👉 Install [Visual Studio Code extension](https://github.com/devchat-ai/devchat-vscode) from [Visual Studio Marketplace](https://marketplace.visualstudio.com/items?itemName=merico.devchat) and enjoy DevChat 👏
DevChat is an open-source platform that empowers developers to leverage AI for code generation and documentation. We aim to go beyond simple code auto-completion and limited operations on code snippets. DevChat offers a highly *practical* and *effective* way for developers to interact and collaborate with large language models (LLMs).
While there are many AI coding tools available, we created DevChat based on our insights gained from generating tens of thousands of lines of code. If you agree with our perspectives outlined below, DevChat could be the perfect choice for you.
- **The value of prompt "engineering" is often overstated**. While a well-crafted prompt template can be beneficial, it doesn't justify spending days or weeks of study. Instead, dedicate an hour or two to create a few effective templates and share them with your team.
- The art of writing prompts is a skill honed through practice. It's not about templates or engineering, but about **refining individual prompts for specific tasks on a case-by-case basis**.
- **The bottleneck in harnessing AI's capabilities lies in how to embed the right context in a prompt**. This isn't merely about the token limit of an AI model's input. Even with an infinite number of tokens, existing AI models would struggle to yield satisfactory results without a proper separation of concerns.
- **Use AI only when it truly adds value**. Our misconception about AI's capabilities in reality is even a greater issue than hallucination of LLMs. What we need is a tool that boosts productivity, not merely an experiment.
In alignment with our insights, DevChat incorporates the following design choices:
- **Precise control over the context embedded in a prompt**. This isn't a feature to be overlooked in the quest for greater intelligence or autonomy. Precise control over context is crucial for effective AI use. In our view, most other tools tend to over-guess what a user needs to put into the context of a prompt. This typically introduces more noise than the Claude model it utilizes can effectively manage.
- **A simple, extensible prompt directory**. This allows developers or teams to easily integrate their own predefined prompt snippets into DevChat, avoiding significant engineering effort or a steep learning curve. You don't need [LangChain](https://github.com/langchain-ai/langchain) to make AI work for you.
- The most fundamental operation involves selecting code (either files or snippets) and adding it to DevChat. For instance, you can add a function along with an existing test case to the prompt context, and ask DevChat to generate several test cases for the function. The test case serves as a useful reference for DevChat, enabling it to understand how to write a valid test case in your environment, thus eliminating the need for you to specify every requirement or setup in your prompt.
- You can incorporate the output of any command, such as `tree ./src`, into a prompt with DevChat. For example, you can add the output of `git diff --cached` to DevChat, which can then generate a commit message for you.
- Program analysis can assist in building the necessary context. Suppose you want DevChat to explain some code to you. DevChat can perform better if it's aware of the dependent functions that the code calls. In this scenario, you select the target code to explain and add "symbol definitions" to the context. DevChat will then generate a prompt that explains the target code, taking into account the dependent functions.
DevChat uses a directory to organize predefined prompts (templates). You can add your own or customize them simply using an editor. Let's explore the directory and see how it works.
- **Location**: By default, the directory is named `workflows` and located in the `.chat` folder under your project root. If you've opened a project on VS Code, you can create a new terminal and run `ls .chat/workflows` to check what is in there.
- **Paths**: The directory contains three subdirectories, `sys`, `org`, and `usr`. They determine the priorities of prompt templates with the same name. That means you can overwrite existing prompts provided by the system. Take the following structure for example. If you create `commit_message` in `usr` and define your own `prompt.txt`, DevChat will use yours, instead of the one in `sys`.
```
workflows
├── sys
│ └── commit_message
│ └── prompt.txt
└── usr
└── commit_message
└── prompt.txt
```
Besides `sys` and `usr`, `org` is reserved for team-wise conventions. Suppose your team requires a specific format for a programming language. Your team can maintain a Git repository to store prompts in `org` and every team member can sync `org` with the repository. Those prompts will overwrite `sys` but you can still further customize them for yourself by providing any in `usr`.
- **Names**: You can include a prompt template by typing a "command" with the corresponding name to the DevChat input. The command is input by typing `/` and its name, as shown below for example.
Prompt templates can be organized in the directory hierarchy. Suppose you want to write a general `code` template with specific requirements for different programming languages. The prompts can be organized as the following structure.
```
workflows
└── usr
└── code
├── prompt.txt
├── go
│ └── prompt.txt
├── js
│ └── prompt.txt
└── py
└── prompt.txt
```
The `/`-separated path to the prompt directory in `usr` is corresponding to a `.`-separated command name to cite the `prompt.txt` file. For example, `path/to/dir` is represented as `/path.to.dir` in the command input.
So, with the above directory, you can type `/code.py` in DevChat to include the `prompt.txt` in `usr/py/prompt.txt`. Note that `sys`/`org`/`usr` need not to be included in either path or command. DevChat will first look up in `usr` and then `org` and `sys` in order.
<sub>(This image is licensed by devchat.ai under a <arel="license"href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0 International License</a>.)</sub>
<sub>(This image is licensed by devchat.ai under a <arel="license"href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0 International License</a>.)</sub>
