Matt Pocock's skills repo just hit 264,000 stars on SkillsMP, and "writing-shape" is the breakout of the week — 7,682 stars in the last seven days alone. If you've been ignoring the SkillsMP marketplace, this is the one that should stop you.
I installed it, ran it against a messy transcript of a technical talk I'd been sitting on for weeks, and the result was the most honest writing session I've had with an AI agent. Here's my assessment.
What this skill actually does
"writing-shape" takes a pile of raw material — a transcript, a wall of notes, a jagged list of fragments — and shapes it into a finished article, paragraph by paragraph. It does not generate content from a prompt. It does not summarize. It treats the writing process as an engineering problem: establish prerequisites, draft candidate openings, then grow the piece one grounded block at a time.
The word "exploit" in the description is deliberate. The exploring is already done. You've committed to a structure, and now you mine the pile to fill it. The raw material file is read-only. The skill never edits your source — it only reads from it and writes to a separate article document.
Why this matters
Most AI writing tools solve the wrong problem. They take a prompt and spit out a draft, which is essentially autocomplete on steroids. The gap they leave isn't quality — it's structure. You get prose that reads fine in isolation but has no architectural logic. No grounding. No deliberate progression.
"writing-shape" fills that gap by forcing a conversation about the article's structure before it writes a single paragraph. It asks what the reader already knows, what they need to hear next, and whether each block earns its place. That's not a writing tool — it's a writing partner who refuses to let you skate by with weak transitions.
Key capabilities I found useful
-
The grounding system. Every concept in the article must be grounded — either the reader brings it as a prerequisite, or an earlier block introduced it. The skill keeps a running list of what's grounded and won't let you build on sand. If a block reaches for an ungrounded concept, it names the gap. This is the single most intellectually honest feature I've seen in any AI writing workflow.
-
Forcing you to pick an opening. Before writing anything, the skill drafts 2–3 candidate openings, each implying a different thesis or angle. It makes you choose or compose a hybrid. The chosen opening locks in what the rest of the article must do. Most writers skip this step and wonder why their piece feels directionless.
-
Paragraph-by-paragraph growth with format arguments. After the opening lands, the skill asks "what does the reader need to hear next?" and pulls material from the pile to answer. But it doesn't just write — it argues about format. Prose or list? Table or bold leads? Callout or inline? Each format choice is weighed out loud with you, not decided silently.
-
The grilling conversational style. The skill pushes back. "What does this paragraph do that the previous one didn't?" "If I cut this, what breaks?" "The opening promised X and we drifted to Y." It's a inverted ideation session — instead of asking what you're noticing, it asks what the article is actually arguing and in what order the reader needs to hear it.
-
Gap naming instead of fabrication. If the pile lacks something the article needs, the skill says so explicitly: "We need an example here and the pile doesn't have one. Give me one or we cut this section." It won't hallucinate content to fill a hole. That alone separates it from every other AI writing tool I've tested.
Who should install this
If you regularly turn notes, transcripts, or rough drafts into published articles — technical writers, blog authors, newsletter producers, documentation leads — install this immediately. It's particularly useful for anyone who has raw material sitting in a file and knows it could be a good piece but can't get past the structural mess.
If you want AI to write your blog post in one shot from a single prompt, this skill will frustrate you. It requires engagement. It expects you to make decisions. It will ask you questions and refuse to proceed until you answer. That's the point.
How to install
Via Claude Code plugin: claude plugins install mattpocock-skills. Or via npx skills@latest add mattpocock/skills for Codex or any agent. The skills install into ~/.claude/skills/ or your project's .claude/skills/ directory. Run /setup-matt-pocock-skills once per repo to configure your preferences.
Concerns and limitations
The skill is honest to a fault, which means it will slow you down. If you're on a deadline and just need content, the iterative back-and-forth will feel like overhead. It also requires you to have raw material ready — it's not a brainstorming tool, it's a shaping tool. The "skills-in-progress" tag in the repo path suggests this isn't final, so expect breaking changes. And because disable-model-invocation is set to true, it won't fire automatically — you have to invoke it intentionally, which is a feature, not a bug.
Verdict
This is the most principled AI writing skill I've used. It treats the reader's understanding as a first-class concern and refuses to produce prose that's structurally unsound. It won't write your article for you — it will make you earn every paragraph. For developers who care about how their writing lands, that's exactly what's missing from the AI tooling landscape. Install it. Use it on something you've been avoiding. You'll finish with a better article and a better sense of how your thinking actually works.
Links: - SkillsMP: https://skillsmp.com/creators/mattpocock/skills/skills-in-progress-writing-shape - GitHub: https://github.com/mattpocock/skills/tree/main/skills/in-progress/writing-shape