claude-blog: a 5-gate delivery contract for AI-written posts
Claude Code blog skill suite: 30 sub-skills, 5 agents, 5-gate v1.9.0 Blog Delivery Contract, dual-optimized for Google rankings and AI citations. Active development at AI-Marketing-Hub/claude-blog (AI Marketing Hub Pro community); public releases ship here.
At a glance
- What is it?
- AgriciDaniel/claude-blog is an MIT licensed Claude Code skill suite of 32 skill directories and 30 user-facing /blog commands, covering strategy, writing, schema, audits and multilingual publishing. Its defining feature is a delivery contract that scores every draft on a 100-point rubric and blocks delivery below 90.
- Who is it for?
- Adopt claude-blog if you publish blog content regularly through Claude Code and want the draft gated before you read it, since the 5-gate contract scoring against 100 points and blocking below 90 is the thing you cannot get from a custom prompt. It suits agencies running multi-post workflows across languages and brands.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 18, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the suite covers
claude-blog is a Claude Code skill suite that writes, optimizes, audits, localizes and refreshes blog content. The README describes it as a full-lifecycle engine for strategy, briefs, outlines, writing, rewriting, analysis, schema, AI citation readiness, site audits, topic clusters, multilingual publishing, audio narration and content decay detection.
That is a wider remit than a writing prompt. The pieces around the writing, decay detection and site audits, are the ones that assume you are running a blog over time rather than producing a single post.
The stated audience has three groups. Solo bloggers and creators, who get SEO, schema, source checks and internal linking without hand-holding each step. Marketing teams and agencies, who can run multi-post workflows through cluster, multilingual, persona, brand and discourse commands. And Claude Code skill builders, for whom the README explicitly frames this as a Tier 4 Agent Skills reference architecture with orchestrator routing, sub-skill dispatch, agent handoffs and code-enforced delivery gates.
It requires Claude Code. This is not a standalone tool and not a library you import.
The delivery contract
The part that separates this from a prompt pack is the gate. The README states the core promise as the user never being the first reviewer: a 5-gate Blog Delivery Contract scores every draft against a 100-point rubric, blocks delivery below 90, verifies artifacts and links, and iterates up to 3 times before escalation.
Read the numbers carefully. Ninety out of one hundred is a high bar, and it is a bar set by the project's own rubric rather than by any external standard. A draft that fails three times escalates to the user rather than shipping, which is the right behaviour and also means you will be reviewing failures.
The output side is specified too. Every draft ships as an artifact folder containing the markdown source, rendered HTML, PDF, a real hero image, three viewport screenshots, a review.md and a preflight-report.json. The renderer handles JSON-LD in an XSS-safe way and uses dark-mode-aware CSS, with one source for every output format.
That artifact convention is more useful than it sounds. A folder with a preflight report and screenshots is something you can review quickly and archive, and it is what makes the gate auditable rather than a score you have to trust.
Scale and architecture
The README publishes a component table, and the numbers describe a large system rather than a skill file.
There are 32 skill directories: one orchestrator at skills/blog/SKILL.md plus 31 sub-skills at skills/blog-*/SKILL.md. Thirty /blog commands are user-facing, and one sub-skill, blog-chart, is internal and not a command. Five specialized agents live in agents/blog-*.md. There are 14 root Python scripts, 22 references, 12 templates and 252 tests.
The orchestrator parses /blog input, detects the target platform, loads only the references it needs, routes to a sub-skill, and coordinates agents and scripts through the delivery contract. Loading only needed references is the part that matters at this size: a system with 22 references cannot load them all into context and stay useful.
The repository also carries a bundled Claude Blog Brain at ./brain, which version 2.2.0 describes as promoted to market-ready after an executable audit and an independent rubric review. The same release reviewed all 125 Brain source records and noted that URL availability alone cannot advance a review date, which is a pointed statement about how such knowledge bases rot.
Installing a Claude Code skill plugin
The README answers the install-path question directly and keeps the answer in the README on purpose, to demonstrate the pattern the suite generates: answer first, explicit paths, source-ready structure, and a compact section an AI system can quote without extra context.
The paths a user-installable Claude Code skill plugin should ship to are:
~/.claude/skills/<name>/
~/.claude/agents/<name>.md
~/.claude/scripts/<helper>.pySkill content goes under the skills directory, agent markdown under agents, and Python helpers under scripts. The repository root carries install.sh and install.ps1 for the two platforms, with matching uninstall.sh and uninstall.ps1, so the install is scripted rather than manual.
Python 3.11 or newer is required, per the README badge. The repository also carries a .mcp.example.json, so at least one workflow expects an MCP server to be configured.
