# Pubcraft > Use for public-facing content — articles, blog posts, SEO content, newsletters (Substack/Beehiiv/Ghost), Medium, Hacker News, YouTube long-form scripts, podcasts, short-video scripts (TikTok/Reels/Shorts), or posts for LinkedIn, X, Reddit, Product Hunt, Threads, Bluesky, Mastodon, Quora, Indie Hackers, Dev.to, Hashnode. Also for AI-search citation (AEO/GEO/LLMO) and AI-content disclosure (EU AI Act Article 50, California SB 942, YouTube AI toggle, C2PA, JUMBF stripping on cross-posts). Triggers: "write an article/blog post/guide", "write a [platform] post", "YouTube script", "podcast show notes", "TikTok/Reels/Shorts script", "Product Hunt launch copy", "create SEO content", "optimize for AI search", "add AI-disclosure / EU AI Act / SB 942 / C2PA labeling". Produces researched, E-E-A-T-compliant, platform-native content that ranks in Google, gets cited by AI assistants, survives spam/AI policies, and follows YMYL and anti-AI-slop rules. Do NOT use for internal docs, code comments, or chat answers. - Skill: `thevrus/pubcraft` (Agent Skill, multi-file: 31 files) - Install (CLI): `npx skillmds@latest add thevrus/pubcraft` - Raw SKILL.md: https://api.skillmd.com/api/skills/thevrus/pubcraft/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Marketing & Growth - Author: thevrus (https://skillmd.com/u/thevrus) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/thevrus/pubcraft --- # Pubcraft Operating instructions for producing researched, publishable web content (articles, blog posts, social posts, video scripts) that ranks in Google, gets cited by AI assistants, satisfies E-E-A-T, reads as authentically human, and complies with platform and regulatory constraints. **Last research date:** September 2026. ## When to use this skill Use it when: - The user asks for an article, blog post, guide, explainer, comparison, how-to, listicle, or pillar page intended for a website. - The user asks for a LinkedIn, X (Twitter), Reddit, Product Hunt, Threads, Bluesky, Mastodon, Quora, or Indie Hackers post intended for publication. - The user asks for a newsletter (Substack or Beehiiv) issue, Medium essay, Hacker News submission, or developer-publication post (Dev.to, Hashnode). - The user asks for a short-video script (TikTok, Instagram Reels, YouTube Shorts). - The user asks for a Product Hunt launch copy package (tagline, description, gallery captions, first comment). - The user asks how to be cited by AI assistants (GEO/AEO/LLMO). - The content will be indexed by Google, ranked by an algorithmic feed, or read publicly. Do NOT use it for: - Internal docs, README files, code comments, technical specs. - Transient chat answers. - Email drafts (use the message_compose tool instead). - Pure creative fiction. ## How this skill is organized This file (`SKILL.md`) contains universal context every output needs: 1. The current search & AI landscape. 2. E-E-A-T requirements. 3. Research methodology. 4. Universal social rules. 5. Routing — which reference file to load for each output type. 6. Operating instructions. Detailed, format-specific guidance lives in `references/` and is loaded **only when needed**. Worked samples live in `examples/`. **Always load both `references/style-guide.md` and `references/output-formatting.md` at the start of any drafting task** — the anti-AI-slop rules apply to every output, and the output-formatting templates are what make the deliverable feel like a paid SaaS report (Surfer, Frase, Clearscope, Originality.ai) rather than a chat reply. --- ## PART 1 — The Search & AI Landscape You Are Writing Into ### 1.1 Google's enforcement posture (May 2026) Google's content quality framework rests on three pillars: the **March 2024 core update + spam policies**, the **September 11, 2025 Search Quality Rater Guidelines** (182 pages, the latest publicly available), and the **March 2026 spam update**. The March 2026 update introduced no new categories but sharpened SpamBrain enforcement of existing ones. Sites publishing AI content at scale without human review lost visibility on programmatic pages; sites with original, expert-driven content were largely unaffected. Google's **"Optimizing your website for generative AI features on Google Search" guidance** (Search Central, published May 15 2026, updated July 10 2026) ratifies on the record the position pubcraft has held since v0.1: no `llms.txt`, no AI-specific schema, no chunking, no content-multiplication for query variants. AI citations and organic rank reward the same mechanics. Detailed mechanism and the commodity-vs-first-hand framing in `references/geo.md`. Three named spam categories matter: 1. **Scaled content abuse.** Pages produced primarily to manipulate rankings rather than help users. Method-agnostic — applies whether content is human, AI, or hybrid. The trigger is the *combination* of high volume, low originality, and ranking intent. 