Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.devin.ai/llms.txt

Use this file to discover all available pages before exploring further.

Devin Desktop serves a variety of use cases at a high level. However, we see certain use cases to be more common than others, especially among our enterprise customers within their production codebases.

​
Code generation

Guidance: Devin Desktop should work well for this use case. Devin Desktop features including single-line suggestions, multi-line suggestions, and fill-in-the-middle (FIM) completions.Best Practices: Ensuring usage of Next Completion (⌥ + ]), Context Pinning, @ Mentions, and Custom Context will provide best results.
Guidance: Devin Desktop should work well for this use case. Devin Desktop features including single-line suggestions, multi-line suggestions, and fill-in-the-middle (FIM) completions.Best Practices: Ensuring usage of Next Completion (⌥ + ]), Context Pinning, @ Mentions, and Custom Context will provide best results.
Guidance: Devin Desktop should work well for this use case. Devin Desktop features including single-line suggestions, multi-line suggestions, and fill-in-the-middle (FIM) completions.Best Practices: Ensuring usage of Next Completion (⌥ + ]), Context Pinning, @ Mentions, and Custom Context will provide best results.

​
Unit Test generation

Guidance: Basic usage of Devin Desktop for generating unit tests should reliably generate 60-70% of unit tests. Edge case coverage will only be as good as the user prompting the model is.Best Practices: Use @ Mentions. Prompt Engineering best practices. Examples include:Write unit test for @function-name that tests all edge cases for X and for Y (e.g. email domain).Use @testing-utility-class to write a unit test for @function-name.
Guidance: Good for low-hanging fruit use cases. For very specific API specs or in-house libraries, Devin Desktop will not know the intricacies well enough to ensure the quality of generated sample data.Best Practices: Be very specific about the interface you expect. Think about the complexity of the task (and if a single-shot LLM call will be sufficient to address).

​
Internal Code Commentary

Guidance: Devin Desktop should work well for this use case. Use Devin Desktop Command or Devin Desktop Chat to generate in-line comments and code descriptions.Best Practices: Use @ Mentions and use Code Lenses as much as possible to ensure the scope of the LLM call is correct.
Guidance: Generally the Refactor button / Devin Desktop Command would be the best ways to prompt for improvements. Devin Desktop Chat is the best place to ask for explanations or clarifications. This is a little vague but Devin Desktop should be good at doing both.Devin Desktop Chat is the best place to ask for explanations or clarifications.This is a little vague but Devin Desktop should be good at doing both.Best Practices: Use the dropdown prompts (aka Devin Desktop’s Refactor button) - we have custom prompts that are better engineered to deliver the answer you’d more likely expect.
Guidance: The best way to do this would be to create the header file, open chat, @ mention the function in the cpp file, and ask it to write the header function. Then do this iteratively for each in the cpp file. This is the best way to ensure no hallucinations along the way.Best Practices: Generally avoid trying to write a whole header file with one LLM call. Breaking down the granularity of the work makes the quality of the generated code significantly higher.

​
API Documentation and Integration

Guidance: This is similar to test coverage where parts of the API spec that are common across many libraries Devin Desktop would be able to accurately decorate. However, things that are built special for your in-house use case Devin Desktop might struggle to do at the quality that you expect.Best Practices: Similar to test coverage, as much as possible, walk Devin Desktop’s model through the best way to think about what the API is doing and it will be able to decorate better.
Guidance: Devin Desktop’s context length for a single LLM call is 16,000 tokens. Thus, depending on the scope of your search, Devin Desktop’s repo-wide search capability may not be sufficient. Repo-wide, multi-step, multi-edit tasks will be supported in upcoming future Devin Desktop products.This is fundamentally a multi-step problem that single-shot LLM calls (i.e. current functionality of all AI code assistants) are not well equipped to address. Additionally, accuracy of result must be much higher than other use cases as integrations are especially fragile.Best Practices: Devin Desktop is not well-equipped to solve this problem today. If you’d like to test the extent of Devin Desktop’s existing functionality, build out a step-by-step plan and prompt Devin Desktop individually with each step and high level of details to guide the AI.

​
Code Refactoring

Guidance: Ensure proper scoping using Devin Desktop Code Lenses or @ Mentions to make sure all of the necessary context is passed to the LLM.Context lengths for a single LLM call are finite. Thus, depending on the scope of your refactor, this finite context length may be an issue (and for that matter, any single-shot LLM paradigm). Repo-wide, multi-step, multi-edit tasks are now supported in Devin Desktop’s Cascade.Best Practices: Try to break down the prompt as much as possible. The simpler and shorter the command for refactoring the better.
Guidance: Ensure proper scoping using Devin Desktop Code Lenses or @ Mentions to make sure all of the necessary context is passed to the LLM.Devin Desktop’s context length for a single LLM call is 16,000 tokens. Thus, depending on the scope of your refactor, Devin Desktop’s context length may be an issue (and for that matter, any single-shot LLM paradigm). Repo-wide, multi-step, multi-edit tasks will be supported in upcoming future Devin Desktop products.Best Practices: Try to break down the prompt as much as possible. The simpler and shorter the command for refactoring the better.