvement actionable (no "make it better")
• Separate MUST FIX from nice-to-have
• Calculate the weighted total correctly
• Be honest even when the score is high (5/5 is almost never correct)

Anti-Patterns to Avoid

• Vague scores: Giving 4/5 without explaining what's good or what's missing
• Generic improvements: "Improve clarity" without specifying what's unclear
• Missing evidence: Claiming accuracy issues without quoting the specific error
• Grade inflation: Scoring 4+ when issues are present to seem positive
• Omitting dimensions: Skipping freshness because "it seems recent"