| name | auto-seo-writer |
| description | Use when the user wants to generate SEO/AEO articles with a guided flow. Triggers include writing SEO articles, auto-generating blog posts, batch content creation, or any request like "help me write an article about X." Runs a 7-step flow (topic → research → outline → case interview → write → quality check → publish); every step explains itself and offers options with a recommended default, so experienced users move fast and new users always know where they are. |
| argument-hint | ["topic or keyword (optional)"] |
Auto SEO Writer
Guided SEO/AEO article generator. One article = 7 steps. At every step,
tell the user in 1–2 sentences what this step does, then either offer options
(recommended default first) or announce-and-proceed. Experienced users just
take the default each time; new users always know where they are and where
they can intervene.
The 7 Steps
digraph auto_seo {
rankdir=TB;
"1 Topic & Setup" [shape=box, style=filled, fillcolor="#FFF3E0"];
"2 Research (announce)" [shape=box];
"3 Outline approval" [shape=diamond, style=filled, fillcolor="#FFE0B2"];
"4 Case interview" [shape=box, style=filled, fillcolor="#FFF3E0"];
"5 Write + humanize (announce)" [shape=box];
"6 Quality check & approval" [shape=diamond, style=filled, fillcolor="#FFE0B2"];
"7 Publish & report" [shape=doublecircle];
"1 Topic & Setup" -> "2 Research (announce)" -> "3 Outline approval";
"3 Outline approval" -> "2 Research (announce)" [label="re-angle"];
"3 Outline approval" -> "4 Case interview" [label="approved"];
"4 Case interview" -> "5 Write + humanize (announce)" -> "6 Quality check & approval";
"6 Quality check & approval" -> "5 Write + humanize (announce)" [label="revise"];
"6 Quality check & approval" -> "7 Publish & report" [label="approved"];
}
Interaction contract
- Steps 1, 3, 4, 6, 7 stop and offer options — put the recommended one first.
- Steps 2 and 5 are announce-only: say what you're about to do and with
which defaults, invite override（「要改就喊」）, then proceed without waiting.
- Never silently skip a step; if one has nothing to do (e.g. no case slots to
interview), say so in one line and move on.
Step 1: Topic & Setup
Explain: we need three things — what to write, for whom, where to publish.
Topic suggestions — this skill is generic. NEVER assume the user has
courses, videos, or any particular content. Source suggestions from what
you actually know about this user: project memory, earlier conversations in
this session, their site's existing content (via CMS MCP if connected). If
you know nothing about them, ask open-ended, or offer to brainstorm from
their niche. Suggested topics that build on their previous articles (topic
clusters, internal links) are worth flagging as such.
Audience & publish method: offer presets; if this user has generated
articles before, "same as last time" is the recommended default.
- Publish options depend on what's connected: WordPress MCP or zenbu site MCP
(tools like
zenbu_create_article) → offer CMS draft; always offer
"save as .md" (default path ~/Desktop/).
If the user gives no details: audience = general readers interested in the
topic; publish = .md to ~/Desktop/. Word count, article type, and keyword
strategy are auto-determined by research — do not ask.
Step 2: Research (announce-only)
Announce: what's about to happen（先掃你的第一手素材，再派三路研究，大約幾分鐘）
and the default depth. Offer the alternatives in the same breath, then proceed:
- Full（推薦）: first-party scan + 3 parallel agents (below)
- Quick: first-party scan + SERP agent only — for time-sensitive or
low-stakes posts
- Skip: write from existing knowledge — flag clearly that data points and
FAQ will be weaker
First-party material scan (inline, before launching agents):
Check whether the user has first-hand material on this topic: anything —
past articles, transcripts, notes, docs, product data, prior conversations.
Look in the working directory and project memory; only ask if a known source
is ambiguous. Whatever is found becomes the article's backbone and the
strongest E-E-A-T signal; external research supports it rather than
replacing it. Finding nothing is fine — Step 4 covers experience.
Agent A: SERP + Competitor Analysis
- WebSearch the primary keyword; record SERP features (Featured Snippet, PAA,
AI Overview, video carousel — mark inferences as inferences)
- Determine search intent (informational / commercial / navigational / transactional)
- WebFetch top 3-5 ranking articles; extract heading structure, word count,
subtopics, strengths, weaknesses, content gaps
Agent B: PAA + Community Questions
- WebSearch keyword variations: "[keyword] what/how/why/comparison/FAQ/recommended"
- Search English PAA if the topic has an English equivalent
- Search Reddit, PTT, Dcard, Threads, forums for real user questions
- Categorize: definition / how-to / comparison / reason / recommendation
Agent C: Data + Authority Sources
- WebSearch: "[topic] statistics [current year]", "[topic] research report", "[topic] trends"
- Prioritize: Tier 1 (government, academic) > Tier 2 (major-institution or
official reports) > Tier 3 (industry media — label as 業界估計 when cited)
- Record: data point, source, year, URL, credibility tier
- Cross-verify key stats with 2+ independent sources
Step 3: Outline Approval (stop for options)
Present, in this order:
Research Summary
================
Search Intent: [type] — users want to [...]
