anti-ai-blog-writer

Anti-AI Blog Writer

Produces blog articles that pass AI-detection tools while remaining strong at SEO and GEO. The skill orchestrates research, drafting (under the Oblique Copywriting Framework), image generation, AI-detector verification with rewrite loop, and markdown plus handoff delivery.

Cardinal-7 (must fire every run)

These seven rules are non-negotiable on every blog, in every mode. The detailed mechanics live in the workflow sections below; this is the always-on subset.

1. Score ≤10% AI on the primary detector (QuillBot). Loop-rewrite flagged passages, max 3 attempts, then escalate to Sean. Never silently relax the gate. Log the score in frontmatter.
2. Research only from primary sources. Regulators, government bodies, peer-reviewed work, official corporate filings, named experts. No content-marketing aggregator sites, forums, or AI summaries as primary citations. Every numeric claim carries a primary-source URL.
3. Follow the Oblique Copywriting Framework hard bans. The 10-point list in Step 4 applies to the whole body, no exceptions. No em dashes in the body (commas, colons or periods only).
4. Auto-load client context first. Silently check the vault for the target site or client before asking Sean anything (see Auto-Load Client Context below).
5. Ship the 4-image bundle (hero + 2 inline photographic + 1 data infographic) unless --no-images is set.
6. Ship a per-article handoff doc covering image placement, AI score, cross-link map and source list.
7. Write outputs to Claude Outputs/Blog Articles/ using the [NN]-[slug].md convention; respect --client-facing provenance (no visible Claude markers).

Auto-Load Client Context

Before asking Sean for a brief, run a silent vault check so the run starts from what already exists. Do not announce the search; only surface what you found when you state the run plan.

1. If a target site or client is named (or inferable from the topic), look for its client note in 10 - Oblique/Clients/ and check 10 - Oblique/Clients/Client Index.md.
2. Check Claude Outputs/Blog Articles/ for prior articles on the same site for the cross-link map, house voice and the topic taxonomy already in use.
3. For aiproperty.sg or property work, also read the Property Proposals Standing Instructions per CLAUDE.md.
4. Pull voice, positioning and ICP from the client note or 10 - Oblique/Oblique Brand.md if the blog is Oblique's own.

If nothing is found, proceed and ask only for what is genuinely missing.

When to use this skill

Trigger phrases:

• "write a blog", "draft a blog post", "write an article", "blog post about [topic]"
• "blog for [site/client]", "we need a blog about", "write SEO content for"
• "pillar article on", "long-form blog about", "anti-ai blog on", "anti-detection blog"
• "do a blog on [topic]", "write a blog for our website"

Also trigger when Sean uploads a topic brief, a discovery call note, or a competitive-research output and asks to turn it into a blog.

Do NOT use for:

• One-line social captions, headlines, or product descriptions (use copywriting)
• Internal SOPs or documentation (use docx for Word, or write directly)
• BrandedbySean spoken-word video scripts (use the BrandedbySean voice profile per CLAUDE.md)
• Already-published blog edits — for those, use copywriting to revise specific sections

The 6-step workflow

State a brief numbered plan to Sean before generating the article (per Karpathy Principle 4 in CLAUDE.md). Pause for redirect at step 1 if anything is ambiguous.

Step 1 — Parse the request, state the run plan

Confirm or infer:

• Topic / working title. Specific enough that two competing blogs on the same topic wouldn't read identical.
• Target audience. Who reads this and what they walk away knowing.
• Length mode. Default: pillar (1,500–2,500 words). --short flag: 800–1,000 words.
• Image bundle. Default: on (4 images). --no-images flag: skip.
• Detector choice. Default: quillbot. Options: zerogpt, both.
• Target site / client. For cross-link resolution and provenance markers (--client-facing strips visible Claude markers).
• Primary keyword + 2–4 secondary keywords if SEO-driven; secondary if the brief is more editorial.