Concretely, adopting this means installing a plugin into your Claude Code configuration and then driving it with /blog commands, not importing anything into a project.
The AI citation score is a heuristic
This is the correction worth knowing about, because it is easy to misread the marketing. The description mentions dual optimization for Google rankings and AI citations, and the README says every article is evaluated for Google-aligned usefulness and internal AI citation readiness heuristics.
Release v2.1.1 reframed that score. The notes say the existing AI score was renamed an AI citation readiness heuristic, explicitly not a calibrated probability and not a Google ranking factor. The same release reweighted the internal 0 to 100 scoring toward originality.
That is an honest correction and it changes how you should use the number. It is a signal for improving a draft, not a prediction that an AI assistant will cite your page and not a proxy for where Google will rank you. Nobody outside the search engines knows the second, and no tool should claim the first.
If you adopt this, treat the score the way the project now describes it, as a checklist that pushes a draft toward being quotable and well-sourced, and keep your own judgement about whether the article is worth publishing.
Where the development happens
The repository description states something adopters should read carefully: active development happens at AI-Marketing-Hub/claude-blog, inside an AI Marketing Hub Pro community, and public releases ship to this repository.
So the public repository is a release channel, not the workshop. Issues and pull requests you open here are against a mirror of the work rather than against the place the work is done, and the README points at GitHub Discussions for community contact.
That is a common arrangement for projects with a paid community tier, and it has consequences. Bug fixes may land privately first. Features you ask for may be prioritised against a paid roadmap. You can still fork under the MIT licence, which is the real safeguard here.
Version 2.2.0 was published on 2026-08-25 and the last push was on 2026-09-11, so the public repository is current even if it is not where the commits originate.
Writing your own skill as the alternative
The alternative for most people is a custom slash command or a short SKILL.md of their own, and the difference is enforcement.
A hand-written skill can encode your voice, your required sections, your citation rules and your publishing format in a few hundred lines. It will not score the draft, will not block delivery, and will not produce an artifact folder with screenshots and a preflight report. Whether that matters depends on whether you were going to read the draft anyway.
Commercial SEO content tools take the other position: they score content against a SERP analysis inside a CMS workflow, with recommendations on term coverage and structure, and they have no notion of an agent producing the draft. They assess what you wrote; this gates what the agent produced before you see it.
Choose claude-blog if you want the gate and you work in Claude Code. Choose a custom skill if you want control and a small surface. Choose a commercial SEO tool if your writing is human and your problem is on-page optimization.
Licence and upkeep
claude-blog is MIT licensed, with a LICENSE and a NOTICE file at the root, so forking is unrestricted and commercial use is permitted.
Maintenance is real work at this size. Thirty-two skill directories, five agents and 252 tests mean an upstream change to Claude Code's skill format can break several files at once, and the release notes show the project doing exactly that kind of upkeep: version 2.2.0 updated reviewed GitHub Actions to SHA-pinned current releases, corrected analyzer handling for JSON-LD @graph and HTML image alt text, and added Turkish readability support with language-aware analyzer reporting.
Version 2.2.0 also added weekly live Google-currentness checks with deterministic offline validation and synchronized public and Brain ledgers, which tells you the project treats its own source records as a maintenance burden worth automating.
The practical cost of adopting it is keeping up with that. Expect to re-run the installer after Claude Code changes, and expect the gate to reject drafts you thought were fine.
Editorial conclusion
Adopt claude-blog if you publish blog content regularly through Claude Code and want the draft gated before you read it, since the 5-gate contract scoring against 100 points and blocking below 90 is the thing you cannot get from a custom prompt. It suits agencies running multi-post workflows across languages and brands. Skip it if you do not use Claude Code, if you want a tool you can contribute to at the source, since active development happens in a separate community repository with public releases shipped here, or if you are tempted to read the AI citation score as a ranking prediction, which release v2.1.1 explicitly says it is not. Verify the gate on your own drafts first: run one article through it and read the preflight-report.json before deciding whether a 90 threshold is the right bar for you.
Frequently asked questions
Is Claude good for blog writing?
claude-blog is built on the assumption that it needs a gate. It scores every draft on a 100-point rubric, blocks delivery below 90 and iterates up to 3 times before escalating to the user, so the model writes and the contract reviews.
Is there a blog post about Claude Skills?
The README embeds a condensed generated article answering where a Claude Code skill plugin should install itself, as a specimen of the output format, and links to a post on the author's own blog about how claude-blog works.
How many commands does claude-blog provide?
The README table lists 30 user-facing /blog commands routed by the orchestrator, 32 skill directories made of one orchestrator plus 31 sub-skills, and 5 specialized agents.
Community notes