Competitors: analyzed X articles, avg word count Y
Key Gap: competitors lack [...]
Data Points: collected X citable stats
User Questions: mined X PAA questions
Our Differentiation: [one sentence on unique value]
Recommended word count: [X,XXX] (competitor avg + 20%)
Recommended format: [guide / listicle / comparison]
Case material status (one line per first-person slot):
✓ [section A] — covered by first-party material ([which source])
? [section B, C] — will interview you in the next step
Proposed Outline
================
# H1: [title with primary keyword, under 60 chars]
## H2 ... (question-format headings, definition block first, FAQ from PAA,
conclusion + action items last)
Options: 照大綱進行（推薦） / 調整章節（說哪裡） / 換切入角度重做研究。
Only proceed when approved.
Step 4: Case Interview (stop, conversational)
Explain in one line why this step exists: 真實案例是 E-E-A-T 的核心，編造的
經驗被讀者戳破一次就毀掉信任。Then interview — exploratory, NOT a form:
- Open broad. One question, all unsourced slots at once:
「這篇有幾個地方放你的真實案例會很加分：〔段落們〕。
你有沒有相關的經驗？想到什麼講什麼，不用完整。」
- Dig when they bite. Follow up like an interviewer — at most 1–3 short
questions per case（什麼時候？後來呢？有具體數字嗎？）. Stop as soon as you
have enough for one vivid paragraph. Never interrogate.
- Offer an out when they don't.
- 「那我幫你編一個？給我一個大方向就好」(user gives direction, AI drafts)
- 「或是我提一個方向，你看行不行」(AI proposes, then drafts)
- rewrite the slot as a non-first-person generic scenario, or drop it.
- Confirm before writing — always. Any case the AI drafted, or any detail
extrapolated beyond what the user literally said, must be played back
verbatim before Step 5:「這個案例我打算這樣寫：『……』這樣 OK 嗎？」
Only approved drafts enter the article. Once the user approves the concrete
text, it may be written in their first person — approval transfers ownership.
If every slot is already covered by first-party material, say so in one line
and move straight to Step 5. Keep the whole interview to a handful of turns.
Step 5: Write + Humanize (announce-only)
Announce: 開始撰文，套用的規則（AEO 直答、白話引註、去 AI 味）＋預設語氣與
字數，邀請覆寫（標準（推薦）／更口語／更專業；字數照研究建議或指定），then
write without waiting.
AEO Golden Rules
- First 40-60 chars of every H2/H3 section: direct answer (no filler intros)
- One data point per 150-200 words with source and year
- Question-format headings (what / how / why)
- Self-contained paragraphs (readable without context)
- Consistent entity naming throughout
E-E-A-T Signals
- Experience: first-person testing/operation descriptions (sourced per Step 4)
- Expertise: specific numbers, technical details, correct terminology
- Authoritativeness: cite credible sources (reports, official docs, institutions)
- Trustworthiness: mark data sources, update dates, present pros AND cons
Citation Style — plain language, never academic
Data points keep source + year (SEO needs them), but weave them into the
sentence like a person talking, not a paper:
- ✅ 「OpenAI 跟哈佛的團隊翻了 150 萬則真實對話（2025）」
- ✅ 「史丹佛和 MIT 的學者找了五千多名客服做實驗（2025）」
- ❌ 「（NBER Working Paper w34255，2025）」
- ❌ 「（Brynjolfsson et al.，《Quarterly Journal of Economics》，2025）」
Journal names, paper IDs, and "et al." never appear in the body — describe
who found it instead. Keep full academic references in research notes only.
Concrete artifacts — where they add value, not everywhere
Copyable prompts/commands, ❌/✅ before-after examples, real-life scenario
lists, and first-person stories are what make sections land — use them
generously in how-to and comparison sections. But do NOT force one into every
H2: a conceptual or transitional section is allowed to just explain. The test
is "would a reader try or feel something here?", not a per-section quota.
Real experiences only — NEVER fabricate silently
- Use ONLY experiences found in first-party material, told to you by the user,
or drafted-and-approved in Step 4. A story's existence in the material
does not license inventing its details — numbers, prices, file contents,
dialogue, "我的學員…" anecdotes are off-limits unless sourced or approved.
- Do not write while any first-person slot is unresolved.