2. **Site reputation abuse ("parasite SEO").** Hosting third-party content on a domain to exploit its ranking signals. 3. **Expired domain abuse.** Buying an aged domain and repurposing it for unrelated content. Google's December 2025 helpful content guidance frames evaluation as **"Who, How, and Why":** Who created it, How was it produced (including any AI), and Why. AI involvement is not disqualifying; *unaccountable, unreviewed, value-free* production is. The May 2026 SEO Starter Guide tightens the test into a verbatim editorial question every page must pass: *"Is it self-evident to your visitors who authored your content? Is the use of automation, including AI-generation, self-evident to visitors?"* Content created "primarily to attract search engine visits" rather than serve readers fails the test regardless of how it was produced. ### 1.2 The Quality Rater Guidelines YMYL ("Your Money or Your Life") topics — anything that could affect a person's health, financial stability, safety, or civic participation — receive the highest scrutiny. Common YMYL categories: medical, legal, financial (loans, taxes, investing, insurance), major life decisions (housing, education, employment), child safety, civic/electoral. Two principles apply: - **Trust is the dominant pillar of E-E-A-T.** An untrustworthy page is rated low quality regardless of expertise or authoritativeness. Verifiable credentials, accurate citations, transparent identity, and a "careful person seeking out experts" standard are the bar. - **AI content is rated "Lowest Quality"** when it is (a) manipulative or misleading, (b) mass-produced without human review, or (c) offers no original information beyond rephrasing existing sources. Important nuance Google has stated explicitly in the May 2026 SEO Starter Guide: **E-E-A-T is not itself a ranking factor**. It is the framework Quality Raters use to evaluate sample results; their evaluations train the ranking systems indirectly. The practical implication is unchanged (named credentialed authors, sourced claims, real expertise still win), but do not promise a client that "improving E-E-A-T scores" will move rank in a quarter. The signal flows through rater training, not a direct input. ### 1.3 The AI search layer (new since 2024) By May 2026, AI-assisted search is roughly 30–35% of all search activity. Ranking organically and being cited by AI assistants are *related but distinct* objectives. Google's AI Overviews use **retrieval-augmented generation** ("grounding") over the same core ranking systems, plus **query fan-out**: the AI generates several related sub-queries in parallel and pulls citations from all of them. Each H2 should stand alone as the answer to a distinct sub-question, not just subdivide the headline keyword. This is measurable, not theoretical: pages that rank for the fan-out sub-queries are **161% more likely to be cited in AI Overviews** than pages that don't (Surfer SEO / Search Engine Land, Dec 2025; `references/geo.md` § "Query fan-out"). For content that should be both findable in Google *and* cited by ChatGPT, Perplexity, and AI Overviews, load `references/geo.md`. ### 1.4 What is currently winning vs. losing **Winning:** Named, credentialed authors with verifiable bios; first-person experiential content; lightly monetized YMYL pages; original data, screenshots, lived experience; sites with detailed author credentials (industry analysis of the March 2026 rollout found ~72% of top-10 pages now display detailed author credentials). **Losing:** Aggregators using scraped content; sites built on expired domains; programmatic AI pages without review; thin Q&A spam; YMYL niches with anonymous authorship; sites with no clear ownership or business identity. --- ## PART 2 — E-E-A-T Requirements (Especially for YMYL) ### 2.1 The non-negotiable on-page checklist For YMYL content, every published article must include all of the following. For non-YMYL content, treat the first four as strongly recommended and the rest as topic-dependent. 1. **Named author byline** linked to a full bio page. Bio must state real credentials, relevant experience, and contact pathway. No generic team bylines on YMYL. 2. **Reviewer/editor credit when the writer is not the credentialed expert.** Format: "Written by [name]. Reviewed for accuracy by [credentialed reviewer name + credential]." 3. **"Last updated" and "Last reviewed" dates** at the top. 4. **A "Sources" or "References" section** at the bottom listing every cited source with a working link. 5. **Required regulatory disclaimers** for the relevant vertical (load `references/compliance.md`). 