SuperAPI AI Directory

paste

Office & productivity

View repository

Updated 10-15 16:37

About this skill

/paste — Clean content for any destination

Cleans terminal output (ANSI, box-drawing, prompt artifacts, etc.) and reformats per destination rules. Reads clipboard by default, writes back to clipboard + previews in chat.

Step 1 — Get the content

In order:

  1. If the user included content in the prompt (pasted, or referenced from earlier in the conversation), use that.
  2. Else, read the clipboard (commands below).
  3. If both empty, ask the user what to paste.

Clipboard commands by platform (read / write):

Platform Read Write
macOS pbpaste pbcopy < "$f"
Linux (Wayland) wl-paste -n wl-copy < "$f"
Linux (X11) xclip -selection clipboard -o xclip -selection clipboard < "$f"
Windows powershell Get-Clipboard -Raw powershell "Get-Content -Raw '$f' | Set-Clipboard"

$f is the temp file from the write procedure in Step 6. Always copy from a file, never from a heredoc or echo.

No clipboard (a remote, headless, or chat-bot agent): skip the clipboard, take content from the prompt, and return the cleaned text in a code fence.

Step 2 — Parse destination

Invocation Destination
/paste plain (default)
/paste plain plain
/paste slack slack
/paste notion notion
/paste twitter or /paste x twitter
/paste linkedin or /paste li linkedin
/paste email email (plain)
/paste email rich or /paste email html email (rich/HTML)
/paste github or /paste gh github
/paste markdown or /paste md markdown render
/paste html render to HTML file + open in browser

If ambiguous, ask. Per-destination rules live in references/destinations.md.

Step 3 — Scan for secrets

Before any cleaning or copying, scan against patterns in references/secret-patterns.md.

If any match:

  1. Stop. Don't copy yet.
  2. Report what was detected with the value masked (first 4 + last 4 chars only — e.g. sk-12...wxyz).
  3. Ask: "Spotted [type] in the content. Continue, redact (replace with [REDACTED]), or abort?"
  4. Default to abort if no clear answer.

Step 4 — Universal cleaning

Strip from the content:

  • ANSI escape codes — \x1b\[[0-9;]*[a-zA-Z] and related (color codes, cursor movement, screen control)
  • Box-drawing chars — ╭ ─ ╮ │ ╰ ╯ ╞ ╡ ╤ ╧ ┌ ┐ └ ┘ ├ ┤ ┬ ┴ ┼ ━ ┃ ┏ ┓ ┗ ┛ etc. Replace with simpler equivalents (-, |) or remove entirely depending on context.
  • Terminal prompt artifacts — leading $ , > , ❯ , % , # (when first non-whitespace on a line that looks like a shell prompt)
  • Carriage returns (\r) — replace with \n or delete
  • Trailing whitespace per line
  • Excess blank lines — collapse 3+ consecutive blank lines to 2

Step 5 — Apply destination rules

Read references/destinations.md and apply the relevant transform.

Step 6 — Output

Clipboard write procedure (every destination that writes to the clipboard):

  1. Write the cleaned content to a temp file with your agent's file-write tool, e.g. /tmp/paste-<slug>.txt. Never pipe it through a shell heredoc or echo: agent harnesses can hard-wrap heredocs (~70 chars) and indent every line, so the paste comes out broken while the chat looks fine.
  2. Copy from the file with the write command from Step 1.
  3. Verify the round trip. Read the clipboard back and compare it to the file, e.g. pbpaste | cmp -s - "$f" && echo match (swap in the platform's read command; on Windows, powershell "(Get-Clipboard -Raw) -eq (Get-Content -Raw '$f')"). On a mismatch, report it and fall back to showing the text in a code fence.
  4. Build the chat preview from what you read back, not from what you meant to copy. A preview from memory is how a mangled copy ships looking clean.
  5. Delete the temp file once verified. It can hold anything the user copied, including text the secret scan let through.
Destination Output behavior
plain / slack / notion / twitter / linkedin / email / github (a) Show preview in a code fence in chat. (b) Clipboard write procedure above; preview comes from the read-back.
email rich Render to HTML, write to /tmp/paste-<timestamp>.html, open it in the browser (open on macOS, xdg-open on Linux, start on Windows). Skip clipboard (the user copies from browser to preserve rich text).
markdown Render the cleaned markdown directly in chat (most agent UIs render it). Also copy raw markdown to the clipboard via the write procedure. Offer: "Open as HTML too?" — if yes, write and open.
html Render to HTML, write to /tmp/paste-<timestamp>.html, open it in the browser. Skip clipboard.

