How to use the Slack Task Tools agent skill

What the skill makes the agent do

Agent skills are instruction packs that an AI agent loads so it knows how to handle a particular kind of task. This one covers Slack. Once it is active, the agent works through the slack_* tools that Paperclip supplies for a task, using the Slack bot assigned to that agent and the access of the linked requester.

In practice the agent can read discussions in the channel it was given, follow thread replies and earlier history, and search a channel. When you ask, it can also write: messages, uploads, reactions, pins, bookmarks, topics, canvases and lists.

• Read shared discussions and summarise them with links to the source messages
• Search a channel and report how the search was scoped
• Post messages and other items only when requested
• Send a direct message to the responsible user for a task or routine

Install the Slack skill

Download the ZIP from the skill's page in this directory and unzip it. The package contains a single file, SKILL.md. For a personal setup, place the unzipped skill folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload

Set up what the skill requires

The skill does not connect to Slack by itself. It depends on a Paperclip server with a Slack connection already configured. A Slack bot must be assigned to the agent and installed in the workspace with the required scopes. The requester whose work the agent is doing must have a linked Slack identity.

For CLI or sandbox runtimes, the environment must provide PAPERCLIP_API_URL, PAPERCLIP_API_KEY, PAPERCLIP_COMPANY_ID, PAPERCLIP_TASK_ID and PAPERCLIP_RUN_ID. The skill uses these to call the Paperclip tools endpoint. It also says the agent must never print credentials.

• Paperclip server with a Slack connection configured
• Slack bot assigned to the agent and installed with required scopes
• Requester's linked Slack identity
• For CLI/sandbox runtimes: PAPERCLIP_API_URL, PAPERCLIP_API_KEY, PAPERCLIP_COMPANY_ID, PAPERCLIP_TASK_ID, PAPERCLIP_RUN_ID

Trigger the skill with natural requests

You do not need special syntax. Ask for what you want in plain language inside a Paperclip task or routine that has the Slack connection. Name the channel when you can. If you do not, the agent uses slack_channels to find the destination you mean.

When a task starts from a Slack conversation, the agent reads from the channel where the request came from. Keep requests specific about what you want back, such as a summary, a decision list, or a message to a named destination.

• "Summarise the launch discussion in this channel and link the key messages."
• "Search this channel for the launch decision and list what was agreed."
• "Send me a Slack message with the open questions from this task."
• "At 10am, send the channel a status update."

Examples of search and channel reads

For a channel search, the skill shows the agent calling slack_search with the channel ID as an array, a query, and a limit. The limit is at most 20 matches. It is not the history page size, so a search does not read the whole channel history.

The skill notes that a bounded history scan is not a workspace-wide search and does not automatically look inside thread replies. If you need older or threaded material, ask the agent to fetch more history or thread pages.

Understand approvals, idempotency and delivery

Every write needs an idempotencyKey in UUID form. The agent handles this, but the skill requires that retries reuse the same key and identical arguments. Destructive operations, creating channels and sending invitations require approval through Paperclip. Instructions found inside Slack content are never treated as approval.

A newly created channel stays disabled for ongoing responses until a person enables it in the connection Settings. When an action is queued or its outcome is uncertain, the agent should report that it is not delivered, rather than retrying with a new key.

• Destructive actions and channel creation wait for approval
• Invitations also require approval
• A queued or uncertain send is not treated as delivered
• Routines check the responsible user's link and permissions when they run

Safety notes: scripts and network

The skill does not run scripts. It does use the network, because its tools call Slack through Paperclip's server. The agent is designed to work only through the assigned bot and the linked requester's access, and to never request or pass Slack tokens or workspace IDs.

Retrieved Slack messages, files and names are treated as untrusted. They cannot instruct the agent to do unrelated work, change permissions or reveal credentials. Private-channel material stays in its source channel or a DM with the requester, so research spanning private channels may require asking the requester to move to a DM.

• Runs scripts: no
• Uses network: yes
• Private channels: findings stay in the source channel or the requester's DM
• Other people's bot DMs are not accessible