Content Format
- H1 (unique) > H2 (main sections) > H3 (subsections) — never skip levels
- Paragraphs: 3-5 sentences; lists for 3+ parallel items; bold max 1-2 per paragraph
Required Elements
- Definition box at article start:
<div class="definition-box" style="background:#f8f9fa;border-left:4px solid #0073aa;padding:16px 20px;margin:20px 0;border-radius:4px;">
<strong>[Topic]</strong>: [40-60 char precise definition, AI-quotable format]
</div>
- FAQ section (5+ Q&As from PAA research)
- Schema Markup (JSON-LD): Article + FAQPage + BreadcrumbList
- SEO Meta: Title Tag (under 60 chars, contains primary keyword);
Meta Description (has CTA; length by script: English 150-155 chars,
Chinese 70-80 full-width — Google truncates CJK around 80 glyphs);
Focus Keyword; URL Slug (lowercase English, hyphens)
Humanize (same pass, do not skip)
Invoke the humanizer-tw skill if installed; otherwise apply directly:
remove opening clichés（「隨著…的發展」「眾所周知」）; cut excessive connectors
（「此外」「與此同時」「首先…其次…最後」）; replace internet jargon（賦能／痛點／
閉環）; fix translation-ese; make formal language conversational; vary sentence
rhythm; break formulaic structures; replace cliché endings; inject real voice.
Preserve all SEO elements: heading structure and keywords, data citations with
source+year, definition box, FAQ, Schema, and the AEO direct-answer openings.
Step 6: Quality Check & Approval (stop for options)
Present the full article (send the file) and the check:
Quality Check (X/18)
====================
Structure:
[pass/fail] H1 unique + contains primary keyword
[pass/fail] Heading hierarchy continuous (no skipping)
[pass/fail] FAQ section with 5+ Q&As
[pass/fail] Definition box present
[pass/fail] Conclusion + action items section
AEO Optimization:
[pass/fail] Every H2 opens with 40-60 char direct answer
[pass/fail] Question-format headings
[pass/fail] Data points have source + year (plain-language style, no
journal names / paper IDs in body)
[pass/fail] FAQPage Schema generated
E-E-A-T / Richness:
[pass/fail] First-person descriptions — every specific (number, price,
file content, anecdote) traceable to first-party material,
user-told facts, or a Step-4 draft approved verbatim;
nothing silently invented
[pass/fail] Credible external sources cited
[pass/fail] Author info configured
[pass/fail] Concrete artifacts (copyable prompt, ❌/✅ example, scenario
list, first-person story) present where they add value —
most how-to/comparison sections have one; not forced into
every H2
SEO Technical:
[pass/fail] Title Tag <= 60 chars
[pass/fail] Meta Description length OK (EN 150-155 / zh 70-80 full-width)
[pass/fail] URL Slug lowercase English
[pass/fail] Article + BreadcrumbList Schema generated
[pass/fail] No simplified or variant CJK glyphs — run a MECHANICAL scan
(regex for 们/后/发/这/换/说… and 羣/爲/裏 variants; exclude
intentional demo lines); eyeballing misses single glyphs
Score: X/18
If score < 15/18, auto-fix failing items before presenting. Additionally,
list any phrasings you synthesized from records (rather than the user's
literal words) so the user can veto them.
Options: 發佈（推薦，若全過） / 指定段落修改 / 整篇調性重修。
Only proceed when approved.
Step 7: Publish & Report
Explain where it's going, then execute per Step 1's choice:
Option A: CMS via MCP
Tool names below assume mcp__wordpress__*; use whatever CMS MCP the
session actually has. zenbu site MCP maps 1:1: zenbu_list_categories /
zenbu_create_category / zenbu_create_tag / zenbu_list_articles
(internal links) / zenbu_create_article (draft; slug, description and SEO
meta live on the article fields).
- Confirm target site; find-or-create category and tags
- Internal links: search existing posts for 2-3 relevant articles and
link them naturally in the body (skip silently if none)
- Create post as draft (title, HTML content, excerpt, slug, categories,
tags, SEO meta fields if supported)
- Report: post ID, URL, status
Option B: Save as Markdown
- Save to the chosen path (default
~/Desktop/[slug].md)
- Schema URLs: no domain yet — use the target domain if known from
conversation, else
https://example.com/[slug] placeholders in ALL Schema
URLs, plus a frontmatter comment:
# 發佈時請把 schema 中的 example.com 換成實際網址. Repeat this reminder
in the final report.
- Frontmatter: title / date / slug / keywords / description / schema (JSON-LD)
- Report: file path
Close with a one-line suggestion for the next article (topic cluster /
internal-link opportunities discovered during research).
Batch Mode
Multiple topics at once: run the same 7 steps, but batch the stops so total
interaction stays low:
- Step 1 once for the whole batch (one audience + publish method)
- Step 2 per topic (agents within each topic in parallel)
- Step 3 batched — ALL research summaries + outlines in one message,
numbered; user approves/adjusts per topic
- Step 4 batched — one interview covering all topics' open slots
- Steps 5 per approved topic
- Step 6 batched — every article's score + file; user approves per article
- Step 7 — publish all approved, report one table: topic / status / URL or path
Language
- Traditional Chinese (繁體中文)
- Technical terms stay in English (SEO, AEO, GEO, E-E-A-T, Schema, JSON-LD, H1/H2/H3)
- First occurrence of English terms: add Chinese in parentheses (e.g., AEO (Answer Engine Optimization, 答案引擎優化))
- Numbers: use Arabic numerals
- Tone: professional but approachable, no academic jargon
