Article

article — generalized SEO article generator

This skill composes the upstream content-brief, write-content, and eeat-audit skills with site-specific brand context loaded from site-config.yaml, then emits a markdown file ready to drop into your content directory.

It is the generalized form of FakturaFlow's fakturaflow-article skill. The pipeline shape is the value; the FakturaFlow-specific bits are config.

Prerequisites

The upstream anthropic-skills plugin must be installed:

/plugin install anthropic-skills

If it isn't installed yet, tell the operator to run that command first, then re-invoke this skill.

Required pre-checks

Before starting, READ these files (they are the brand context and MUST be respected):

1. site-config.yaml at project root — brand voice, audience,

terminology, output path, language profile selection.

1. .claude/skills/article/language-profiles/.md — the

active language profile referenced by site-config.yaml. Falls back to english.md if unset.

1. / — existing articles. Don't duplicate angles.

Cross-link where relevant.

1. CLAUDE.md at project root, if present — supplemental project

conventions, do-not-do list, hard rules.

If site-config.yaml is missing, STOP and tell the operator to copy examples/site-config.example.yaml from this repo into their project root. Don't draft against defaults — you'll write the wrong voice.

Inputs

The operator provides ONE of:

1. A target keyword — e.g. "engineering onboarding checklist". The

skill picks the angle from the SERP.

1. A keyword + angle — e.g. `"engineering onboarding checklist, what

the first-week meeting cadence should be"`.

1. No input — read the project's keyword pool if one exists (common

paths: .claude/seo/keywords.md, content/keywords.yaml, seo/queue.md). Pick the highest-priority unwritten keyword. Confirm with the operator before drafting.

Pipeline

Run these steps in order. Each step uses the named upstream skill; inject the site-config constraints as additional system context.

Step 1 — content-brief

Invoke content-brief with:

• target keyword = the input keyword
• language = from site-config.yaml language field (defaults

to en)

• market = from site-config.yaml audience.market field
• search intent = let the skill auto-detect from the SERP
• product context = site-config.yaml site.name and

site.positioning

Capture the brief output: SERP gap analysis, recommended H2/H3 outline, target word count, must-cover entities.

Step 2 — write-content (CORE)

Invoke write-content with the brief from Step 1, plus these config-driven overrides:

Language and register:

• Language: from site-config.yaml language (default en).
• Register: from audience.register. Honor the value verbatim — do not

paraphrase to a "nicer" register.

• Pronouns: from voice.pronouns.
• Tone: must hit every adjective in voice.tone. Must avoid every item

in voice.avoid.

Terminology rules (HARD):

• Apply every terminology.preferred "use X, not Y" pair.
• Reject every entry in terminology.banned from the draft. If the

topic genuinely requires a banned term, STOP and ask the operator before continuing.

Pricing rules (if pricing block present in config):

• Pricing source of truth = site-config.yaml pricing block.
• Never invent prices. Never reference legacy prices.

Phasing dates (if phasing_dates block present in config):

• Phasing dates source of truth = site-config.yaml phasing_dates.
• Always quote the full sequence when relevant. Never paraphrase to a

single date when the reality is staged.

• This block exists for regulatory rollouts (KSeF, GDPR, accessibility

acts, etc.) where misquoting a date is a liability.

Brand positioning:

• Use site.positioning once, naturally, in the article. Not a sales

pitch. One paragraph, mid-article.

• Closing CTA: from cta.text linking to cta.url.

Article structure (defaults — config can override):

• Word count: 1,500–2,500 (write-content's default range).
• Lead paragraph (~80 words) framing the reader's question/problem.
• 4–7 H2 sections; H3 for sub-points where helpful.
• Concrete examples relevant to audience.primary.
• One natural product mention per article. Not a sales pitch.

Anti-AI-slop ruleset: The write-content skill's built-in anti-slop ruleset (banned vocabulary, structural pattern detection, the Horoscope Test) is mandatory. Do NOT skip it. The voice.avoid list in site-config is ADDITIONAL, not a replacement.

Language profile constraints: Apply every rule from the active language profile (loaded in pre-check step 2). These are language-specific style rules: punctuation density, diacritics, plural agreement, smart quotes, surface-conditional rules.

Step 3 — eeat-audit

Invoke eeat-audit on the draft. Check for:

• Experience signals: real, specific examples for the target

audience. Not generic.

• Expertise: correct technical details. Domain terms used

accurately.

• Authoritativeness: primary-source links where the draft makes

claims of fact (statutes, statistics, official documentation).

• Trustworthiness: disclaimers where appropriate (legal/medical/

financial topics). Source attribution.

Apply suggested fixes inline.

Step 3a — Regulatory fact-check (CONDITIONAL HARD GATE)

If site-config.yaml has regulatory_topic: true (e.g. legal, tax, medical, financial, compliance content), audit the draft for:

1. Dates — every date claim must match phasing_dates exactly OR

cite a primary source. Paraphrasing to a single date when the reality is phased is a must-fix.

1. Numbers — every penalty/fine/threshold must be cited with source.

No bare claims.

1. Statute or schema versions — current version only. Never claim a

future or deprecated version is current.

1. Official URLs — verify before citing.
2. Required disclaimer — if disclaimer_template is set in config,

append it to the article. Verbatim.

If a claim cannot be verified against a citable source, EITHER cite the source or remove the claim. If neither is possible, STOP and tell the operator. Do not auto-publish content with unverifiable regulatory claims.

Step 4 — Language review

Invoke the bundled language-review agent (in .claude/agents/copy-reviewer.md) with:

• The article draft
• The active language profile path
• The site-config terminology rules

Apply all must-fix and should-fix items. Surface nits to the operator in the final report.

Step 5 — Emit markdown file

Write the final article to /.md with frontmatter matching output.frontmatter_format. Slug = kebab-case of the keyword unless output.slug_style says otherwise.

Frontmatter formats:

astro / next (default):

---
slug: 
title: ""
description: ""
date: 
dateModified: 
readMin: 
category: ""
author: ""
authorBio: ""
---

jekyll:

---
layout: post
title: ""
date: 
categories: []
author: ""
description: ""
---

plain:

---
title: ""
date: 
---

Article body: plain markdown only. No JSX. Cross-references to other posts use path-based links ([text](/blog/other-slug)), not hash fragments — Google doesn't index hash-only URLs as separate pages.

Step 6 — Report

Tell the operator:

1. Slug + filename created.
2. Target keyword + detected intent.
3. Word count.
4. Any nits the language-review agent flagged that weren't auto-fixed.
5. Any claims the eeat-audit step flagged for human verification.
6. Suggested cross-links to existing posts (if found).

Do NOT auto-commit. The operator reviews the .md file and commits when ready.

What to do when stuck

If the keyword has no clear search intent for the configured audience (e.g. consumer queries when targeting B2B), STOP and tell the operator. Don't write thin content to fill quota — better to write 0 articles than 1 that erodes brand.

If write-content returns content that violates hard rules (terminology, banned vocabulary, language profile, phasing dates) and a single regen doesn't fix it, STOP and ask. Don't mechanically rewrite — the input might need adjusting.

If site-config.yaml is incomplete (missing voice.tone, missing audience.primary, etc.), STOP and ask the operator to fill it in before drafting. Drafting against defaults produces generic output that will read like exactly what it is.

Caching

The system prompt for the drafting step uses cache_control: ephemeral — sequential calls in one session reuse the prefix and cost ~10% of an uncached call. Keep multi-article batches in a single session to benefit.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: TrendTweekers
• Source: TrendTweekers/article-autopilot
• License: MIT

Install and usage instructions live in the source repository linked above.