slack

name	slack
description	Read and post to Slack from OMP (or any harness without a Slack connector) by shelling out to the `claude` CLI, which has the managed Slack MCP authed and connected. ALWAYS trigger this skill whenever a Slack URL appears anywhere in the chat -- any `slack.com/archives/...` link (e.g. `shedul.slack.com/archives/C.../p...`), a `slack://` link, or a message/thread/channel permalink -- even if the user only pastes the link with no explicit instruction. Also trigger when the user wants to read a Slack thread/channel, post or reply to a message, add a reaction, or search Slack, and no native Slack tool is present in the current session.
allowed-tools	Bash

slack

OMP has no Slack connector. Claude Code does -- its claude.ai Slack managed MCP is authed and ✔ Connected. So borrow it: shell out to the claude CLI in headless mode and let it drive the Slack tools. Same trick works for any managed connector Code has (Gmail, Notion, Calendar, Drive, Atlassian, Figma...).

Model split: brain reads & writes, arm carries

This skill runs in the current foreground model (Opus, or whatever the session is on) -- that's deliberate, hence no model: in the frontmatter. Keep the split clean:

• The FG model (brain) does all comprehension and composition. Understanding the thread, deciding whether/what to reply, drafting the exact message text -- that's judgement, it stays here. Never delegate the thinking.
• claude -p --model haiku (arm) is dumb transport only. It just carries a fully-formed instruction to the Slack MCP and pipes the result back: "read thread X verbatim", "post this exact text to thread Y". It never decides content.

So the brain reads the raw thread (via a haiku shell-out that returns it verbatim -- no summarizing), comprehends it here in full context, writes the reply word-for-word, then hands haiku a "post exactly this" command. Haiku is the wire, not the writer.

When a Slack URL shows up: fetch context FIRST, unprompted

A pasted Slack link is a request for context, even with no instruction attached. The moment one appears in the chat, before replying to the user or doing anything else:

1. Parse CHANNEL_ID and the ts from the URL (see "Parsing a Slack permalink" below).
2. Read it immediately -- slack_read_thread for a p.../thread_ts link, slack_read_channel for a bare channel link. Pull the full thread verbatim, not a snippet.
3. If the thread references another thread, channel, user, or file, follow it (slack_read_thread / slack_read_user_profile / slack_read_file) until you actually understand what's being discussed.
4. Only then respond to the user, grounded in what the thread actually says.

Do not ask "want me to read it?" -- just read it and report what's there. Reading is free and non-destructive; posting is the only thing that needs care. Never post, react, or reply back into Slack off a pasted link unless the user explicitly asks -- default to read-only.

Do it quietly -- no narration

This is plumbing. Say nothing while you work. No "Slack link dropped, gotta pull that thread", no "plan's trivial here", no "fetching now", no announcing the shell-out. Just silently run the read and go straight to the answer.

• Zero preamble before the tool call. The tool call itself is the only signal the user needs that something's happening.
• No meta-commentary about what the skill is, what you're about to do, or how simple/trivial the task is.
• After fetching: lead with the answer to what the user actually asked (e.g. "do we need lots?" -> answer it), grounded in the thread. Not "here's what the thread says" -- just the answer, with thread detail only as needed to support it.
• The FG model still does the thinking; it just does it silently. Think, fetch, answer. No live diary.

The one-liner

claude -p --model haiku --dangerously-skip-permissions "<prompt telling it which slack_* tool to use and with what args>"

• -p -- headless: prints the result and exits, so it behaves as a plain CLI tool.
• --model haiku -- cheap; Haiku drives the MCP calls fine. Bump to sonnet only if it flails.
• --dangerously-skip-permissions -- no interactive tool-approval prompts (required for non-interactive use).
• The flag is --model, not -m (-m errors with unknown option).

Confirm the connector is live first if unsure: claude mcp list -> look for claude.ai Slack ... ✔ Connected.

Slack tool names (namespaced mcp__claude_ai_Slack__*)

You don't pass these directly -- you name them in the prompt so the sub-agent picks the right one. Verified available:

• slack_read_thread -- read a thread (channel id + thread_ts)
• slack_read_channel -- read recent channel messages
• slack_send_message -- post a message; pass thread_ts to reply in-thread
• slack_add_reaction / slack_get_reactions
• slack_search_public / slack_search_public_and_private / slack_search_channels / slack_search_users
• slack_read_user_profile / slack_list_channel_members
• slack_schedule_message, slack_send_message_draft
• slack_create_canvas / slack_read_canvas / slack_update_canvas / slack_read_file / slack_create_conversation

Parsing a Slack permalink

https://<workspace>.slack.com/archives/<CHANNEL_ID>/p<TS>

• CHANNEL_ID = the path segment after /archives/ (e.g. C08963VPK1A).
• Message ts = the p number with a . inserted before the last 6 digits: p1784906364468409 -> 1784906364.468409.
• A reply's thread_ts is the parent message's ts (in the URL as ?thread_ts=...).

Recipes

Read a thread verbatim:

claude -p --model haiku --dangerously-skip-permissions \
  "Use slack_read_thread to read channel C08963VPK1A thread_ts 1784906364.468409. Print every message verbatim with author and text. Do not summarize."

Reply in-thread (single-quote the outer arg; embedded literal text goes in double quotes):

claude -p --model haiku --dangerously-skip-permissions \
  'Use slack_send_message to post a threaded reply in channel C08963VPK1A to thread_ts 1784906364.468409. Send exactly this text: "your message here". Then report the resulting message ts and permalink.'

Search:

claude -p --model haiku --dangerously-skip-permissions \
  "Use slack_search_public to find messages matching 'deploy freeze' in the last week. List channel, author, ts, text."

Gotchas

• Posts as the authed user, not a bot -- messages appear from whoever owns the Code CLI's Slack auth, tagged Sent using @Claude. Don't post anything you wouldn't say under that name.
• Always read before you write. Never post a reply until you've read the thread and know what you're responding to.
• Verify the write landed: re-run slack_read_thread and confirm your message is present before declaring done.
• Quoting: wrap the prompt in single quotes and put the literal message body in double quotes so shell expansion doesn't mangle it. Escape apostrophes in the body as '"'"'.
• Cold MCP start makes the first call take ~10-15s. Give it a generous timeout.