Claude Skills for Content Writing: The Four-Skill Blog Engine

Claude skills for content writing are installable folders that package a single content competency — a system prompt, reference materials, and example outputs — so any Claude session can load it on demand. The best blog content stack assembles four of them: a brief skill, a draft skill, a humanizer skill, and a publish skill. Together they take a target query to a Sanity Draft without losing editorial control.

Anthropic shipped the Agent Skills primitive on October 16, 2026 (Anthropic engineering blog). The timing matters because the McKinsey State of AI 2026 survey found that more than 80% of companies adopting generative AI still report no meaningful improvement in productivity (McKinsey). The gap is not the model. It is the workflow. Claude skills for content writing close that gap by replacing one-shot prompting with a versioned, testable pipeline.

TL;DR

• Claude skills for content writing are installable folders, not prompts you re-paste.
• The four-skill blog engine has Brief, Draft, Humanizer, and Publish stages.
• Each skill enforces one publish-requirements gate so the post ships clean.
• Anthropic's Agent Skills primitive (October 2026) made this workflow portable across Claude, Claude Code, and the API.
• Skills complement Frase and Clearscope rather than replacing them outright.

Why Claude skills changed how blog content gets written

Through 2024, writing a blog post with Claude meant a long prompt, a long file paste, and a long conversation. Every project rebuilt the same scaffolding in chat history that disappeared after a week.

The Agent Skills primitive ended that pattern. A skill is a folder containing a `SKILL.md` file with a `name`, a `description`, and the body of instructions. Anthropic's public skills repository documents the format. Claude only loads the skill when the request matches the description, which is what makes the pattern scale.

Content writing is the canonical place to apply this. A blog post is not one task. It is a four-stage pipeline: decide what to write, draft against a structure, humanize the result, and publish into a CMS. Claude skills for content writing let you encode each stage once and reuse it across every post. See our primer on how to create Claude skills if the format is new.

Kevin Indig has made the parallel argument from the SEO side. In his 2026 reflection, he frames the shift as content engineering: modular systems, not one-off campaigns. Claude skills for content writing are the operational form of that idea.

The four-skill blog engine: brief, draft, humanize, publish

The most reliable Claude skills for content writing stack is four skills. The naming matters because composition is the unlock.

The Brief → Draft → Humanize → Publish skill stack maps cleanly to four stages most teams already run by hand. The Brief skill produces a structured `brief.json` with SERP gaps, an information gain score, an H2 outline, and a primary keyword. The Draft skill consumes that brief and writes the post against the ski ramp method. The Humanizer skill runs a cadence-and-cliché pass. The Publish skill enriches the post, runs QA gates, and writes a Sanity Draft.

Four skills, not one, for three reasons. Each stage has a different quality signal. You want to swap one without touching the others. The human review point sits between skills, not inside them.

This piece is the proof. The post you are reading was produced by the same four-skill blog engine. Brief, draft, humanizer, publish — four artifacts on disk that the publish script stitches together.

Inside the brief skill: SERP gaps, information gain, JSON contract

The brief skill is where most pipelines fall over. Run it badly and every downstream skill compounds the mistake.

It takes a target query and a persona. It scrapes the top ten SERP results, maps what every page already covers, and writes an intent-gaps array of the angles the SERP is missing. It then scores information gain on a five-dimension rubric (proprietary data, first-hand evidence, original framework, expert attribution, freshness) out of nine. Score below seven, the skill refuses to advance.

That refusal is the whole point. Brief stage is where you can change the angle cheaply. Once a draft exists, sunk cost takes over.

The output is a JSON contract with a fixed shape: target query, primary keyword, meta title (45–65 chars), meta description (140–165 chars), H2 outline, intent gaps, information gain plan, external link targets, and asset plan. A validator enforces the schema. This piece used the validator the same way our broader Claude skills for SEO cluster does.

A brief skill that returns prose instead of JSON is the most common failure mode. Downstream skills need machine-readable input, otherwise the draft skill becomes a prompt and the chain collapses.

Inside the draft skill: ski ramp opening and section orchestration

The draft skill is the part of the Claude skills for content writing stack everyone wants to start with, and the part you should write second.

It accepts the brief and produces a markdown draft. The opening must follow the ski ramp method: a 40 to 80 word direct answer AI Overviews can extract verbatim, a 60 to 100 word evidence paragraph with one stat and a named source in the first 800 characters, and a TL;DR of three to five markdown bullets. After that, one H2 per outline section, an H3 wherever a sub-point earns its own block, and bullets wherever a paragraph would otherwise read as a list.

The draft skill must enforce two structural rules most prompts ignore. The first is a minimum of two markdown tables per post. Tables earn citations from AI Overviews and serve as anchors a Claude or Perplexity answer can quote. The second is a GEO tactic per H2: a cite, a stat with attribution, a named-expert quote, or precise technical terminology. The Ahrefs Claude skills explainer makes the same point about citation-worthy structure.

A good draft skill writes one H2 at a time. It does not introduce new claims in later passes. Connective edits and humanizer work happen as separate skill calls. Mixing them inside one skill is how voice drift and fact drift get into the post.

Inside the humanizer skill: cadence, AI tells, and Flesch

The humanizer skill is unglamorous and load-bearing. It is what makes a Claude draft shippable without a full human rewrite.