After output, report a one-line summary of what was cleaned (e.g., "Stripped 12 ANSI codes, 4 box-drawing chars, 1 prompt artifact. Character count: 248 / 280 (X). Clipboard round-trip verified.").

Length warnings

Twitter and LinkedIn have practical limits. the user said "let me trim" — don't refuse or auto-truncate.

  • Twitter/X: warn if >280 chars. Note the overage.
  • LinkedIn: warn if >400 chars (comfort line, not hard limit).
  • Other destinations: no length warning.

URL handling for social destinations

If the content has URLs and destination is twitter or linkedin:

  • Surface the URLs separately (don't include them in the body)
  • Remind the user: links go in a first comment (LinkedIn) or reply (X), not the body
  • See makerskills:jab-hook and feedback_social_link_placement.md in your memory store (your agent's memory, or ${MAKERSKILLS_MEMORY:-${MAKERSKILLS_CONFIG:-$HOME/.config/makerskills}/memory}/)

Examples

/paste slack

→ Reads clipboard, strips ANSI, converts **bold** to *bold*, copies to clipboard, shows preview.

/paste twitter
[pasted error log]

→ Uses pasted content (ignores clipboard), strips everything, warns at 412/280 chars, copies, asks if the user wants to trim or pick a different destination.

/paste html
[pasted markdown table]

→ Renders table as HTML, opens in browser. the user selects + copies into Notion/email with formatting preserved.

/paste
[pasted output with sk-anthropic-key in it]

→ Stops. "Spotted what looks like an Anthropic API key: sk-an...3kf9. Continue, redact, or abort?"

Composes with

  • jab-hook — clean output for pasting into Typefully drafts when the MCP path doesn't fit; also enforces the link-placement rule for X/LinkedIn destinations
  • social-fetch — clean a fetched post before quoting it in a draft or newsletter
  • second-brain — paste output can be captured cleanly into raw/note-<slug>.md for future compilation
  • watch-video — transcript / summary output often flows through paste for platform-specific reformatting

Notes on quality

  • The clipboard write procedure is not optional. File → copy from file → read back and compare → preview from the read-back. Origin: 2026-07-31, a heredoc-piped copy shipped hard-wrapped and indented while the chat preview looked clean; the user caught it, not the verification.
  • Secret detection runs first, always. Before any formatting or destination logic, paste scans for sk-*, AKIA*, xoxb-*, JWT-shaped strings, and long-hex tokens. Prompts before proceeding on any hit. Better to false-positive occasionally than to leak a real key.
  • Destination-aware transforms, not one-size-fits-all. Slack wants *bold*; LinkedIn wants unicode-styled bold; email wants HTML. Same input, 9 different valid outputs. The destination flag is not optional.
  • ANSI codes get stripped for every destination. Terminal escapes (\033[31m etc.) render as garbage everywhere except the source terminal.
  • Character limits are enforced, not warnings. X at 280, LinkedIn at 3000, Twitter at 25000 — paste refuses to copy over-limit output and offers a trim strategy.
  • Links go in first comments for social destinations — not the body. Enforced when destination is twitter or linkedin. Documented in feedback_social_link_placement.md in your memory store.
  • HTML destination opens in a browser tab, not the clipboard. Formatted tables + code blocks need a rendered surface to select-and-copy from with formatting preserved. Copying raw HTML to the clipboard produces garbage in Notion/email.