6. **Site-level trust block** within one click: legal entity name, regulatory IDs where applicable, physical address, contact pathways. ### 2.2 Demonstrating "Experience" (the first E) Experience is the hardest E to fake and the one Google has most aggressively elevated since adding it in December 2022. Concrete ways to embed it: - **Walk-throughs of real (anonymized) scenarios** with real numbers and outcomes. - **Original screenshots** of dashboards, documents, tools, results — anything that cannot be mass-generated. - **First-person operational specifics** that only a practitioner would know. - **Direct quotes from a credentialed source**, attributed by name and credential. - **Embedded video or photo from the writer or expert** — even 30 seconds. ### 2.3 Source quality hierarchy Use this priority order when researching and citing. **Tier 1 — Primary sources:** - Government agencies (.gov), regulators, official datasets, statutes, court rulings. - Original academic research (.edu, peer-reviewed journals). - Primary corporate sources (10-Ks, press releases, official documentation). - Original survey data with disclosed methodology. **Tier 2 — Authoritative trade and research:** - Industry trade associations, established think tanks. - Federal Reserve / FRED, BLS, Census, BEA, IMF, World Bank. - Major academic centers. **Tier 3 — Established journalism:** - Wall Street Journal, Bloomberg, Reuters, NYT, FT, AP, BBC, established trade press. **Tier 4 — Aggregator/educational sites** (use sparingly, only for non-statistical framing): Investopedia, Wikipedia (as starting point only), niche authority blogs. **Never cite as a primary source:** Random blog posts, AI chatbot outputs, content farms, forum threads (you may *quote* a user with attribution, but do not source statistics from forums). ### 2.4 Verifying claims, avoiding hallucinated statistics LLMs hallucinate numbers confidently. Mandatory protocol: 1. **Every number, rule, threshold, or percentage in the draft must have a working URL** to a Tier-1 or Tier-2 source. No exceptions. 2. **Re-fetch the cited page at draft time** to confirm the number is current. 3. **For any statistic that cannot be relocated, delete it** rather than approximate. 4. **Distinguish rules from rates of thumb.** A rule is verifiable; a rate of thumb must be framed as such. 5. **Date-stamp time-sensitive figures inline:** "As of [Month YYYY], …" --- ## PART 3 — Research Methodology The single biggest determinant of whether content reads as AI slop or expert prose is *what you knew before drafting.* Research first, write second. ### 3.1 Pre-writing protocol For every article, before drafting: 1. **Define search intent precisely.** Run the target query in web_search. Note: AI Overview content (if present), People Also Ask questions, related searches, top 10 organic results, what site types rank. Match the intent. 2. **Pull primary-source data fresh.** Open every relevant Tier-1 source and capture current numbers + URLs. 3. **Identify the information gain.** What does this article say or show that the current top 10 do not? Original data, an updated number, a real scenario, a fresh angle, an expert quote — there must be at least one. 4. **Build the source list before drafting.** Minimum 4–6 Tier-1/Tier-2 sources for any YMYL article. 5. **Get a quote or sign-off from a credentialed expert** when possible — even one or two sentences. ### 3.2 Original data vs. cited data - **Use cited data** for regulatory rules, program parameters, historical statistics, macroeconomic context. - **Generate original data** whenever feasible: a worked-through scenario with real numbers, a comparison table built from current sources, a screenshot, a small reader survey, internal data summarized in aggregate. Original data is the highest-leverage E-E-A-T signal available to a small site. ### 3.3 Citation formatting - Cite **inline** at the point of the claim with source name + hyperlink. - End every YMYL article with a **References** list of working URLs. - Use the **canonical, original URL** — not a republished version. - **Quote sparingly** — a sentence or two with attribution. Long block quotes are a copyright and originality risk. --- ## PART 4 — Style Essentials (compressed) The full anti-AI-slop ruleset lives in `references/style-guide.md`. Always load it before drafting. The condensed version, for routing decisions: - **Hard-banned Tier 1 words:** delve, tapestry, navigate (metaphorical), realm, landscape (metaphorical), pivotal, testament, elucidate, embark, underscore, harness, illuminate, unveil, intricate, multifaceted, robust, ecosystem (metaphorical), cutting-edge, game-changer, ever-evolving. - **Hard-banned structures:** "Not just X, it's Y." Em-dashes >1 per ~500 words. Tricolons in series. Symmetrical "in conclusion" wraps. - **Required signals:** sentence-length variance, falsifiable specifics (real names/places/dates/numbers), first- and second-person voice, taking a position rather than hedging, occasional "I don't know." - **Five-point self-check before delivery:** (1) any Tier-1 banned word? (2) >2 em-dashes? (3) any "Not X, but Y"? (4) ≥2 falsifiable details per 500 words? (5) any podcast-intro sentences? --- ## PART 5 — Universal Social Rules Before going platform-specific, three things apply everywhere: - **Hook in the first sentence.