Three checks do most of the work. The first is sentence-length standard deviation. A draft with stdev under six reads as flat AI prose, and the skill rewrites until variance hits six or higher. The second is an AI tell pattern library that flags the usual offenders: dead metaphors, hype verbs, and the hedge phrases that signal AI confidence. The set has about thirty entries. The third is Flesch reading ease, targeting a band of 50 to 70 that sits between HBR cadence and Justin Welsh's LinkedIn voice.

A fourth check, often missed, is em-dash budget. AI drafts overuse em-dashes the way humans overuse exclamation points. The Metaflow humanizer caps em-dashes at one per 400 words. The constraint forces clearer sentences.

The skill is not a black box. It outputs a metrics block: stdev, Flesch, tell count, em-dash count, filler-word count. The operator sees the numbers before approving the next pass. That visibility is the difference between trusting the skill and second-guessing it. Our piece on common Claude Code content mistakes catalogues the failure patterns this skill catches.

Inside the publish skill: enrich, QA, and Sanity write

The publish skill takes the humanized draft and produces a Sanity Draft document, ready for human review.

Three stages. Enrichment first: the skill injects internal links by topical relevance, embeds the brief's chosen YouTube video after the configured H2, and appends FAQ stubs from `faq_questions`. Then QA hard gates: a credible stat in the opening 800 characters, at least five internal links for any post over 1,500 words, at least three external links for BOFU posts (two for informational), primary keyword density between 0.5% and 1.5%, sentence-length stdev of six or more, and Flesch between 50 and 70. Any failure blocks ship. Finally the Sanity write: markdown converts to Portable Text (tables and the YouTube embed included), the hero image uploads to Sanity assets, and the document lands as a Draft for editorial review.

The Sanity write lives in its own skill for reversibility. A failed publish is one Studio click to delete. A failed draft is a wasted Claude session. Separating them keeps the cost of a mistake low. Our piece on how to build AI agents covers the orchestration patterns this publish skill follows.

How to install Claude skills for content writing today

Installation is the part everyone overcomplicates.

In Claude.ai, skills live under Settings, Capabilities, Skills. Upload a zipped folder containing `SKILL.md` and any reference files. In Claude Code, skills live in `.claude/skills/` inside your repo and the CLI auto-discovers them. In the Anthropic API, you pass a `skills` array on the message create call. The official Agent Skills documentation walks through each surface.

To bootstrap, use Anthropic's pre-built skills-creator. Enable it, paste a structured prompt describing the skill you want, attach reference files, and Claude returns a zipped skill ready to upload. The YouTube walkthrough above is the cleanest demo of that flow.

Test the skill against three real cases before trusting it. One known-good case where the right output is already documented. One adjacent case where the skill should adapt. One adversarial case where the skill should refuse. The test set is part of the skill, not separate from it. The same discipline shows up in our best Claude Code setup for marketing teams guide.

Version the skill in git. Tag releases. Roll back when a model update breaks the prompt. Treat the skills folder like code.

Claude skills for content writing vs Frase, Clearscope, and Surfer

Three SaaS tools dominate the content optimisation category. None of them are direct competitors to Claude skills for content writing. They cover one stage of the pipeline well.

Frase and Clearscope are excellent brief-stage tools. They surface SERP terms and content gaps faster than any DIY skill. Surfer's content editor is a strong drafting environment. None ship a humanizer or a CMS publish path.

The right pattern for most teams is hybrid. Run Frase or Clearscope as the SERP-research input into the Brief skill. Use the Draft, Humanizer, and Publish skills downstream. That keeps the SaaS spend where it earns its keep and the workflow ownership inside your repo. Improvado has documented similar hybrid Claude marketing stacks.

Claude skills for content writing win when you control the editorial standard. SaaS wins when you control the research surface. Treat them as a stack, not a choice. The full Metaflow library lives in our learning center.

Frequently Asked Questions

What is a Claude skill for content writing?

A Claude skill for content writing is an installable folder containing a `SKILL.md` file plus reference materials. It packages a single competency — drafting, briefing, humanizing, or publishing — that any Claude session can load on demand. Claude only activates the skill when the user request matches its description.

How do Claude skills for content writing differ from prompts?

A prompt is one-shot text you paste into a conversation. A skill is versioned, discoverable, and reusable across sessions and tools. Skills bundle reference documents (tone guides, sitemap CSVs, brand rules) plus example outputs. The model auto-loads the skill when needed, which removes the recurring paste-and-rebuild work prompts require.

Where do I install Claude skills for content writing?

In Claude.ai, upload a zipped skill under Settings, Capabilities, Skills. In Claude Code, drop the folder into `.claude/skills/` inside your repo. In the Anthropic API, pass a `skills` array on the message create call. The official Agent Skills documentation covers each surface.

Can Claude skills for content writing replace Frase or Clearscope?

For the brief stage, partially. The Brief skill can scrape the SERP, score information gain, and output a structured JSON brief. Frase and Clearscope are still faster for surfacing the long tail of SERP terms. Most teams run them together rather than substituting one for the other.

How many Claude skills do I need for a blog pipeline?

Four is the minimum: Brief, Draft, Humanizer, and Publish. Fewer means you are stitching prompts at one stage. More means you are over-decomposing the pipeline. Start with four. Add a fifth (distribution or repurposing) once the first four are stable.

How do I test a Claude skill for content writing before trusting it?

Run three real cases. A known-good case where the right output is documented. An adjacent case where the skill should adapt to a new vertical. An adversarial case where the skill should push back. Version the skill in git. Re-run the cases on every model update.