Blog Writer Fleet v3.0

# /blog-writer-fleet — Blog Writer Fleet Team

> Version: 3.0 | Created: 2026-04-23 | Fleet: standalone
> Purpose: Produce industry-reference-quality articles with P0 regulatory citations, Angle Gate, and 3-Model QG

## Objective (O)

客戶品牌的目標受眾在用 ChatGPT、Perplexity、Google 搜尋該領域專業問題時，客戶品牌網站是被引用和排名最前面的來源 —
不是因為 SEO 技巧，而是因為每篇文章的法規引用密度和專業深度讓 AI 引擎和搜尋引擎都認定它是最可信的權威。

## Goal (G)

客戶品牌網站的 AI 爬蟲訪問量每週增加 20%，SEO 自然流量月增 15%，AI citation（被 ChatGPT/Perplexity 引用次數）每月 ≥5 次。

Team-specific angle: Through industry-reference-quality articles with P0 regulatory citations.

State Change: Reader trust level goes from "skeptical of another SEO blog" → "bookmarking this as a reference source."

## Fleet-Level KPI

| KPI | Baseline | Target | Frequency |
|-----|----------|--------|-----------|
| Articles published/week | ~1 (manual) | ≥2 | Weekly |
| Avg 3-Model QG score | ~6.5 (pre-SOP) | ≥8.5 | Per batch |
| P0 citation density | ~0.5/article | ≥3/article | Per article |
| AI crawler citation count | ~5/month | ≥15/month | Monthly |
| Persona Q7 YES rate | — | 100% | Per article |
| First-pass QG rate | — | ≥70% | Monthly |

## Phase Gate 1: After Agent 1 (Research Complete)

Condition: Claim Bank exists with ≥1 claim AND research context has ≥1 P0 citation.
Expected pass rate: ~95%

# Gate 1 check
CB_COUNT=$(python3 -c "import json; d=json.load(open('docs/blog-writer-fleet/{slug}/claim-bank.json')); print(d.get('returned',0))")
P0_COUNT=$(grep -cE "(NFPA|ADA|ANSI|IBC|UL).*[0-9]" docs/research/{slug}-context.md)
echo "Claims: $CB_COUNT, P0 citations: $P0_COUNT"
# PASS if CB_COUNT > 0 AND P0_COUNT > 0

If FAIL: Do NOT proceed to Agent 2. Fix research gaps first.

---

## Phase Gate 2: After Agent 4 (Quality Validated)

Condition: All 3 validator scripts PASS + all 3 Personas Q7 = YES + 0 CORRECTED + 0 UNVERIFIABLE.
Expected pass rate: ~70% first attempt. Average 1.3 attempts per article.

# Gate 2 check
python3 tools/validators/citation_backcheck.py {seo-draft} && \
python3 tools/validators/causal_inference_scan.py {seo-draft} && \
python3 tools/validators/trend_claim_scan.py {seo-draft} && \
echo "Scripts: PASS"

CORRECTED=$(grep -c "CORRECTED" docs/blog-writer-fleet/{slug}/fact-check.md)
UNVERIFIABLE=$(grep -c "UNVERIFIABLE" docs/blog-writer-fleet/{slug}/fact-check.md)
# PASS if all scripts exit 0 AND CORRECTED=0 AND UNVERIFIABLE=0

If FAIL: Return to Agent 2 or Agent 4 depending on failure type.

---

## Phase Gate 3: After Agent 6 (Score Approved)

Condition: 3-Model median score ≥ 8.0.
Expected pass rate: ~80% first attempt. Average 1.2 attempts per article.

# Gate 3 check
grep "Median" docs/blog-writer-fleet/{slug}/qg-scores.md
# PASS if median >= 8.0

If FAIL: Return to Agent 2 with feedback from lowest-scoring model. Max 2 full pipeline loops.

Effective timeline with retries: ~3.2 hours/article (vs 2.5 hours first-pass only).

### Agent 0: Team Leader

O: watersonusa.ai is the most cited, highest-ranked source for door hardware questions —
   because P0 regulatory depth earns trust from AI engines and search engines.

G: Every task output makes the target reader feel "this is the most professional,
   most credible answer I've ever read." Final 3-Model QG median ≥8.0.
   O alignment check: Leader can state in one sentence how the article serves O.
   Example: "This article helps facility managers choose the right fire door closer
   by citing NFPA 80 Section 6.1.3, directly building the 'most credible answer'
   trust the O requires."

S:
- Receive topic/task → Judge relevance to O
- Assign agents in correct sequence (1 → 2 → 3 → 4 → 5 → 6)
- Monitor Phase Gates 1, 2, 3 — intervene before failure
- Final acceptance: read as skeptical architect reading for first time
- Feedback loop: record one lesson to docs/blog-writer-fleet/leader-feedback.md

M:
- Final output 3-Model QG median ≥8.0 (Gate 3 passes)
- Leader can articulate O-alignment in one sentence
- If cannot → article not aligned → return to Agent 2

Model: Claude Opus (judgment requires highest capability)
Input: Topic string + slug name
Output: Orchestration decisions + final acceptance + feedback log entry

### Agent 1: Knowledge Researcher

O: Reader feels "every claim in this article has a paper trail."
   Research foundation so thorough no reviewer finds an unsourced statement.

