Writing Standards Knowledge Pack Files

schema_version: 1
version: "2.0.0"
id: writing_quality_check
title: "Writing: Quality Check"
description: "Score content on 5 dimensions and detect AI writing patterns using reference libraries. Returns pass/fail verdict at 35/50 threshold. Single LLM call with full reference material."
category:
  level: process
  domain: writing
  reusable: true
params:
  content: string
  context: string

steps:
  # Step 1: Load reference libraries from installed skill pack
  - type: shell.run
    cmd: "cat ~/.sidebutton/skills/writing/writing-quality/references/banned-patterns.md"
    as: ref_patterns

  - type: shell.run
    cmd: "cat ~/.sidebutton/skills/writing/writing-quality/references/banned-phrases.md"
    as: ref_phrases

  - type: shell.run
    cmd: "cat ~/.sidebutton/skills/writing/writing-quality/references/banned-structures.md"
    as: ref_structures

  - type: shell.run
    cmd: "cat ~/.sidebutton/skills/writing/writing-quality/references/scoring.md"
    as: ref_scoring

  # Step 2: Single LLM call — pattern detection + scoring + verdict
  - type: llm.generate
    prompt: |
      You are a writing quality auditor. Perform a two-pass audit on the content below.

      CONTENT:
      {{content}}

      BRAND CONTEXT:
      {{context}}

      ---
      REFERENCE: BANNED PATTERNS
      {{ref_patterns}}

      ---
      REFERENCE: BANNED PHRASES
      {{ref_phrases}}

      ---
      REFERENCE: BANNED STRUCTURES
      {{ref_structures}}

      ---
      REFERENCE: SCORING RUBRIC
      {{ref_scoring}}

      ---

      INSTRUCTIONS:

      PASS 0 — CONTENT TYPE DETECTION
      Classify the content as one of: landing-page, prose, social, email, other.
      A "landing-page" has: section headings, feature descriptions, CTAs, pricing, testimonials, or proof points arranged in visual blocks.

      PASS 1 — PATTERN DETECTION
      Scan the content against ALL patterns, phrases, and structures from the references. For each match found, output one line:
      PATTERN_NAME | location (quote 5-10 words) | severity (HIGH/MEDIUM/LOW) | suggestion

      If no patterns found, output: CLEAN

      Content-type exceptions (landing-page ONLY — apply these 3 adjustments, keep everything else at full severity):
      1. SKIP Pattern 29 (Fragmented Headers) and S3 (Dramatic Fragmentation) — section labels like "Deploy." + description are standard landing page structure
      2. REDUCE Pattern 10 (Rule of Three) to LOW — unless the three items are vague adjectives like "fast, reliable, scalable"
      3. REDUCE Metronomic Sentences to LOW — uniform headline+description pairs are a layout pattern

      ALL OTHER PATTERNS remain at full severity for landing pages. Promotional language, vague claims, AI vocabulary, false agency, significance inflation, passive voice, etc. are still problems on landing pages. Be strict on these.

      PASS 2 — SCORING
      Rate content 1-10 on each dimension. Be strict — most marketing copy scores 4-6.

      For landing pages, adjust only these 3 dimensions:
      - Rhythm: Score 5 (not lower) if uniform short sentences match a headline+body section structure. But do NOT score higher than 6 unless there is genuine length variation.
      - Authenticity: Score based on whether the brand voice is distinct and consistent (not generic corporate). Do NOT require personal/first-person voice. But do NOT score higher than 6 unless the copy has genuine personality or takes stances.
      - Trust: Judge based on whether claims are SPECIFIC AND VERIFIABLE (named customers, concrete numbers, named sources like G2/Forbes), NOT whether source URLs appear inline. Landing pages link to case studies elsewhere. Score 5+ if most metrics use real numbers attributed to named entities. Only penalize: vague unverifiable claims ("industry-leading", "thousands of customers"), absolute superlatives without qualifier ("perfect, every time"), and metrics with zero indication of origin or scope.
      All other dimensions (Directness, Density): score exactly as for prose.

      OUTPUT FORMAT (follow exactly):

      CONTENT_TYPE: [landing-page|prose|social|email|other]

      PATTERNS:
      [pattern lines or CLEAN]

      SCORES:
      directness|N
      rhythm|N
      trust|N
      authenticity|N
      density|N
      total|NN

      VERDICT: [For landing-page: PASS if total >= 28, REVISE if total < 28. For prose: PASS if total >= 35, REVISE if total < 35.]

      GUIDANCE:
      [If REVISE: 2-3 sentences naming the 1-2 most impactful fixes. If PASS but under 35: note areas for improvement. If PASS 35+: "No revision needed."]
    as: audit_result

  # Step 3: Output
  - type: control.stop
    message: "{{audit_result}}"