State the plan in numbered form:

1. Research: [list of primary sources you'll consult] → verify all numbers current
2. Outline: [list of section headings, word-count targets, cross-link slots]
3. Draft: [length target] under Oblique Copywriting Framework
4. Generate images (in background): hero + 2 inline + 1 infographic
5. AI-detector loop: target ≤10% on [detector]. Rewrite flagged passages up to 3x.
6. Deliver markdown to Claude Outputs/Blog Articles/ + handoff doc.

Wait briefly for Sean to redirect. If silent or "go", proceed.

Step 2 — Research from primary sources

Apply research-sourcing skill rules: every factual claim cited, sources logged at the draft footer.

What counts as primary:

• Regulator and government sites (IRAS, MAS, URA, HDB, MOF, MOE, CPF Board, LTA, IMDA, ACRA, etc. for Singapore; equivalent bodies for other jurisdictions)
• Peer-reviewed journals
• Official corporate filings, annual reports, press releases
• Named expert publications and interviews
• URLs ending in .gov, .gov.sg, .edu, .edu.sg, .org for established institutions

What does NOT count:

• Content-marketing aggregator blogs ("homejourney.sg" style sites that paraphrase primary sources)
• Forum threads, Reddit, Quora
• AI-generated summaries from other sites
• Sites without an author byline and date

Cross-check every number against at least one secondary source. If they disagree, surface the disagreement in the draft (don't pick silently).

Verify gate: every numeric claim has a primary-source URL. Sources dated within the last 18 months unless the fact is historical (e.g., a policy change date).

Step 3 — Outline

Write the outline before drafting. The outline should list:

• Section headings (H2 level), in order
• Approximate word count per section
• Key data points each section will use
• Cross-link slots (which other articles or pages this links to)

Verify gate: outline word count totals to within ±10% of the target range.

Step 4 — Draft

Apply the Oblique Copywriting Framework hard bans below. Write the body. The opening hook is the most-flagged passage in AI detection — give it the most rhythm variation and the most specificity.

Anti-AI hard bans (inline from CLAUDE.md)

1. No antithetical pivots. Never write "not just X — it's Y" / "this isn't just A; it's B." Say it directly.
2. No filler openers. Never open with "In today's world…", "Now more than ever…", "In a crowded marketplace…", "At the end of the day…", "It's worth noting that…"
3. No hollow power words. Avoid: game-changer, revolutionary, transformative, disruptive, world-class, cutting-edge, synergy, leverage (as a verb), seamlessly, holistic, bespoke, robust, streamline, elevate, foster, landscape, navigate (metaphorically), delve, tapestry, multifaceted, nuanced (as filler).
4. No em dashes in the body. Period, colon, or comma. Em dashes are allowed only inside citation entries in the Sources block.
5. No restating the same point twice. No "in other words," "what this means is," or any sentence that confirms the one before it.
6. No explaining the hook. A strong opening line doesn't need a defensive follow-up.
7. No bullet points in prose unless the content genuinely requires a list.
8. No excessive bold, headers, or formatting in flowing paragraphs.
9. No performative fragmentation. No "Twelve sections. Thirteen answers." style number-juxtaposition. No subject-dropped verb-first fragments. Full sentences with explicit subjects.
10. Presenter-style headlines. Question-led ("Where we see opportunity"), phrase-led ("Three deliverables at a glance"), or full-sentence headlines. No billboard-style two-clause headlines outside of actual ad copy.

Specificity rules

• Numbers and outcomes over adjectives.
• Named cases, specific dates, real organisations.
• Vary sentence length aggressively. If three sentences in a row are the same length, vary them.
• The Competitor Test: could someone copy-paste this sentence into any other AI chat and still have it make sense? If yes, rewrite to be more specific.

Verify gate: scan the full draft against the 10-point hard ban list before moving to step 5. Reject and rewrite any violation.

Step 5 — Generate images (background, runs in parallel with step 6 if scheduling allows)

Skip entirely if --no-images.