G: Raw topic → structured Claim Bank JSON + P0 research context file.
   State change: "we have a topic" → "we have verified, citable evidence."

Model Routing:
- Claim Bank build: Script (no AI) — python3 tools/validators/build_claim_bank.py
- P0 standard research: Gemini Pro via /ai-fallback chain:
    gemini-2.5-pro → gemini-2.5-flash → claude-sonnet

M (Measures):
| Metric | Threshold | Verification |
|--------|-----------|-------------|
| Claim Bank exists | returned > 0 | build_claim_bank.py --topic "{topic}" --gt docs/waterson-product-facts.md --top 15 |
| P0 citation count | ≥1 P0 source | grep -cE "(NFPA|ADA|ANSI|IBC|UL).*[0-9]" {context} |
| Research file size | >1KB | test $(wc -c < {context}) -gt 1000 |
| Product mentions | ≤2 Waterson mentions | grep -c "K51M|K51L|Waterson" {context} ≤ 2 |

Input: Topic string + slug name
Output:
- docs/blog-writer-fleet/{slug}/claim-bank.json
- docs/research/{slug}-context.md (>1KB)

Error Handling:
- Claim Bank returns 0 → Topic too narrow. Broaden keywords OR run /product-knowledge-update.
- P0 research fails (all fallback exhausted) → STOP. Report to user.
- Ground truth <5KB → Run /product-knowledge-update first.

### Agent 2: SEO Draft Writer

O: Reader feels "this article understands my real-world problem and gives me
   a clear path forward with Waterson." Authoritative yet approachable.

G: Claim Bank + Research → 1000-1500 word narrative SEO article with Angle Gate PASS.

Model Routing: Claude Opus — direct session or subagent. Opus required for
   nuanced tone control and narrative quality.

M (Measures):
| Metric | Threshold | Verification |
|--------|-----------|-------------|
| Word count | 1000-1500 | wc -w < blog-seo-draft-{slug}.md |
| Claim Bank citations | ≥3 | grep -c "waterson-product-facts.md:L" {draft} |
| Human Layer markers | ≥2 | grep -c "HUMAN LAYER TODO" {draft} |
| Waterson density | ≥3/500 words | python3 density check script |
| Angle Gate | PASS | NEG=$(grep -ciE 'NOT:|don.t|wrong|mistake|avoid' {draft}); test $NEG -le 3 |
| Opening hook | Scenario/story NOT product spec | first H2 must NOT start with model number |
| CTA present | /solutions/ link at end | grep -c "/solutions/" {draft} |

Input: claim-bank.json + {slug}-context.md
Output: docs/blog-writer-fleet/{slug}/blog-seo-draft-{slug}.md

Error Handling:
- Word count off → revise. Max 2 revision rounds.
- Angle Gate FAIL → rewrite opening to positive framing.
- <3 citations → escalate to Agent 1.

### Agent 3: AEO Optimizer

O: AI engines (ChatGPT, Gemini, Perplexity) cite watersonusa.ai as primary source
   for door hardware questions.

G: SEO draft → structured AEO article (≤1000 words) optimized for AI engine extraction.
   State change: "human-readable article" → "AI-citable structured knowledge."

IMPORTANT: AEO is NOT traditional Q&A formatting.
  - Uses H2/H3 + paragraph narrative
  - Maintains flow while using semantic HTML and atomic facts
  - Consistent with EN/ZH content — same claims, same citations
  - Restructured for machine readability

M:
| Metric | Threshold |
|--------|-----------|
| Word count | ≤1000 |
| Format | Structured article, NOT Q&A list (no Q:/A: pattern) |
| Consistency | Same claims and P0 citations as EN |
| Waterson integration | ≥50% sections mention Waterson |

Input: blog-seo-draft-{slug}.md + claim-bank.json
Output: docs/blog-writer-fleet/{slug}/blog-aeo-draft-{slug}.md

Error Handling:
- >1000 words → trim. Keep all P0 + Waterson.
- Q&A format detected → rewrite as structured narrative.
- Inconsistent claims → cross-reference claim-bank.json.

### Agent 4: Quality Validator

O: Every published article survives adversarial review — 3 real-world personas find value,
   every fact checks out, 3 deterministic scripts find zero violations.

G: Draft articles → validated articles with 0 CORRECTED, 0 UNVERIFIABLE,
   3 Persona YES, 3 Script PASS.

Model Routing:
- Persona Review: Gemini Pro via /ai-fallback: gemini-2.5-pro → gemini-2.5-flash-lite → claude-sonnet
- Fact Check: Gemini Pro via /ai-fallback (same chain)
- Script Validators: Script (no AI) — deterministic Python

3 Personas (simulated by Gemini):
1. Architect — specifying door hardware for a new building project
2. Facility Manager — replacing failing closers in an existing building
3. General Contractor — looking for code-compliant, easy-to-install solutions

Each answers 7 questions. Q7 (verbatim): "After reading, do I want to look up Waterson's product?"

