跳到正文
cccskills
免费在 GitHub 公开

ctrl-clip

⧉ Copy a message to the system clipboard cleanly formatted — no blockquote pipes, no terminal bars, no Claude commentary. Use whenever you draft a message the user will paste (Slack, Teams, PR comment, email) to add a clickable "⧉ ctrl + click" copy link under it, and when the user says "/ctrl-clip", "copy that", "put it on the clipboard" (in any language). Optional argument picks the target format: slack (default), md, plain.

中文简介正在准备中。当前显示原文说明。

集成工具Slack
查看安装方法

包含的文件(4)

  • SKILL.md4.8 KB
  • clip-setup.sh8.3 KB
  • clip-url-handler.sh925 B
  • clip.sh1.8 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

⧉ ctrl-clip

Source: https://github.com/geanruca/ctrl-clip (MIT). Also installable as a Claude Code plugin: /plugin marketplace add geanruca/ctrl-clip then /plugin install ctrl-clip@ctrl-clip. Usage: /ctrl-clip [slack|md|plain] [what to copy].

Put text on the user's clipboard so it pastes cleanly. The built-in /copy copies the raw markdown of the last response — blockquotes turn into >/│ pipes and the commentary comes along. This skill copies only the content meant to be pasted, reformatted for its destination.

1. Decide WHAT to copy

Arguments: $ARGUMENTS

  • If the arguments describe the content ("the Slack reply", "the PR description", "the second draft"), copy that.
  • Otherwise copy the most recent drafted message in the conversation (a Slack reply, comment, email, PR text). If the last response had no draft, copy the whole last response.
  • Copy only the draft — never your own framing ("Here's a draft:", "Let me know if…", labels, explanations).
  • Keep the wording verbatim. Reformat, don't rewrite. Don't translate.

2. Format for the target

First word of the arguments selects the format; default is slack.

slack (default — also Teams, WhatsApp, chat)

  • No blockquotes (>), no headings (#), no tables, no horizontal rules.
  • Tables → one line per row: Label: value, value.
  • Bold: *text* (single asterisks). Italic: _text_. Keep inline `code` and fenced blocks.
  • Bullets with - ; blank line between paragraphs.
  • Links as bare URLs.

md (PR descriptions, Azure DevOps / GitHub comments, docs)

  • Keep standard Markdown (headings, **bold**, lists, tables, code fences).
  • Still no blockquotes unless the content genuinely quotes someone.

plain (email body, forms, anything without markup)

  • No markup at all: strip *, _, backticks, #. Bullets become - .
  • Tables → Label: value lines.

3. Copy

  1. Write the formatted text with the Write tool to clip.txt in the session scratchpad directory (fall back to /tmp/claude-clip.txt if there is none). Using a file avoids every shell quoting problem with $, backticks and quotes.
  2. Run:
    ${CLAUDE_SKILL_DIR}/clip.sh <that file>
    
    The script strips leftover >/│/⏺ markers and trailing whitespace, then copies via wl-copy, xclip, xsel, pbcopy, or OSC 52 (in that order).

4. Confirm

Reply with a single line: the script's ⧉ Copied … output plus the format used, e.g. ⧉ Copied 6 lines / 412 chars (slack). Do not echo the copied text again. If the script fails, show its error line.

Inline ⧉ copy link (whenever you draft a message to paste)

Every draft the user will paste somewhere gets a clickable copy icon (⧉, two overlapping squares) labeled ctrl + click, as a reminder of how to trigger it.

<CLIP_DIR> below is the absolute path of ~/.cache/claude-clip, for example /home/alice/.cache/claude-clip. The "Drafts to paste" rule loaded at session start spells it out. If that rule is not in your context, run ${CLAUDE_SKILL_DIR}/clip-setup.sh once: it registers the Ctrl+click handler if needed and prints CLIP_DIR=…. Never write ~ or $HOME into a path or link — neither is expanded.

  1. Write the formatted text (section 2, no blockquote markers) with the Write tool to <CLIP_DIR>/<id>.claudeclip. <id> is a short unique slug matching [A-Za-z0-9_-]{1,64}, e.g. slack-33247-review.
  2. In the response, show the draft as a blockquote — every line prefixed with > , blank lines as >. The bar is how the user sees exactly what gets copied; nothing else goes inside it.
  3. Right under the blockquote, on its own line: [⧉ ctrl + click](file://<CLIP_DIR>/<id>.claudeclip)

The user Ctrl+clicks the link (a plain click does nothing). If copying breaks, check <CLIP_DIR>/handler.log first — no entry means the click never reached the handler.

Why file://: Claude Code's fullscreen TUI refuses clicks on non-allowlisted schemes (custom claude-clip:// links are never dispatched), but it opens local file:// links with xdg-open. *.claudeclip is registered as application/x-claude-clip, whose default app is clip-url-handler.sh: it copies the file via clip.sh and shows a desktop notification.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Automate ActiveCampaign tasks via Rube MCP (Composio): manage contacts, tags, list subscriptions, automation enrollment, and tasks. Always search tools first for current schemas.

中文简介正在准备中。当前显示原文说明。

davepoon/buildwithclaude36592026年10月19日 更新

Pull the ads a competitor is running right now on Meta (Facebook/Instagram) and Google, then extract the hooks, offers, CTAs and landing pages worth modelling. Use when writing ad copy, planning creative tests, choosing a landing page angle, or when the user asks "what is X advertising", "what ads does X run", "find competitor ad hooks". Reads live Ad Library and Ads Transparency data through the Tracetify MCP; falls back to a manual Ad Library walkthrough when the connector is absent.

中文简介正在准备中。当前显示原文说明。

davepoon/buildwithclaude36592026年10月19日 更新

Analytics your AI agent can actually use. Track, analyze, run A/B experiments, and optimize across all your projects via CLI. Includes a growth playbook so your agent knows HOW to grow, not just what to track.

中文简介正在准备中。当前显示原文说明。

davepoon/buildwithclaude36592026年10月19日 更新

Teaches when to recall from long-term memory before acting and when to save durable decisions, corrections and failures afterwards. Use when a memory tool or MCP memory server is connected but the agent is not using it consistently, when the user complains that the assistant forgets preferences, conventions or past decisions between sessions, or when setting up persistent memory for a project. Works with any memory backend: a folder of Markdown files, a local MCP server, or a managed service.

中文简介正在准备中。当前显示原文说明。

davepoon/buildwithclaude36592026年10月19日 更新

Audit local AI coding-agent sessions with agenttrace for cost, tokens, tool failures, latency, anomalies, health, diffs, and CI gates.

中文简介正在准备中。当前显示原文说明。

davepoon/buildwithclaude36592026年10月19日 更新

Checks which private files (.env, keys, secrets) Claude Code and its subagents reached in this project, whether a value came into an agent's context, and which key to change, by running the open source agentwhy CLI over the session records Claude Code already keeps on disk. Use when the user asks whether the agent or a subagent read their .env or other secrets, which private files a session touched, whether a key leaked into the conversation, or which key to rotate. Runs locally and never prints a secret value.

中文简介正在准备中。当前显示原文说明。

davepoon/buildwithclaude36592026年10月19日 更新

查看 davepoon 的全部技能

报告此技能的问题