https://blog.replit.com/building-my-first-slack-bot

 Features Careers Blog Pricing
[?] Sign up
Back to blog

I'm not a programmer, and I used AI to build my first bot

Thu Oct 5 2023 by Jeff Burke

[building-m]

Before I joined Replit, I had never written a line of code.

A few weeks ago, I built a Slack Bot that uses Google LLMs to provide
daily channel summaries. You can fork the template or get started
with 100 Days of Code.

None of this was possible without Replit. Let me explain why:

 1. Getting started with no setup
 2. Using Replit AI capabilities
 3. Learning LLMs with ModelFarm
 4. Simple hosting and deployments

Getting started with no setup

I studied Finance, Marketing, and Math. I never took a single
Computer Science class. During my time playing professional baseball,
I tried to learn to code to earn some money to support my training.
At Boston Consulting Group, I attempted to learn Python to automate
some basic data manipulation and scraping tasks we did.

Every time, I would get stuck early on. It would take far too long to
see any output. The process was too slow, high-friction, and boring.
Half of my time felt like it was just spent getting set up.

Fast forward to last year, and we released 100 Days of Code on Replit
. It was the best experience I had ever had:

  * No wasted time to set up my development environment
  * Daily, digestible lessons
  * Video and written content directly in the editor, so there was no
    tab and context switching
  * Ghostwriter helped with the boilerplate code to make me super
    efficient in later lessons

Within days and weeks, I was building legit projects, and I was using
these skills to improve my day-to-day work at Replit.

Once you learn how easy getting started can be, you won't go anywhere
else. Replit is the only place I will ever code.

Using Replit AI capabilities

Replit's AI helped me learn in 100 Days of Code faster, but it is
also super helpful with creating new projects and understanding other
projects. For example, for my Slack bot, I began with a simple
prompt:

Generate code that allows me to call on the Slack API and log the
messages from a specific Slack channel

I knew nothing about the Slack API. I had not visited the docs yet. I
wanted to start with Ghostwriter.

import os
from slack_sdk import WebClient
from slack_sdk.errors import SlackApiError

# use your own oauth token that you receive from Slack, remember not to reveal it to others
slack_token = os.environ["SLACK_API_TOKEN"]
client = WebClient(token=slack_token)

channel_id = 'your_channel_id' # insert your channel id here

try:
    result = client.conversations_history(channel=channel_id)
    messages = result.data['messages']
    for msg in messages:
        print(msg)
except SlackApiError as e:
    assert e.response["ok"] is False
    assert e.response["error"]
    print(f"Got an error: {e.response['error']}")

This was a great starting point, and instead of copy and pasting from
another tab, I was able to just one click insert into the development
environment.

One of the tricky components of Slack is managing the bot
permissions. You can read the docs, but in this case, I just asked:

[ghostwrite]

Replit's AI was able to give a pretty clear answer, so I simply added
those permissions.

I was able to use Edit Code to make small adjustments like change the
message logging to only messages for that day. In cases where I
generated new code, I was also able to ask the AI to explain specific
lines, which helped me learn.

For someone who was building this project alone, Replit's AI was
basically able to not only speed up the process but also act like a
tutor.

Learning LLMs with ModelFarm

I wanted this Slack bot to provide utility. For me, I am a part of so
many Slack channels at Replit, and it is a struggle to track them
all. Summarizing the channels would provide a lot of utility. To do
that, I would need to use LLMs.

Replit recently launched the ModelFarm, and for a month, the cost of
the Google Vertex models is free. So I decided to use these for
summaries.

# Only log today's messages
          if message_timestamp.date() == date.today():
            user_id = message.get('user', 'N/A')
            user_info = client.users_info(user=user_id)
            user_name = user_info["user"]["name"]
            timestamp_str = message_timestamp.strftime('%Y-%m-%d %H:%M:%S')
            message_content = f"{user_name} said {message['text']}"
            channel_content += f"{message_content}\n"

This block of code:

  * Logs every message from that day
  * Reformats to say "{username} said {message}" (ex., "Jeff said I
    am OOO")
  * Combines into one massive message labeled channel_content

Then we had one paragraph of all the messages. Using Model Farm, I
created this:

  response: TextGenerationResponse = model.predict(
      f"Read all of the text in {channel_content}. Please make sure to include which user said each discussion point. Then provide a summary in bullets on all the main discussion points",
      temperature=0.8,
      top_p=0.8,
      top_k=40,
      max_output_tokens=256)
  response_text = response.text
  print(response_text)

This is a fascinating part of LLMs. With all of my messages stored in
channel_content, I now want to send them to the model. But within
this prompt, I wanted to give the Vertex model clear instructions, so
I just embedded a prompt that basically says:

  * Read all of the messages in channel_content
  * Include the username
  * Provide a summary of main points
  * And format with bullets

So the input included some long messages from my peers:

[slack-chan]

And when I click run, the Vertex model produced:

[vertex-out]

Much cleaner and more concise. Way easier and faster to digest.

Simple hosting and deployments

Once I completed my Slack Bot, the question was... how do I share it?
There are a variety of cloud providers, but I have never used them.
And Googling them leads to more questions than answers. Everyone
labels everything differently.

Finally, most of them function off Git. Replit can use Git too, so
that is not a huge issue, except for the fact that I do not know how
to use Git.

This is where Replit deployments comes in. I can deploy directly from
the editor. No need for other vendors. Now that I have done the hard
work (creating the project), sharing is made easy. And because I have
Replit Pro, I already have allocation for Autoscale deployments, so I
do not need to pay anything additional.

This is WAY easier than anywhere else. Plus, it includes analytics
and more.

You can fork this template, and if you add your Slack API Keys (in
secrets), it will work right away. No API keys for the models
required. And it's free!

I also attached some dummy data, so people are not seeing our
internal Replit conversations. But it works! If you'd like to try it
out, reach out to me directly on X (@Jeff_Burke14), and I will get
you set up.

Helpful resources

  * Start 100 Days of Code
  * Fork the template

                     Replit Logo Follow @Replit

Legal

Terms and services Privacy Subprocessors DPA US Student DPA

Replit

Blog About Jobs Teams Pricing

Handy links

Create a repl Docs Status Page Python Packages Import From Glitch

Social media

Facebook Twitter Tiktok Instagram

Languages

Clojure Haskell Kotlin QBasic Forth LOLCODE BrainF Emoticon Bloop
Unlambda CoffeeScript Scheme APL Lua Ruby Roy Python Node.js
JavaScript Deno (beta) Golang C++ C C# F# HTML, CSS, JS Rust Swift
Python (with Turtle) Basic (beta) R Bash Crystal Julia Elixir Nim
Dart Reason Node.js Tcl Erlang TypeScript Pygame Love2D Emacs Lisp
(Elisp) PHP Web Server SQLite Java PHP CLI Pyxel Raku Scala (beta)
Nix (beta)

Copyright (c) 2023 Replit, Inc. All rights reserved.