M:
| Metric | Threshold |
|--------|-----------|
| Persona Q7 | All 3 = YES |
| Fact Check | 0 CORRECTED, 0 UNVERIFIABLE |
| citation_backcheck.py | Exit code 0 |
| causal_inference_scan.py | Exit code 0 |
| trend_claim_scan.py | Exit code 0 |

Error Handling:
- Any Q7 = NO → Return to Agent 2. Max 2 loops.
- CORRECTED/UNVERIFIABLE → Fix using ground truth. Max 3 iterations.
- Script fails → Fix specific issue, re-run. Max 3. If still failing → STOP.

### Agent 5: HTML Publisher

O: Article appears on watersonusa.ai in 3 formats (EN, ZH, AEO) with correct
   meta tags, schema markup, sitemap entries — without breaking existing pages.

G: Validated drafts → 3 production HTML files staged in content-plan.
   State change: "validated markdown drafts" → "production-ready HTML staged for owner approval."

Model Routing: Claude Sonnet — HTML assembly + Publishing Strategy report.

M:
| Metric | Threshold | Verification |
|--------|-----------|-------------|
| EN HTML | exists | test -f blog/{slug}/index.html |
| ZH HTML | exists | test -f blog/zh/{slug}/index.html |
| AEO HTML | exists | test -f blog/web/{slug}/index.html |
| No noindex | ABSENT | grep -c "noindex" blog/{slug}/index.html = 0 |
| Content-plan | registered | grep "{slug}" content-plan.md ≥1 |
| NOT pushed to blog/ | staged only | git diff --cached --name-only does NOT include blog/ |

Owner Approval SLA:
- content-plan entry created → owner notified via email/Slack
- No response 48h → auto-escalate (reminder)
- No response 7d → "STALE — needs re-review before publish"

Error Handling:
- /publish-article fails → check draft format, verify template, retry once.
- noindex still present → quality gates inside /publish-article failed. Fix draft issues.
- Publishing strategy = BLOCKED → STOP. Report to user. Do NOT proceed to Agent 6.

### Agent 6: Score Gatekeeper

O: Only articles that independently impress 3 different AI models get published.
   No self-congratulatory scoring — each model evaluates blind.

G: HTML articles → 3-Model blind evaluation with median ≥8.0.
   State change: "we think it's good" → "three independent AI models confirm it's good."

Model Routing:
- Evaluator 1: Claude Sonnet — direct
- Evaluator 2: Claude Haiku — direct
- Evaluator 3: Gemini Pro — via /ai-fallback
All three evaluate the SAME article independently, blind to each other's scores.

Scoring Rubric (each evaluator):
| Dimension | Weight | What to score |
|-----------|--------|--------------|
| Technical accuracy | 25% | Claims verified? P0 citations present? |
| Reader value | 25% | Target audience learns something actionable? |
| Waterson positioning | 20% | Natural integration, not forced advertising? |
| Writing quality | 15% | Clear, professional, proper structure? |
| AEO readiness | 15% | Can AI engines extract and cite this content? |

M:
| Metric | Threshold |
|--------|-----------|
| Sonnet score | Recorded |
| Haiku score | Recorded |
| Gemini score | Recorded |
| Median score | ≥8.0 |
| All scores logged | 3 scores in qg-scores.md |

Error Handling:
- Median <8.0 → REJECT. Return to Agent 2. Max 2 full loops.
- One model unavailable → use /ai-fallback for replacement. Still need 3 scores.
- All models unavailable → STOP. Do NOT publish without 3-Model evaluation.

## Cross-Skill Dependencies

| This Skill | Depends On | Why |
|------------|-----------|-----|
| /blog-writer-fleet | /writing-guide | Agent 2 follows Waterson-primary voice rules, Angle Gate |
| /blog-writer-fleet | /ai-fallback | Agents 1, 4, 6 use multi-model fallback chains |
| /blog-writer-fleet | /publish-article | Agent 5 invokes Path A for HTML assembly |
| /blog-writer-fleet | /product-knowledge-update | Pre-check: ground truth must exist (>5KB) |
| /blog-writer-fleet | /content-scout | Upstream: topic discovery feeds into Agent 1 |
| /blog-writer-fleet | /w-emailcontent | Upstream: email mining feeds topics into this fleet |
| /blog-writer-fleet | /security-check | Must run before any git push in Agent 5 |
| /blog-writer-fleet | /blog-publish-sop | SOP wrapper that orchestrates this fleet step-by-step |
| /blog-publish-sop | /blog-writer-fleet | SOP reads this OGSM for agent definitions and interfaces |

## Capacity Model

| Mode | Throughput | Notes |
|------|-----------|-------|
| Solo operator | ~10 articles/week | ~2.5h/article, limited by API rate limits |
| Team (2-3 people) | ~15 articles/week | Pipeline overlap (A1 researches N+1 while A2 writes N) |
| Full parallel (5 instances) | ~25 articles/week theoretical | Limited by Opus quota, diminishing returns past 3 |