The blog skill chain

The problem

A blog post drafted end to end by one AI agent fails in predictable ways: it states product facts from memory, cites nothing and reads as machine-made.

An AI agent that reviews its own blog draft checks it against its own blind spots, so its drafting errors pass review too.

Every Visualping blog post passes through one sequence of jobs, from keyword research to the CMS entry, and each manual handoff is a chance to skip a step.

The system

The blog skill chain is a set of Claude Code skills, each a written procedure with its own checks, that a pipeline skill called blog-pipeline runs in a fixed order:

1. seo-blog-writer pulls keyword and People Also Ask data from DataForSEO, studies the top 10 search results, flags Your Money or Your Life (YMYL) topics, drafts the post and audits it against a 23-item checklist.
2. humanizer-vp strips AI writing patterns and applies the house style: no em dashes, no negative parallelisms, no invented numbers.
3. blog-image-generator generates the images and has a vision model check the spelling of any text in them.
4. internal-linker adds links to related Visualping pages, ranked on Search Console and Ahrefs data.
5. verify-claims finds and links sources for uncited claims.
6. blog-quality-review scores the post and writes a rewrite brief for whatever fails.
7. strapi-blog-post loads the post into Strapi as a draft.

Five optional skills plug into the blog skill chain. The stat-lookup skill adds Visualping usage data and keeps a statistic only if it is framed as a sample and helps the reader do or decide something. The story-enricher skill offers a story, a recent case study and a quote from a named person, each with a checked source. The other three catch overlapping pages (cannibalization-detector), set exact image text (blog-image-text-overlay) and check product facts (vp-fact-guard).

The prompts inside the blog skill chain are Visualping-specific and stay private. The structure is the transferable part: fixed stages, one skill that owns each kind of fix, and gates a person holds.

The two keys

The first key in the blog skill chain belongs to the five skills that edit the file, from seo-blog-writer through verify-claims. The second key belongs to two checkers that never edit it, blog-quality-review and vp-fact-guard, and to me at three gates.

Gate 1 of the blog skill chain is draft approval. The pipeline stops at the finished draft, and the voice edit starts only after I approve it.

Gate 2 of the blog skill chain is a quality score of 80 or more from blog-quality-review, then my sign-off. The reviewer screens for YMYL risk, harmful content, purpose, accuracy and freshness, then scores 62 criteria in nine weighted categories. One em dash or negative parallelism marks a post DO NOT PUBLISH at any score.

When a post in the blog skill chain scores under 80, blog-quality-review writes a structured rewrite brief that names each failing criterion, the affected sections and the skill that owns the fix. Content gaps go to seo-blog-writer, AI patterns to humanizer-vp, thin linking to internal-linker and uncited claims to verify-claims, then the reviewer re-scores the post. Under 60, the pipeline stops. After one failed retry, I decide.

Claim verification is the soft gate in the blog skill chain. The verify-claims skill searches Exa for each uncited claim and links sources that score 60 or more on domain authority, claim match and recency. The post moves on either way and loses review points for claims left uncited.

The product-fact check in the blog skill chain is vp-fact-guard. It compares every claim-shaped sentence with a canonical table of Visualping product facts, covering plans, AI features, URLs and customer references. A blocking violation fails the post until it’s fixed, and a claim the table can’t settle comes to me marked UNVERIFIED.

Gate 3 of the blog skill chain is CMS confirmation. Once I confirm, strapi-blog-post blocks unresolved image placeholders and duplicate slugs, then checks that the SEO metadata saved. Every post lands in Strapi as a draft, and a person publishes it.

What it produced

The blog skill chain drafts posts for the Visualping blog, where signups that started on a blog post rose 47% in the 12 months to August 2026, against the 12 months to August 2024. The figure counts first landing pages only, so it is a floor, and it covers the whole blog program, so I don’t credit the chain with all of it.

Visualping blog posts also reach AI answer engines. When I asked ChatGPT, Perplexity and Claude “Who is Eric Do Couto?” on October 1, 2026, Perplexity and Claude both cited Visualping blog posts, and Perplexity named me Head of Marketing at Visualping from my author bio.

What I’d change

The blog skill chain trusts agents to run every stage, and in September 2026 I caught subagents skipping the optional skills and the draft gate, hand-applying Gate 2 fixes and estimating scores. I’d have every stage write a sidecar file with its results, as verify-claims and blog-quality-review already do, and have strapi-blog-post refuse a post with one missing.

I’d make claim verification in the blog skill chain a hard gate for numbers. Today an unsourced statistic costs a post points in review and can still pass. Blocking it at the CMS step would match my own standard: if a claim can’t be checked, it doesn’t ship.