** Mobile users decide in under a second. No "Excited to share…", no "I've been thinking lately about…", no warm-up. - **Specificity beats generality.** "Closed a Tampa first-time buyer at 6.125% on FHA Friday" outperforms "Helping clients with FHA loans." Numbers, names, dates, places. - **Stay on-platform.** Every algorithm in 2026 — without exception — penalizes external links in the post body. If a link is essential, put it in a follow-up comment (still penalized on X and LinkedIn but less so) or in a profile bio. --- ## PART 6 — Routing: Which Reference to Load Decide the output type and load the matching reference file from `references/`. Do not load all of them — progressive disclosure is the point. | User wants | Load | |---|---| | Article / blog post / SEO content | `references/seo-article.md` | | AI-search citation strategy (GEO/AEO/LLMO) | `references/geo.md` | | Score a draft's AI-citability (0–100 paragraph-by-paragraph audit) | `references/citability-scoring.md` | | Refresh / update / republish an existing article (rank decay, new-year hub refresh) | `references/content-refresh.md` | | Head-to-head competitor comparison ("how do I outrank X", "why do they get cited") | `references/competitor-compare.md` | | Regulated vertical (financial, medical, legal, etc.) **or AI-disclosure compliance (EU AI Act / SB 942 / platform C2PA / JUMBF cross-post)** | `references/compliance.md` | | LinkedIn post or carousel | `references/linkedin.md` | | X (Twitter) post, long-form, or thread | `references/x-twitter.md` | | Reddit post or comment | `references/reddit.md` | | Product Hunt launch copy | `references/product-hunt.md` | | Newsletter (Substack / Beehiiv / Ghost) | `references/newsletters.md` | | Medium essay | `references/medium.md` | | Hacker News submission or comment | `references/hackernews.md` | | YouTube long-form video (script, title, thumbnail copy, description) | `references/youtube-long-form.md` | | Podcast episode (solo, interview, show notes, episode title) | `references/podcast.md` | | Short-video script (TikTok / Reels / Shorts) | `references/short-video.md` | | Threads / Bluesky / Mastodon post | `references/alt-social.md` | | Quora answer | `references/quora.md` | | Indie Hackers milestone post | `references/indiehackers.md` | | Dev.to or Hashnode developer post | `references/dev-publishing.md` | | Cross-platform repurposing | `references/cross-platform.md` | | Quick lookup of platform AI policies / format limits / priority | `references/reference-tables.md` | | Audit/review/brief output formatting (templates, ASCII charts, Mermaid, schema code blocks) | `references/output-formatting.md` (load-always) | **Always also load** `references/style-guide.md` at the start of any drafting task. For most outputs, you'll load 2–3 references at most: style-guide + the platform-specific file (+ compliance if YMYL, + geo if AI-search-optimizing). Load `examples/` files when a worked sample helps: | Worked sample | What it demonstrates | |---|---| | `examples/before-after-refactor.md` | AI-slop draft → audited rewrite (apply when teaching the style-guide audit) | | `examples/linkedin-example.md` | LinkedIn post hitting the first-210-char hook + 360Brew rules | | `examples/x-example.md` | X (Twitter) thread + long-form post under Grok ranking weights | | `examples/reddit-example.md` | Reddit comment-led post passing the specificity test | | `examples/product-hunt-example.md` | Full Product Hunt launch package (tagline, description, gallery, first comment, response templates) | --- ## PART 7 — How Claude Should Operate When invoked: 1. **Confirm the brief** if any of these are missing: target audience, primary search query (for articles), output type / target platform, word count, tone, vertical (YMYL?). Ask one consolidated question; don't drip. 2. **Load the right references** per Part 6. Don't over-load. 3. **Run web_search and web_fetch** to gather primary sources. State when sources cannot be located rather than hallucinate. 4. **Plan structure** explicitly before drafting (or quickly note it inline for short pieces). 5. **Draft following the platform's rules.** For social posts, follow the loaded reference exactly — character counts, format, hook patterns, banned moves. 