content-qa

name	content_qa
version	1.0.0
description	Guardrail + plagiarism check for marketing drafts. Runs platform-specific policy rules, a banned-terms list, optional Originality.ai plagiarism scoring, and returns a structured verdict consumed by the n8n fan-out workflow (node "content_qa (OpenClaw)").
inputs	[{"name":"platform","type":"string","required":true,"enum":["facebook","zalo_oa","tiktok","linkedin"]},{"name":"body","type":"string","required":true},{"name":"topic","type":"string","required":false},{"name":"hashtags","type":"array","required":false}]
outputs	[{"name":"policyOk","type":"boolean"},{"name":"plagiarism","type":"number","description":"Similarity score 0..1 from Originality.ai (0 when check skipped)."},{"name":"risk","type":"string","enum":["low","medium","high"]},{"name":"violations","type":"array","description":"Array of rule IDs triggered."},{"name":"notes","type":"string"}]
requirements	{"tools":["http.request","fs.read"]}
limits	{"concurrency":4,"timeout_seconds":30}

Skill — content_qa

Rule sources (read once, cached)

1. rules/banned_terms.vi.txt — thương hiệu cấm, từ nhạy cảm, competitor.
2. rules/platform_limits.json — max length, max hashtags, disallowed formats.
3. rules/claims_whitelist.txt — các tuyên bố được phép (ví dụ "giảm 20%" chỉ hợp lệ nếu trong whitelist).

Pipeline

1. Normalize body to NFKC, lowercase for matching (keep original for return).
2. Platform limits
   • facebook: len(body) ≤ 5000, hashtags ≤ 5.
   • zalo_oa: len(body) ≤ 500, no URLs containing bit.ly|tinyurl|cutt.ly.
   • tiktok: len(body) ≤ 150, hashtags ≤ 8.
   • linkedin: len(body) ≤ 1500, hashtags ≤ 3. Fail → add rule PLATFORM_LIMIT, risk = high.
3. Banned-terms scan
   • Substring match against banned list. Hit → BANNED_TERM, risk = high.
4. Claim validation
   • Regex ((\\d{1,3})\\s*%|\\d+\\s*(nghìn|triệu|tỷ)) — mỗi khớp phải xuất hiện trong topic hoặc claims_whitelist; nếu không → UNVERIFIED_CLAIM, risk = medium.
5. Plagiarism (optional)
   • If ORIGINALITY_API_KEY is set, POST body to https://api.originality.ai/api/v1/scan/ai with timeout 10 s.
   • score_original < 0.6 → PLAGIARISM, risk = medium.
   • On HTTP error or missing key → plagiarism = 0, no violation.
6. Aggregate
   • policyOk = violations has no high-risk rule.
   • risk = max(level) across violations, default low.

Output contract

{
  "policyOk": true,
  "plagiarism": 0.08,
  "risk": "low",
  "violations": [],
  "notes": "banned=0; claims_ok=2; originality=0.92"
}

Wiring

• Called from n8n workflow 06-content-fanout node content_qa (OpenClaw).
• Request body schema: { platform, body, topic }. Extend with hashtags when editorial requests richer validation.
• Non-200 from OpenClaw → n8n continues via onError: continueErrorOutput; Shape draft + QA falls back to { policyOk: true, risk: 'low', notes: 'qa_skipped' }.

Failure modes

Symptom	Behaviour
Originality.ai timeout / 5xx	Mark plagiarism = 0, add note plagiarism_skipped
Rules file missing	Return policyOk=false, risk=high, note=rules_missing
Input too long (> 10 KB)	Return policyOk=false, risk=high, note=input_too_big

Maintenance

• Rules live in automation/openclaw/skills/content_qa/rules/. Review weekly during editorial sync. Commit changes to git so rollback is trivial.