Content Quality Evaluator

Content Quality Evaluator

Purpose

This skill acts as a strict quality gatekeeper for content, evaluating articles against industry-standard criteria before publication. It produces a detailed scorecard with pass/fail determination and actionable feedback for AI or human correction.

Use this skill when:

• Reviewing AI-generated content before publication
• Evaluating blog posts, articles, or long-form content
• Ensuring consistent quality standards across all published content
• Auditing existing content for improvement opportunities

Quick Reference

Evaluation Categories (26 Criteria)

Decision Thresholds

Note: Gate failures (Ethics & Compliance) result in immediate FAIL regardless of other scores.

Usage Patterns

Automatic Loading

The skill loads when Claude detects:

• Content review or evaluation requests
• Pre-publish quality checks
• Requests to "evaluate," "review," or "score" content
• References to content quality or E-E-A-T

Manual Invocation

Invoke with: /content-quality-evaluator

Or ask:

• "Evaluate this article for quality"
• "Score this blog post"
• "Is this content ready to publish?"
• "Run a quality check on this draft"

Required Brief Metadata

Derive from the brief and the client context (ask ONLY if absent from both):

• author/byline and bio: signature rule in the client guidelines-donts.md
• disclaimer_required: from guidelines-donts.md regulatory constraints (none|affiliate|finance|legal|medical)

MCP Tools Required

For Fact Verification (CT-7), harness-aware: In Claude Code: ToolSearch: "select:mcp__crawl4ai__crawl_url". In other harnesses: any URL-fetch tool. Use crawl4ai to fetch and parse source websites for verifying factual claims (pricing, stats, features, dates).

Fallback: If crawl4ai unavailable, mark CT-7 as N/A with note "Fact verification skipped - crawl4ai unavailable"

Reference Loading Sequence

This skill uses progressive loading to minimize token usage while maintaining thoroughness:

Phase 1 - Initialization: Load evaluation-criteria.md and scoring-methodology.md to understand the 26 criteria and scoring logic.

Phase 2 - Gate Check: Load the client context (clients/<slug>/context/): brand-voice.md, vocabulary.md, guidelines-donts.md, proof-points.md, writing-guide.md, to evaluate EC-1 through EC-5.

Phase 3 - Fact Verification: Load crawl4ai MCP tool to verify factual claims in CT-7 before full evaluation.

Phase 4 - Full Evaluation: Score all criteria using loaded references. Load the client linking-clusters.md when evaluating internal linking.

Phase 5 - Feedback Generation: Load feedback-templates.md to produce structured, actionable feedback for failing criteria.

Phase 6 - Output: Use quality-scorecard.md template to format final results.

Evaluation Process

Step 1: Gate Check (Category 6)

Evaluate pass/fail criteria before proceeding to full scoring:

• EC-1: Originality - Verify no plagiarism patterns exist
• EC-2: Legal compliance - Confirm required disclaimers are present per brief flag
• EC-3: Brand voice alignment - Check adherence to brand-voice.md guidelines
• EC-4: Guardrails compliance - No violation of the HARD rules in the client guidelines-donts.md (promised results, denigration, forbidden claims...)
• EC-5: Client proof compliance - Every claim about the client exists in proof-points.md; invented testimonials or numbers = FAIL

Gate failure triggers immediate FAIL verdict with specific issues documented. Do not proceed to full evaluation if gate fails.

Step 2: Fact Verification (CT-7)

Verify all factual claims before scoring:

1. Extract verifiable claims (prices, statistics, features, dates, quotes)
2. Identify source URLs for each claim
3. Use crawl4ai to fetch and parse source content
4. Compare article claims against source data
5. Document verification results (Pass/Fail per claim) in scorecard

If crawl4ai unavailable, mark CT-7 as N/A and note "Fact verification skipped - crawl4ai unavailable" in scorecard.

Step 3: Full Evaluation (Categories 1-5)

Score each criterion using the 4-point scale defined in scoring-methodology.md. Mark criteria as N/A when not applicable (these are excluded from final score calculation).