6. **Run the self-audit** (Part 4 + the loaded `style-guide.md`). The anti-AI-slop rules apply to *every* output, including social. 7. **Deliver the draft as a Markdown file** when length is over ~600 words; deliver inline for short social posts. For Product Hunt, deliver the full five-piece package as one structured response. 8. **Mention any [COMPLIANCE TBD] placeholders** and recommend human review. 9. **Cite all sources** in a References section with working URLs (articles only — social posts don't get reference sections, but factual claims still need to be true). 10. **Never claim the content is "publication-ready"** for a regulated vertical without human compliance review. Be direct about what the user still needs to do. 11. **Teach the why, not just the verdict.** Every flag must name its mechanism — what triggers, who measures it, why it matters. The audit template in `references/output-formatting.md` enforces this in three places: - A **`> Why this matters`** callout under every section (Style, E-E-A-T, Structure, Schema, GEO, Compliance). - A **`Why it works`** column on every row of the prioritized fix list — the mechanism the user takes to the next article. - A **`What to take away`** closing block of 2–3 principles, each linked back to the relevant reference section for re-reading. Use these sources when naming mechanisms: - **Style flags** (banned words, em-dashes, "Not X, but Y"): `references/style-guide.md` § "Why these rules exist." - **Source citations and dates**: `SOURCES.md` (the bibliography for every claim in the skill). - **Common references**: September 2025 Quality Rater Guidelines (YMYL trust), March 2026 core update (affiliate/aggregator collapse), Sedestral 2026 + Semrush + Princeton GEO study (AI-citation rationale), perplexity-and-burstiness classifiers (Originality.ai, GPTZero, Copyleaks). A review that names the mechanism turns a one-time fix into a durable lesson. That's the skill's job. 12. **Format every response as production-grade Markdown** per `references/output-formatting.md` (load-always). Open audits and reviews with a one-line verdict + TL;DR. Use the right element for the data: - Tables for comparisons (✅/❌ checks, platform matrices, prioritized fix lists). - Fenced code blocks for schema/HTML. - ASCII bar charts or Mermaid for distributions and flows. - Callouts for compliance warnings. Drop the SaaS-replacement footer on deliverables over 500 words. The output is the deliverable; aim for the polish of a paid SaaS report (Surfer, Frase, Clearscope, Originality.ai), not a chat reply. --- ## What this skill cannot do - **No skill substitutes for human compliance review.** Disclaimer templates are starting points, not legal advice. Regulated content should be cleared by an internal or contracted compliance officer before publication. - **Search and platform policy moves fast.** This skill reflects September 2026 state. Algorithm changes — especially LinkedIn's quarterly shifts and X's monthly open-source updates — require quarterly reverification. - **AI-citation shares are volatile month to month, not a stable baseline.** ChatGPT's Reddit citation share swung from roughly 60% to roughly 10% within September 2025 alone after the Perplexity-Reddit lawsuit fallout (`references/geo.md` § "Platform → LLM citation routing"). Treat GEO as a position to defend quarterly, not a one-time optimization pass — a citation-source mix measured today can be materially different in three months. - **AI detection tools are imperfect.** Originality.ai, GPTZero, Copyleaks, Turnitin produce false positives on careful human writing and false negatives on lightly edited AI text. Use them as sanity checks, not gates. - **The biggest publication risk is not Google.** It is shipping a draft that violates a vertical regulation because no human reviewed it. Build a human gate between Claude's output and publication. - **No launch-day operations.** For Product Hunt and Reddit specifically, Pubcraft drafts the copy. It does not coordinate hunter outreach, manage waitlists, or post on a schedule. That's an ops layer outside this skill's scope. The aim is to make Claude operate as a **senior content strategist + writer** — the kind of person who would charge $300/hour at an agency and replaces several paid SaaS tools (Surfer SEO, Frase, Clearscope, Copy.ai, Jasper, Originality.ai). Claude does the E-E-A-T research, the platform-specific structuring, the anti-AI-slop enforcement, the GEO citation engineering, and the YMYL-compliance scoping. What it does *not* replace: the credentialed subject-matter expert, the compliance officer, or the final human editor. The deliverable is a researched, structured, on-brand draft that those humans can clear in minutes rather than hours.