GitHub

_A code assistant_
_(Formerly known as Clippy)_
Twitter thread
Getting started
1. Install Poetry.
2. Clone this repository.
3. Add the api key (OpenAI) to .env file: OPENAI_API_KEY=.... Optionally, you can add your SerpAPI
key to allow the model to use search: SERPAPI_API_KEY=
4. Install ctags.
5. For pylint, install it and pylint-venv.
6. Install dependencies: poetry install.
7. Run: poetry run clippinator --help. To run it on a project,
use poetry run clippinator PROJECT_PATH
8. You can stop it and then it will continue from the last saved state. Use ^C to provide feedback to the main agent.
Details
The purpose of **Clippinator** is to develop code for or with the user.
It can plan, write, debug, and test some projects autonomously.
For harder tasks, the best way to use it is to look at its work and provide feedback to it.
The tool consists of several agents that work together to help the user develop code. The agents are based on GPT-4.
Note that this is based on GPT-4 which runs for a long time, so it's quite expensive in terms of OpenAI API.
Here is the thing: it has a reasonable workflow by its own. It knows what to do and can do it. When it works, it works
faster than a human.
However, it's not perfect, and it can often make mistakes. But in combination with a human, it is very powerful.
Obviously, if you ask it to do something at very low levels of abstractions, like "Write a function that does X", it
will do it. It poses tasks like that to itself on its own, to a varying degree of success.
But combined with you, it will be able to do everything while only requiring a little bit of your intervention.
If the project is easy, you will just provide the most high-level guidance ("Write a link shortener web service"),
and if it's more complicated, you will be more involved, but **Clippinator** will still do most of the work.
Taskmaster
This tool has the main agent called _Taskmaster_. It is responsible for the overall development. It can use tools and
delegate tasks to subagents. To be able to run for a
long time, the history is summarized.
Taskmaster calls the specialized subagents (minions), like _Architect_ or _Writer_.
The taskmaster first asks some questions to the user to understand the project.
Then it asks the Architect to plan the project structure, and then it writes, debugs, and tests the project by
delegating tasks to the subagents.
Minions
All agents have access to the planned project architecture, current project structure, errors from the linter, memory.
The agents use different tools, like writing to files, using bash (including running background commands), using the
browser with Selenium, etc.
We have the following agents: _Architect_, _Writer_, _Frontender_, _Editor_, _QA_, _Devops_. They all have different
prompts and
tools.
Architecture
The architecture is just text which is written by the Architect.
It is a list of files with summaries of their contents in the form of comments, important lines (like classes and
functions).
The architecture is made available to all agents. Implementing architecture is the goal of the agents at the first