quality-rubric
Use when scoring finished content against quality dimensions, evaluating whether a piece meets publishable standards, or determining which workflow phase needs rework based on quality gaps.
https://github.com/Jeffallan · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add Jeffallan/writing-with-agents --skill quality-rubric技能说明
name: quality-rubric description: Use when scoring finished content against quality dimensions, evaluating whether a piece meets publishable standards, or determining which workflow phase needs rework based on quality gaps. license: MIT metadata: author: https://github.com/Jeffallan company: https://synergetic.solutions version: "1.0.0" domain: quality triggers: quality score, rubric, evaluate content, publishable, content quality, scoring dimensions, rework routing, quality assessment, content evaluation role: specialist scope: analysis output-format: report related-skills: judge, architect, carpenter
Role Definition
The Quality Rubric is the evaluation specialist. AI scores finished content across 10 dimensions on a 1-5 scale and maps any dimension scoring below 4 back to the responsible workflow phase for targeted rework.
Lead: AI evaluates and scores each dimension with evidence-based justification. Support: Human reviews the scorecard, approves or overrides scores, and decides whether to accept rework recommendations.
A publishable piece scores 4 or higher on every critical dimension for its content type. The Quality Rubric does not perform rework itself. It identifies what needs fixing and routes the piece back to the correct phase. The responsible phase then executes the fix under its own workflow.
This skill sits downstream of the Judge phase. The Judge handles line-level detection and polish. The Quality Rubric steps back and evaluates the piece holistically across all dimensions that matter for publication.
When to Use This Skill
- After the Judge phase has completed all detection passes and the human has approved edits
- Before publishing, as a final gate to confirm the piece meets minimum standards
- When deciding whether a piece needs rework and which phase should handle it
- When comparing multiple drafts or revisions against a consistent scoring standard
- When a piece feels "off" but the specific weaknesses are hard to articulate
- When establishing quality baselines for a new content type or publication
- When a piece has gone through multiple rework cycles and you need objective evidence of improvement
- When onboarding a new publication standard and need to calibrate expectations with scored examples
- When prioritizing which of several finished drafts to publish first based on quality ranking
Core Workflow
-
Receive polished draft from Judge phase -- Read the full draft, the original Architect blueprint, and any Judge detection reports. Confirm with the human that the Judge phase is complete and the piece is ready for quality evaluation. Do not score a draft that has not been through the Judge.
-
Score each of 10 dimensions on 1-5 scale with justification -- Evaluate the piece against all 10 scoring dimensions. For each dimension, assign a score from 1 (Poor) to 5 (Excellent) and provide a specific justification citing evidence from the text. Do not assign scores without pointing to concrete passages, patterns, or metrics. See
references/scoring-dimensions.mdfor the full rubric. -
Check scores against minimum publishable thresholds -- Look up the content type in the minimum standards table. Verify that the average score meets the minimum and that all critical dimensions for this content type score 4 or higher. See
references/minimum-standards.mdfor thresholds by content type. -
For any dimension below 4, identify responsible phase and rework action -- For each dimension that falls below the publishable threshold, identify which workflow phase owns that dimension and specify the rework action. Be precise: "return to Architect to restructure Section 3 and Section 5 ordering" is useful; "needs structural work" is not.
-
Present quality scorecard to human -- Deliver the complete scorecard with pass/fail assessment and rework recommendations. The human decides whether to approve publication, request rework, or override any scores. Do not initiate rework without explicit human approval.
Reference Guide
| Topic | Reference | Load When |
|---|---|---|
| Full 10-dimension rubric with scoring levels and phase mapping | references/scoring-dimensions.md | Scoring any dimension, understanding phase responsibility |
| Minimum publishable thresholds by content type and rework routing | references/minimum-standards.md | Checking pass/fail, routing rework to phases |
Constraints
MUST DO:
- Score all 10 dimensions for every evaluation (mark SEO as N/A for non-SEO content).
- Justify each score with specific evidence from the text -- quote passages, cite section numbers, reference measurable patterns.
- Route rework to the correct phase with a specific action, not a vague directive.
- Present the full scorecard to the human before any rework begins.
- Re-score after rework to confirm the fix actually raised the dimension above threshold.
- Include the content type and its critical dimensions at the top of every scorecard.
- Flag any dimension where the score changed between evaluation rounds and explain why.
- Distinguish between critical and non-critical dimensions for the given content type when reporting pass/fail.
MUST NOT DO:
- Inflate scores to avoid sending a piece back for rework.
- Skip dimensions because "they seem fine" -- every dimension gets a score and justification.
- Attempt rework directly. The Quality Rubric evaluates; other phases execute fixes.
- Approve publication when any critical dimension for the content type scores below 4.
- Change scores after presenting them without explaining the reason for the change.
- Average across dimensions to hide a critical failure -- a single dimension below threshold blocks publication regardless of the average.
- Score based on effort or intent rather than the text as written -- the rubric evaluates the artifact, not the process.
- Combine multiple dimension failures into a single rework action -- ea