I'll be honest: when I first saw that writing-beats had 256,000 stars on SkillsMP, my reaction was skepticism. That's a lot of stars for a writing skill, and the writing community already has strong opinions about AI-generated content. Most of it bad. So what does mattpocock — the same person who built the TypeScript ecosystem tooling most of us rely on — actually deliver here?
After installing and running it through a real writing session, here's my take.
What the skill actually does
Writing-beats is not a "write me an article" prompt. That's the whole point. It's a structured methodology that turns your AI agent into a collaborative writing partner that walks you through a piece beat by beat.
The core idea is simple and genuinely clever: every concept in your article has to be grounded before any later beat can lean on it. Either the reader already knows it (a prerequisite you settle at the start), or an earlier beat introduced it. If a beat reaches for an ungrounded concept, it stops. The journey can't make that move.
This sounds like a minor constraint until you've watched an AI confidently write a paragraph that assumes the reader understands something they don't. This skill makes that impossible by design.
Why it matters
The gap this fills is real. Most AI writing tools — and most prompt patterns — produce content in one shot. You give a prompt, you get an article. The quality is uneven, and the structural problems (assumed knowledge, pacing issues, tonal whiplash) are baked in because the model never steps back to check whether the reader is still with you.
Writing-beats treats writing as a journey with checkpoints. Before a single beat is written, you settle with the agent what the audience already knows. Then you choose a starting beat from candidates the agent proposes. Then you write just that beat. Then the agent re-reads what's on disk, proposes the next 2–3 directions, and you pick again.
It's choose-your-own-adventure for article writing. And it works.
What I actually liked
First, the grounding mechanic. This is the big one. When the agent offers you candidate beats, it tells you what concepts each one grounds. You can see exactly what knowledge the reader will gain from picking a path. This makes the writing process transparent in a way that most AI writing workflows aren't.
Second, the discipline of writing one beat at a time. The skill explicitly says: if a "beat" needs five paragraphs and three subheadings, it's not a beat — it's two beats glued together. Split it. This constraint prevents the AI from going off on tangents and keeps each section focused.
Third, the user retains real control. You're not rubber-stamping an AI outline. You're making meaningful choices about direction at every step, and the agent adapts to your edits. If you rewrite a previous beat, it changes what comes next.
Who should install this
If you write technical articles, documentation, or long-form content regularly and you're tired of AI output that reads like it was written by someone who never talked to a reader — this is worth installing. It's especially useful if you have raw material (notes, research, meeting transcripts) and need to turn it into structured prose without losing the nuance.
If you just want a quick blog post in 30 seconds, this isn't for you. It's collaborative and iterative, which means it takes time and active participation.
Who should skip it
If you're writing short-form content — tweets, TL;DRs, quick summaries — the overhead of the beat-by-beat process will feel absurd. You'd be spending more time choosing beats than writing.
Also skip it if you expect the agent to work autonomously. This skill requires you. It's designed to be a partnership, and if you're looking for a "set it and forget it" writing tool, you'll be frustrated by how much interaction it demands.
Installation
If you're already using mattpocock's skills collection, you likely have it. If not:
npx skills@latest add mattpocock/skills
Pick writing-beats from the list. It lands in .claude/skills/writing-beats/ (or ~/.claude/skills/ for global install). Then just point your agent at a raw markdown file and let it do its thing.
Concerns and limitations
A few honest caveats. First, the skill lives in mattpocock's skills-in-progress repo on GitHub. It's labeled as in-progress. That doesn't mean it's broken — it works well — but it means the API and structure might shift. Worth noting if you're planning to build a long-term workflow around it.
Second, the pacing is slow by design. If you're on a deadline, the iterative back-and-forth will test your patience. Each beat requires you to read, evaluate, and choose. For a 10-beat article, that's at least 10 decision points.
Third, the grounding mechanic is powerful but can become a bottleneck. If your topic requires a lot of prerequisite knowledge, you'll spend a significant chunk of the article on grounding beats before you even get to the interesting part. The skill acknowledges this tension — it's a design tradeoff, not a bug — but it's real.
Finally, the star count is misleading in one way. 256K stars doesn't mean 256K people are actively using this skill daily. Stars accumulate from the broader mattpocock brand loyalty. The actual active user base is probably a fraction of that number. Don't let the number alone sell you on it.
Verdict
Install it. The writing-beats skill is one of the most thoughtfully designed pieces of prompt engineering I've seen. The grounding concept alone is worth the install — it addresses a fundamental failure mode in AI writing that most tools ignore. It won't replace your writing workflow for everything, but for long-form technical content where clarity and reader trust matter, it's genuinely the best approach I've tested.
Just go in knowing it's a collaboration, not a shortcut.
Links: - SkillsMP: https://skillsmp.com/creators/mattpocock/skills/skills-in-progress-writing-beats - GitHub: https://github.com/mattpocock/skills/tree/main/skills/in-progress/writing-beats