CRAFT for Claude Cowork · The Case for Curation & Provenance
What CRAFT cookbooks are — the hierarchy, the anatomy — is documented on the cookbooks page. This post is about the argument underneath them: why we bet on a curated shelf instead of an open pile — what the curation consists of, and who stands behind it.
Most AI ecosystems answer the capability question with a marketplace: thousands of plugins, extensions, or prompts, contributed by anyone, ranked by popularity. The pile is genuinely impressive, and it answers a real question — breadth. Whatever your task, something in there touches it. But popularity tells you what’s common, not what’s ordered toward your job — and assembling a workflow from parts that were never designed to meet makes you the integrator.
CRAFT for Claude Cowork took the opposite bet: fewer things, arranged on purpose. An official cookbook is a set of recipes curated around one job and sequenced so each step hands to the next — not a search result, an itinerary. Look at one from the inside: the Branding & Identity cookbook walks a single guided path from defining who you are to a brand profile, campaign plan, and content calendar that speak with one voice, each step handing its work to the next. The Journalist Process cookbook does the same from idea to published piece. Neither asks you to assemble anything — the assembling was the authors’ job, done once, for everyone.
Curation is a promise about what’s inside
“Official” on a CRAFT cookbook is not a badge on a store listing — it’s the shelf that ships with the framework, built and revised under the same process discipline as everything else in it. And the discipline has teeth: the framework carries its own verification layer — recipes whose whole job is checking work, with fact-check gates and 0–100 confidence scoring — and a published account of that layer catching a real mistake — and when a cookbook falls short, it gets revised. The mistake was small and plausible: draft copy in nine content files calling the framework open source, when the license says otherwise. The pass flagged the claim, and all nine files were corrected before any of it shipped — an error a human reviewer might miss on the tenth read, caught on the first fact-check pass.
A cookbook comes with its own experts
Curation covers who does the work, too. The branding and journalist cookbooks above do not arrive as bare instructions for you to hand a generalist assistant: each ships with its own team of specialist AI personas — subject-matter experts, plural, each with its own part of the job to mind. In the branding cookbook the bench spans who you are, how you will reach people, and what you publish; the journalist cookbook is staffed the same way, from finding what is worth covering to judging whether a piece is ready.
Why staff a job with several experts instead of one capable generalist? Because a bench disagrees. Any single voice, human or AI, reviewing its own work tends to approve it; different lenses on the same work catch different mistakes — the process lens catches what the language lens misses, and the other way around. That is the same reasoning behind the verification layer above: quality comes from more than one perspective with standing to object. The personas are how a cookbook packs that principle into the box — you open it and the bench is already seated.
The other question: who made this?
Before you ask what an add-on does, there is a question worth asking first: who made it, and how would you know? In open marketplaces, the honest answer is usually an account name. Independent security researchers who audit AI-agent marketplaces keep documenting the same pattern — useful add-ons shelved next to malicious ones, published under account names with no identity you can check behind them. That is not a knock on any one platform; it is what an open pile permits, structurally: breadth without provenance.
The receipts: what independent audits found
- One security vendor audited 3,984 publicly shared agent skills: 36.82% carried at least one security flaw, 13.4% with critical issues — Snyk, February 2026
- 76 payloads in that corpus were human-confirmed malicious — and a coordinated campaign pushed 30+ malicious skills into one public marketplace — reported in Snyk’s February 2026 audit
- A live skill was found in the wild stealing developer credentials while disguised as a usage-report upload — Datadog Security Labs, May 2026
- In the public agent-tool registries surveyed, moderation was reactive and community-reported, with no pre-publication code review and no standardized publisher-identity verification — academic surveys, 2025–2026
Point-in-time figures from named third-party research into open AI-agent ecosystems; none of these studies examined CRAFT. The pattern, not any single number, is the argument.
CRAFT’s answer is small enough to check. Every official cookbook comes from one named person — Richard Ketelsen, a builder who brings a safety-first mindset from his day job in cybersecurity and has been developing AI methods in public, week after week, at ketelsen.ai. The recipes are plain instructions you can read, not opaque code you have to trust: the whole shelf ships in the open, and the methodology, cookbook structure, and recipe format are documented in the public repo. That is what building openly buys you — trust you can check instead of trust you are asked to take.
What this feels like on day one
The practical difference shows up in your first hour: you start at a stocked workbench — the what-you-get tutorial shows what ships and how to invoke it in plain language. Pick the cookbook that matches your job, open it, run the first recipe. The choice you make is “which itinerary,” not “which pile of parts.”
And curation is a floor, not a ceiling: the framework is built for you to write your own recipes and grow your own collections on top of it — your additions ride the same structure the official shelf uses. Curated to start, extensible by design.
Start from a stocked workbench
The button below downloads CRAFT for Claude Cowork itself — the framework this post argued for, official cookbooks and their personas included. The fastest way to test the argument is to open one cookbook, see how it’s arranged, and check who made it.
No account needed. Free during beta. The download is ungated.
The honest bounds. CRAFT for Claude Cowork is in beta, and we say so plainly. One honest prerequisite: CRAFT runs on Claude Cowork, so you need Cowork available in your Claude plan — CRAFT adds no cost, account, or gate on top of it. Curation is a promise about care, not perfection: cookbooks get revised when they fall short, AI can make mistakes, and the verification layer exists because of that — review anything consequential yourself. This post reflects the framework as it is today and gets updated as CRAFT evolves — current contents live on the pages linked above.
