菜单
Turn the current chat into a gi canvas, or open an existing one. Use when the user says "gi canvas", "start a canvas", "переведи в канвас", "продолжим в канвасе", "open the canvas", or asks to move this conversation into a canvas document.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Verify a cozy build by running `cozy verify` against a target you name — a locally-built docker container (default), a running sbx sandbox, or a host checkout. The checks live in cozy-module/verify.nu. Use when you say "verify the build", "build-check", "smoke test the sandbox", "is everything wired up", "/verify-cozy docker", after building the image, or after creating a sandbox.
This skill should be used when working alongside a user in a shared Nushell terminal (e.g. the cozy sandbox) — coaching Nushell in everyday tasks and practicing literate programming with numd (executable markdown), dotnu (.nu scripts that embed their own output as `# =>` comments), REPL capture, and claude-nu session archiving. It governs interaction style: build a common space where the user runs Nushell themselves and grows fluency, instead of the agent executing everything for them. Relevant when the user says "numd," "dotnu," "teach me," "show me how," "let me run it," "literate programming," "run the code blocks in this markdown," "update the embeds," "capture this session," "keep this doc in sync," "export this claude session," "self-updating README," or asks how to document, share, practice, or verify Nushell work.
Reduce a canvas document to its current-state meaning, letting spent rejected paths and process scaffolding fall into git history (recoverable via `git log -p -- <doc>`). Use when the user says "distill", "git-intent-distill", or "reduce the canvas".
This skill should be used when inspecting, querying, or rewriting the user's Nushell command history. Relevant when the user asks "what did I run", "show my recent commands", "what was I trying", "find commands I ran in <dir>", "remove these history entries", "move/retag these history entries to <path>", "fix the cwd on these rows", or when you need to understand what the user has been doing in a directory by looking at their shell history.
Cut a new release for the cozy project: finalize CHANGELOG.md, create a git tag, and summarize what to do next. Use when the user says "make a release", "cut a release", "release", "bump version", "tag a release", or "prepare release".
Re-vendor modules from upstream into cozy/vendor/ and prepare atomic commits grouped by repo. Use when the user says "vendor updated versions", "I vendored, commit", "nu toolkit/vendor.nu then commit", "refresh vendor", "bump vendored modules", or "vendor the latest, commit by groups".
| name | gi-canvas |
| description | Turn the current chat into a gi canvas, or open an existing one. Use when the user says "gi canvas", "start a canvas", "переведи в канвас", "продолжим в канвасе", "open the canvas", or asks to move this conversation into a canvas document. |
| argument-hint | ["canvas path"] |
| allowed-tools | Bash(nu -c *), Bash(ls *), Read |
The user wants a canvas session but is inside a chat session, where typing nushell into Bash is awkward. Run the module command for them and hand back one line to paste into a terminal.
A canvas session cannot be started from within a session: the Canvas style, the Stop hook, and $env.GI_CANVAS all arrive with claude at launch. So this skill always ends the same way — the user runs one command themselves, in their terminal, and that new session is the canvas one.
claude-nu is a nushell module command, not a program on PATH: it is loaded in the user's interactive REPL, and a one-shot nu -c does not load it. So every command you run has to import it first:
nu -c 'use ~/repos/claude-nu/claude-nu/; claude-nu gi'
If that path does not exist on this machine, find the module (ls ~/repos/claude-nu ~/git/claude-nu) and use the real one.
The commands the user runs go into their REPL, where the module is already loaded — hand those over bare: claude-nu gi open <doc>.
Read $ARGUMENTS and the conversation:
gi/ → open it.A canvas keeps one session for life, and gi open handles both halves of that: a canvas with no session: in its frontmatter gets one minted and written in, one that has it is resumed. You do not have to tell them apart.
gi on its own tells you what is seeded here and whether this session is already bound (canvas non-null → it already is a canvas session; say so and stop). ls gi/*.md lists the repo's canvases.
nu -c 'use ~/repos/claude-nu/claude-nu/; claude-nu gi enable --from-session' # → gi/session-<id>.md, seeds style + skills
nu -c 'use ~/repos/claude-nu/claude-nu/; claude-nu gi enable --from-session --tools' # ...keeping tool calls as one-line placeholders
nu -c 'use ~/repos/claude-nu/claude-nu/; claude-nu gi enable notes/x.md --from-session' # ...at a chosen path
It refuses to overwrite an existing doc — that is deliberate, do not delete the old one to get past it; name another path instead (gi enable <doc> --from-session).
Then tell the user, in one line: exit this session and run claude-nu gi open <doc>. The import wrote this session's id into the canvas, so that reopens this same session (same id, full context) with the canvas bound, and the missing tail — the turns after the import, which the session log cannot contain yet — is still in context and can be appended there.
gi open launches Claude Code itself, so the user runs it — not you. Hand over one of these, as is:
claude-nu gi open # new timestamped canvas, then launch
claude-nu gi open gi/plan.md # a named one; created if new, resumed if it holds a session
claude-nu gi open <doc> --no-hook # style only, no Stop-hook floor
One line with the exact command to run. No explanation of the protocol, no summary of what the canvas is; the style file does that once the session starts.