The 4-image standard bundle:

1. Hero (Nano Banana 2, 16:9, 2K). Editorial-photography aesthetic. No people unless contextually essential. No text overlays, no logos, no recognisable brand names.
2. Inline 1 (Nano Banana 2, 16:9). Supporting photographic context, complements the section it sits next to.
3. Inline 2 (Nano Banana 2, 16:9). Same rules. Different angle / subject from inline 1.
4. Data infographic (GPT-Image-2 gpt-image-2-text-to-image, 16:9, quality: high, size 1792x1024). The article's central data point rendered as clean typography on cream off-white. No clip art, no decorative icons, just text and minimal layout. Use muted terracotta highlights for emphasis.

Use the batch script at templates/generate-images.js. Run with run_in_background: true so the AI-detector loop can proceed in parallel.

Verify gate: all 4 files exist at Claude Outputs/Blog Articles/images/[slug]/ and each is >100 KB.

Step 6 — AI-detector verification loop

Run the body through the chosen detector (default QuillBot via Chrome MCP). The plain-text body for testing should be:

• Markdown-stripped (no ## headers, no | table delimiters, no ** bold markers)
• Body only — exclude YAML frontmatter, Further Reading, and Sources sections
• Save the plain-text version to Claude Outputs/Blog Articles/images/[slug]/quillbot-body.txt for repeatability

QuillBot flow (full mechanics in templates/quillbot-detection-flow.md):

1. Free anonymous tier caps at ~1,200 words per run. If the body is over 1,200 words, split via templates/split-body-for-quillbot.js into Chunk A (≤1,050 words) and Chunk B (remainder).
2. Copy the chunk to OS clipboard via PowerShell: Get-Content [path] -Raw | Set-Clipboard.
3. Navigate Chrome MCP to https://quillbot.com/ai-content-detector. Click the editor textbox (coordinate-based click, around the centre of the visible textbox). Send Ctrl+V. Wait 3 seconds.
4. Click Detect AI button. Wait 6–8 seconds. Read the score from the result panel.
5. If the QuillBot page hangs or rate-limits (login modal appears, page shows spinner indefinitely): clear cookies via JS (localStorage.clear(); sessionStorage.clear(); plus a cookie-loop deletion), reload the page, retry.
6. If QuillBot is rate-limited after cookie clear, fall back to ZeroGPT at https://www.zerogpt.com/ (full-body, 15,000 char limit).

Rewrite loop:

• If above 10% AI: identify flagged passages (highlighted in QuillBot's result panel). Rewrite them for more sentence-rhythm variation, more specific detail, and more first-person voice where natural. Preserve all facts.
• Re-test the same chunk. If still above 10%, rewrite again.
• Max 3 rewrite attempts. After 3, escalate to Sean with the flagged passages and the current score.

Note on detector calibration:

• QuillBot v7.1.0 (current at time of writing) calibrates conservatively. Articles with parallel declarative structure typically score 0–6% AI.
• ZeroGPT calibrates aggressively. The same content often scores 25–50% AI because ZeroGPT treats polished policy-prose as AI-like. Treat ZeroGPT as supplementary, not authoritative, unless --detector zerogpt is explicitly chosen.

Verify gate: the final AI score is ≤10% on the primary detector. Score is logged in the frontmatter ai_detector_score field.

Step 7 — Deliver

Save the article to Claude Outputs/Blog Articles/[NN]-[slug].md where [NN] is a two-digit batch number (use 01 for single articles).

The frontmatter must include:

---
title: "..."
slug: ...
topic: ...                  # one of the available topics on the target site
excerpt: "..."              # ~160 chars
publishedAt: YYYY-MM-DDTHH:MM:SSZ
coverImage: images/[slug]/hero.jpg
inlineImages:
  - images/[slug]/inline-1.jpg
  - images/[slug]/inline-2.jpg
  - images/[slug]/infographic-[descriptor].png
generated_by: claude
generated_date: DD/MM/YYYY
source_skill: anti-ai-blog-writer
ai_detector_score: "X% (QuillBot v7.1.0, DD/MM/YYYY, [details])"
---

Write a per-article handoff at Claude Outputs/Blog Articles/[NN]-HANDOFF.md using templates/handoff.md as the base. Include:

• File map (article + image filenames)
• Suggested image placement (where each image goes in the article body)
• AI-detector check table (which chunks were tested, scores, rewrite count)
• Cross-link map
• Source list summary

Verify gate: article markdown opens cleanly, all 4 image references resolve to existing files, handoff doc is in the same folder.

