CRAFT for Claude Cowork โ€” Features

CRAFT NEWS

CRAFT for Claude Cowork · Free Public Beta

Everything CRAFT ships today, on one page — so you don’t have to hunt for it across posts and tutorials.


How to read this page. Every feature below is in the free download right now — nothing here is planned, coming soon, or aspirational. It’s a starting point, and it will grow: the framework changes faster than marketing copy does, so counts stay approximate on purpose and this page gets updated as CRAFT evolves. Where a live tutorial demonstrates a feature, the entry links to it.

If you want the shorter, argued version — what CRAFT adds to Claude Cowork and where Cowork alone is enough — that’s the honest comparison page.

Session continuity & working memory

  • Written handoffs. Every session ends by writing its state — decisions, open items, exactly where the work stopped — and the next session boots from that file. tutorial →
  • Session resume. Start a new session and pick up mid-stride, days or weeks later. tutorial →
  • Mid-session checkpoints. Save a restore point before anything risky, and roll back to it if needed. tutorial →
  • Device switching. Close the project on your desktop, continue on your laptop — files and working state both move. tutorial →
  • Compaction resilience. Because the working state lives in files, hitting a long conversation’s context limits doesn’t cost you the project’s memory.
  • Lessons that compound. A lessons-learned file records mistakes and discoveries once; sessions re-read it at boot, so the same mistake doesn’t get made twice.
  • Organized session naming. A naming standard keeps your Recent Chats menu findable across a long project.

Zero-fear upgrades & self-repair

  • Two-folder architecture. The framework lives in a read-only folder; your work lives in a separate writable one. The two never mix. how it works →
  • Drag-and-drop upgrades with proof. Upgrading is a folder swap — and a built-in verifier proves your work was untouched, with rollback if you want the old version back. how upgrades work →
  • The Two-Folder Doctor. Built-in diagnostics that detect and repair the framework’s known failure modes — a doctor to run before there’s a mystery to solve.

A named team of personas

  • 19 official personas. An orchestrator routes work to named specialists — research, brand architecture, campaigns, content strategy, editing, writing, technical, business review, and more. tutorial →
  • Multi-perspective reviews. Work can be reviewed by several persona lenses, with recorded confidence scores — not one voice agreeing with itself. go deeper →
  • A defensive-only security pair. Two personas who find and harden, weigh and accept — never weaponize.
  • Create your own. Recipes for building and customizing your own personas ship in the box.

Recipes & Official Cookbooks

  • ~90 recipes and growing. Repeatable written procedures — session lifecycle, git safety, QA, research, content, delegation — organized into volumes, every recipe versioned and indexed. The current count lives on the downloads page. tutorial →
  • What you get, spelled out. A plain-language tour of the recipe library and how to use it. read it →
  • Multi-recipe workflows. Chain recipes into larger repeatable procedures. tutorial →
  • Downloadable Official Cookbooks. Curated recipe libraries the AI fetches and ingests in one command — two live today (Branding & Identity, Journalist Process), more coming.
  • Honesty built into the catalog. The framework knows your local install is a curated subset — ask “is there more?” and it routes you to the live catalog instead of guessing.

Build your own

  • Author recipes, cookbooks, and workflows. A protected personal namespace that upgrades never overwrite, with validation to keep your creations well-formed. tutorial →
  • Manage your library. Organize, version, and maintain what you build. tutorial →

Quality gates & drift control

  • A written quality bar. Validator recipes, review gates, and consistency checks that every session can reach for — output quality stops depending on how well you prompted today. how it catches mistakes →
  • Factual-claim validation. Recipes that check claims against live sources before they ship.
  • Cross-session consistency. Checks that today’s work doesn’t quietly contradict last month’s decisions.
  • Full project state audit. A comprehensive health check for the whole project.

Security & provenance

  • No secret ever appears in chat. A written credential discipline: tokens live in a dedicated ignored folder and are passed out-of-band — never typed into a conversation, never in files that leave your machine.
  • Signed, verifiable releases. Every release ships with SHA256 checksums and a GPG signature, from byte-reproducible builds — you can verify your download is exactly what was published. security page →
  • Personalization with a true OFF. Brand-aware mode or zero-personalization mode — and off genuinely means off, with consent gates on every personal surface. tutorial →

Git safety & disaster recovery

  • Checkpoints and backups. Git checkpoints, pre-modification backups, snapshot bundles, and off-site sync — designed for people who aren’t git experts. tutorial →
  • Project archiving. Retire a finished project cleanly, with its full history intact. tutorial →

Learning CRAFT

  • Tutorial mode. Say “turn on tutorial mode” and CRAFT teaches itself to you, step by step, inside your own project. tutorial →
  • A beginner series. Plain-language setup and basics pages on this site, with official install links. start here →
  • A structured human–AI language. A typed comment protocol for giving the AI unambiguous instructions — and reading its status back the same way.
  • Diagnostic mode. A verbose-trace toggle for when you want to see exactly what the framework is doing.

Working beyond one AI

  • Cross-AI delegation. Structured paste-packets hand research out to your other AI subscriptions and fold results back in — with a wall that keeps your project’s private material out of outbound prompts. tutorial →
  • Sub-agent delegation with contracts. Parallel work runs on written contracts — what each agent must deliver, in what form, to which file — so it comes back mergeable.
  • Research, structured. A front door for research tasks with sourcing discipline. tutorial →

The project-management spine

  • A standing to-do list. Shared between you and the AI, surviving across sessions.
  • Subprojects with charters and phase gates. Long arcs of work get structure: a goal, phases, and a close that has to be earned. tutorial →
  • Institutional memory. Lessons-learned capture is part of the lifecycle, not an afterthought — your project gets smarter as it gets older.

The honest fine print

Everything above is plain Markdown and text files you can open, read, and verify — that’s the point of the whole design. This page describes the framework as it is today, kept deliberately conservative: when in doubt, we leave a feature off this page rather than stretch a claim. If your Claude use is occasional and self-contained, Cowork alone may be genuinely enough — here’s our honest accounting of that, and the deeper story on memory is in What Claude Remembers Now — and What CRAFT Adds.

Try it against this list. CRAFT for Claude Cowork is in free public beta through December 31, 2026 — no login, no paywall, no email gate. Download it, open it in Cowork, and check any entry yourself. Get the latest release →

This page reflects the framework as it is today and gets updated as CRAFT evolves — counts and capabilities on this site are kept approximate on purpose, because the framework changes faster than marketing copy does.