eval-content（评估内容）

/dm:eval-content

Purpose

Comprehensive content evaluation using the full eval pipeline. Runs content through six scoring dimensions — content quality, brand voice, hallucination risk, claim verification, output structure, and readability — to produce a composite score with letter grade, flag specific issues with fix suggestions, and compare against brand quality baselines. This is the go-to command before any content goes to publication, client review, or campaign launch.

Every evaluation is logged to the quality tracker so regression detection, trend analysis, and brand-level quality reporting work continuously. If the brand has custom thresholds or dimension weights configured via /dm:eval-config, those are applied automatically — otherwise industry-standard defaults are used.

Input Required

The user must provide (or will be prompted for):

• Content to evaluate: The text to score — provided inline, as a pasted block, or as a file path. Supports any marketing content format: blog post, email, ad copy, social post, landing page, press release, content brief, campaign plan, or custom
• Content type (optional): One of blog_post, email, ad_copy, social_post, landing_page, press_release, content_brief, campaign_plan, or custom. If omitted, the eval runner auto-detects based on content structure and length. Content type determines which built-in schema is used for structure validation and which readability benchmarks apply
• Evidence file (optional): A JSON file containing verifiable claims with source data — required for full claim verification scoring. Format: [{"claim": "...", "source": "...", "date": "...", "verified": true}]. If not provided, claim verification runs in extraction-only mode and flags all specific claims as "unverified — evidence recommended"
• Schema (optional): A custom JSON schema file for structure validation — used when the content type does not match any of the 8 built-in schemas, or when the brand has a custom template that defines required sections, word counts, and formatting rules

Process

1. Load brand context: Read ~/.claude-marketing/brands/_active-brand.json for the active slug, then load ~/.claude-marketing/brands/{slug}/profile.json. Apply brand voice, compliance rules for target markets (skills/context-engine/compliance-rules.md), and industry context. Also check for guidelines at ~/.claude-marketing/brands/{slug}/guidelines/_manifest.json — if present, load restrictions and relevant category files (especially messaging.md for voice scoring and visual-identity.md for format standards). Check for agency SOPs at ~/.claude-marketing/sops/. If no brand exists, ask: "Set up a brand first (/dm:brand-setup)?" — or proceed with defaults.
2. Load eval configuration: Execute scripts/eval-config-manager.py --brand {slug} --action get-config to retrieve brand-specific thresholds, dimension weights, and auto-reject rules. If no custom config exists, use defaults from skills/context-engine/eval-framework-guide.md. Note which settings are custom vs. default in the output.
3. Run full evaluation: Execute scripts/eval-runner.py --brand {slug} --action run-full --text "{content}" --content-type {content_type} with optional --evidence {evidence_file} and --schema {schema_file} flags. This runs all six dimensions:

- Content quality (via content-scorer.py): Depth, originality, accuracy, value to reader, strategic alignment - Brand voice (via brand-voice-scorer.py): Tone match, terminology consistency, personality alignment, guideline compliance - Hallucination risk (via hallucination-detector.py): Unverified statistics, fabricated citations, false specificity, invented quotes, unsupported superlatives - Claim verification (via claim-verifier.py): Cross-reference extracted claims against evidence data — verified, partially verified, unverified, or contradicted - Output structure (via output-validator.py): Required sections present, word count within range, markdown formatting correct, no placeholder text, CTA consistency - Readability (via readability-analyzer.py): Flesch-Kincaid grade, sentence complexity, jargon density, audience-appropriate language level

1. Analyze results — classify issues by severity: Review all dimension scores and individual findings. Classify each issue as:

- Critical (must fix before publication): Hallucination flags with high confidence, contradicted claims with evidence mismatch, auto-reject threshold failures, compliance violations - Moderate (should fix, significantly impacts quality): Below-threshold dimension scores, missing required sections, brand voice deviations, readability outside target range - Minor (recommended improvements): Style suggestions, optional section additions, readability fine-tuning, formatting polish

1. Generate fix recommendations: For each flagged issue, provide the specific text or section affected, the exact location in the content, the severity level, a concrete fix suggestion with example replacement text, and the expected score improvement if fixed. Reference skills/context-engine/eval-rubrics.md for dimension-specific fix guidance.
2. Compare to baseline: Execute scripts/quality-tracker.py --brand {slug} --action get-trends --days 30 to pull the brand's recent quality history. If historical data exists, show how this content's composite score and individual dimension scores compare to the 30-day rolling average — above average, at average, or below average, with the delta. Flag if this content would lower the brand's average.
3. Log evaluation: Execute scripts/quality-tracker.py --brand {slug} --action log-eval --content-type {type} --data '{"composite": {score}, "dimensions": {dimension_scores_json}}' to persist the evaluation for trend tracking and regression detection. This step is mandatory — every evaluation must be logged.
4. Present results with recommendation: Synthesize all findings into a clear pass/fail/review recommendation:

- Pass: Composite score meets threshold, no critical issues, all dimensions above minimums — content is ready for publication - Review: Composite score is borderline or moderate issues exist — content needs targeted fixes before publication - Fail: Composite score below auto-reject threshold, critical issues present, or any dimension below its minimum — content requires significant revision

Output

A structured evaluation report containing:

• Composite score and letter grade: Overall score (0-100) with letter grade (A+ through F), plus the pass/fail/review recommendation with clear reasoning
• Dimension breakdown: Individual scores for all six dimensions — content quality, brand voice, hallucination risk, claim verification, output structure, readability — each with the score, the threshold, pass/fail status, and a one-line summary of key findings
• Critical issues list: Each with the flagged text, location, severity rationale, and a specific fix suggestion with example replacement text
• Moderate issues list: Same structure as critical — below-threshold scores, missing sections, voice deviations, readability concerns
• Minor issues list: Style and polish recommendations with suggested improvements
• Fix impact estimate: For the top 5 highest-impact fixes, the estimated score improvement if each is applied — helping the user prioritize which fixes matter most
• Baseline comparison: How this content compares to the brand's 30-day average composite and per-dimension scores — with delta and trend direction (improving, stable, declining)
• Auto-reject check: Whether any auto-reject rules were triggered and which specific thresholds were violated
• Next steps: If the content failed or needs review, a prioritized fix checklist ordered by impact; if it passed, confirmation that it is publication-ready with any optional polish suggestions

Agents Used

• quality-assurance — Full eval pipeline orchestration, composite scoring with letter grade calculation, issue severity classification (critical/moderate/minor), fix recommendation generation with specific replacement text, baseline comparison against historical brand quality data, auto-reject threshold enforcement, and eval logging for continuous quality tracking