100+ clients served
$10M+ client revenue generated
3,200 calls booked
Free Claude skill

Build an AI teammate that lives in your Slack.

Six questions about your business, five stages, and an agent is live in a real channel answering real messages. Runs on your own keys, on infrastructure you own. No subscription.

What you actually end up with

A named agent sitting in a Slack channel that you talk to in plain English. It knows what your business does and who you sell to, and it does one real job you currently do by hand. About 150 lines of code you can read, running on a free tier.

Not a wrapper around someone else's product. You hold the keys, the code and the data.

What you need before you start

No n8n, no Zapier, no monthly tool.

The five stages

1
ScopeSix questions become the agent's brief. The one that matters is "what is the first job" - and it has to be small and real.
2
The Slack appA generated manifest you paste in. You come away holding a bot token.
3
The runtimeOne file, deployed. It bridges Slack to Claude and every non-obvious line says which failure it prevents.
4
Wire it up and prove itA real message in a real channel gets a real reply. Nothing else counts as done.
5
The second jobAdd capability without breaking what already works, and behind an approval step.

The part nobody warns you about

Your first Slack agent will install perfectly, report no errors, and say absolutely nothing.

That is the normal outcome, not bad luck. Almost every layer in this stack returns a success status for a total failure. A 200 from your endpoint is not a reply. A green deploy is not a running version. An installed app is not an event subscription. Each of those will let you believe you are finished while nothing works.

One example, because it costs people a whole evening. The obvious way to stop your bot replying to itself is to ignore any message carrying an app_id. Do not do that. Human messages arriving through relays and integrations also carry an app_id, so that guard silently discards real people. The agent looks completely dead and every log line says success.

The skill ships a catalogue of these, starting with the one question that rules out most causes in a single check. That file is the reason this exists as a skill rather than a blog post.

Get the skill

Download the skill

Prefer a zip? Download the .zip - same contents.

Install it in 60 seconds

  • Unzip into your skills folder. In Claude Code that is .claude/skills/slack-agent-starter/.
  • Start a session and say "let's build my Slack agent".
  • Answer the six questions. It runs the rest with you.

What's in the box

  • The runtime - one commented file that receives Slack events, verifies they are genuinely from Slack, asks Claude for a reply, and posts it back.
  • The failure catalogue - how these break in practice, ordered by how often each one bites, starting with the fork that halves the search space.
  • The law - the guardrails to give an agent before you let it write, send or spend anything. The headline: buttons act, the model never does.
  • A preflight check - proves your endpoint answers Slack's verification challenge before you wire anything, because a failed Request URL gives you no useful error.

A note on safety

The skill refuses to build fully automated messaging on platforms whose terms prohibit it, LinkedIn most notably. Draft plus one-tap approval gets you nearly all of the speed with none of the account risk. Anything customer-facing is a draft for a human to send.