Step 4: Calculate Score

Apply category weights per scoring-methodology.md, excluding N/A criteria and renormalizing weights accordingly. Verify no category falls below minimum thresholds. Determine final verdict: PASS (≥80%), CONDITIONAL (70-79%), or FAIL (<70%).

Step 5: Generate Output

Produce the quality-scorecard.md with all category scores, criterion-level scores, structured feedback for failing criteria using token-efficient format, strengths summary, and final verdict with recommended action.

Post-Evaluation Workflow

The evaluation produces a verdict that determines next steps:

PASS (≥80%): Content meets quality standards. Proceed directly to publication.

CONDITIONAL (70-79%): Content is publishable but has improvement opportunities. Present user with two options:

• Revise now using AI correction flow (recommended)
• Editor override - publish as-is with quality score documented

FAIL (<70%): Content requires revision before publication. Present user with two options:

• AI correction flow (automated batch fixes)
• Manual review (user handles revisions)

AI Correction Decision Tree

[Evaluation Complete]
         |
         v
    [Score ≥80%?]
    /           \
  YES            NO
   |              |
[PASS]      [Score ≥70%?]
   |          /         \
[Publish]   YES         NO
             |           |
      [CONDITIONAL]   [FAIL]
             |           |
      [User Choice]  [User Choice]
       /        \      /        \
   Revise    Publish  AI      Manual
      |         |      |         |
      v         v      v         v
   [Correction Flow] [Correction Flow] [Human Review]
         |                  |
         v                  v
   [Re-evaluate]      [Re-evaluate]
         |                  |
   [Attempt < 3?]     [Attempt < 3?]
      /      \           /      \
    YES      NO        YES      NO
     |        |         |        |
   [Loop]  [Escalate] [Loop]  [Escalate]

Correction Flow Steps

When AI correction is selected:

1. Generate Feedback: Load feedback-templates.md and produce structured, criterion-coded feedback for all failing items
2. Apply Fixes: Batch all corrections in a single revision pass
3. Document Changes: Create revision history showing original vs corrected content with criterion codes
4. Re-evaluate: Run full evaluation on corrected content
5. Loop or Complete: If score still fails and attempts < 3, repeat from step 1. If attempts = 3, escalate to human review with all scorecards and revision logs.

Maximum 3 correction attempts per article. After 3 failed attempts, the content requires human intervention to address persistent issues.

Output Format

Feedback Format (Token-Efficient)

**CT-3: Source Citation** | ¶4: "Studies show..." | Add source link or cite study
**SR-2: Scannable Formatting** | §2: Wall of text | Break into 3 paragraphs

Revision History Format

| Location | Original | Corrected | Criterion |
|----------|----------|-----------|-----------|
| ¶4 | "Studies show..." | "According to [Source]..." | CT-3 |

Integration

With the client context (clients/<slug>/context/)

Automatically loads when available:

• brand-voice.md → EC-3 (Brand Voice Alignment)
• vocabulary.md → Terminology checks (approved AND banned terms)
• guidelines-donts.md → EC-2 disclaimers + EC-4 hard guardrails
• proof-points.md → EC-5 client proof compliance
• writing-guide.md → WQ-5 AI writing patterns
• linking-clusters.md → internal/external link rules

In Content Pipeline

This skill functions as the quality gate between article creation and publication. The typical flow is: Brief Creation (Content Brief Creator skill) produces a comprehensive brief, Article Creation (Article Creator skill) generates the draft, Content Quality Evaluation (this skill) scores and validates the content, and finally Publication occurs only after PASS or approved CONDITIONAL verdict.

An orchestrator (human or agent) can automate this sequence, including the correction loop, requiring human intervention only for CONDITIONAL verdicts or escalated failures.

Reference Files

Maintenance

Review Frequency: Quarterly

• Update criteria based on Google guidelines changes
• Refine feedback templates based on common issues
• Adjust weights if needed based on content performance data

This skill uses progressive disclosure - reference files are loaded on-demand based on evaluation needs.