Defaults and flags

Flag	Default	Effect
length	pillar (1,500–2,500 words)	--short = 800–1,000 words
images	on (4 images)	--no-images skips image generation entirely
detector	quillbot	--detector zerogpt or --detector both
AI gate	≤10% AI	hardcoded, do not relax without Sean's explicit override
output	Claude Outputs/Blog Articles/	per CLAUDE.md Output Conventions
provenance	internal markers in frontmatter	--client-facing strips visible markers, keeps metadata only

Failure modes and fallbacks

Failure	Fallback
QuillBot rate-limited / login modal blocks	Clear cookies + localStorage, reload, retry. If still blocked, switch to ZeroGPT and log both that QuillBot was unreachable and the ZeroGPT score with the calibration caveat.
Image gen network blip (one or more files fail)	Re-run the generate-images.js script. It skips files that already exist on disk and only retries the missing ones.
AI score stays above 10% after 3 rewrite attempts	Stop. Escalate to Sean with: the current score, the highlighted passages, the rewrite history. Do not silently lower the gate.
Primary source unavailable for a key claim	Surface the gap. Either rewrite around it or ask Sean for guidance. Do not fabricate a citation.
Length undershoots after first draft	Add depth via worked examples, additional context, expanded methodology — not filler. If you can't reach the target without filler, the topic isn't pillar-sized; recommend --short.
Length overshoots	Cut the weakest section. Pillar quality matters more than hitting 2,500 exactly.

Provenance rules (per CLAUDE.md)

Internal-only files (drafts staying in the vault): YAML frontmatter is fine. generated_by: claude, generated_date, source_skill: anti-ai-blog-writer.

Client-facing deliverables (--client-facing flag): NO visible "Built by Claude" markers anywhere — not in body, not in footers, not in signatures. Use document metadata (the YAML frontmatter is internal; if the body is being copied into a CMS like WordPress or aiproperty.sg admin, the frontmatter doesn't ship). Folder location (Claude Outputs/) and filename convention provide the system-level marker.

Composed Skills

This skill orchestrates rather than reinvents. Layer 1 primitives and meta/gate skills it calls:

• imagery and nano-banana (Layer 1) — image generation, invoked via the batch script. The hero and inline photos use Nano Banana 2; the data infographic uses imagery. Don't manually call them per-image; let the script orchestrate.
• research-sourcing (meta/gate) — automatically applied when web research is performed; every claim cited.
• copywriting (meta/gate) — voice and Oblique rule enforcement; the Step 4 hard bans are the inline always-on subset. Scope: the full article body, headlines and excerpt are customer-facing copy and pass this gate.
• claude-in-chrome MCP (tooling, not a skill) — QuillBot / ZeroGPT detection automation.

Do not duplicate these skills' logic. If something in this skill's spec drifts from theirs, update this one to align.

Templates folder

All in templates/ next to this SKILL.md:

• frontmatter.yaml — YAML frontmatter template with required and optional fields
• handoff.md — per-article handoff doc template
• generate-images.js — Kie.ai image batch script (Node.js). Customise ITEMS per article. Run via node generate.js with KIE_API_KEY env var set.
• split-body-for-quillbot.js — body chunker. Splits at a paragraph boundary near the 1,050-word mark.
• quillbot-detection-flow.md — step-by-step Chrome MCP flow for the QuillBot detector, with screenshots of common UI states (login modal, rate-limit spinner, result panel layout) and the recovery patterns.
• ai-tell-ban-list.md — full Oblique Copywriting Framework hard bans, copy of the inline list above plus extended examples.

Maintenance

Dated history and migration/validation notes live in CHANGELOG.md.

Per CLAUDE.md, after modifying this skill, keep Skills Map.md and AI Vault Map.md in sync.

If QuillBot's model version changes or their UI changes such that the Chrome MCP flow breaks, update templates/quillbot-detection-flow.md accordingly. If ZeroGPT or any successor detector becomes the new authority, the AI-detector loop in step 6 changes; everything else stays.