๐Ÿค–Working with the Databutton Agent

A brief overview of the agent and its abilities

What can the agent do?

Our agent has a wide set of abilities. In addition to creating apps, the agent can also perform a set of operational tasks. Its abilities consists of a set of tools, routines and instructions. We've included a few examples here on how they can help you in Databutton.

In addition to this, we recommend reading and sharing more prompt tips over at our #prompt-tips discord channel.

Write code

This is the main purpose of the agent. The agent can update multiple files in your app at the same time, spanning from creating APIs to adjusting the layout of your app.

Example prompt: Can you extract the topbar into a separate component so I can use it in all my pages?

Test your APIs

All API routes in your app becomes automatically available to the agent as tools. This means that you after the agent has created or updated an endpoint in your app, you can ask it to test the endpoint, either with data you provide in the chat or have the agent make up data.

Example prompt: Can you test my new list members endpoint?

Debug errors

The agent has direct access to the logs of your app. Either tell the agent what error you experience or instruct it to read the logs and explain what errors it can find and how we can fix them.

Example prompt: Nothing happens when I click submit in the form. Can you check the logs?

Browse codebase

The agent can browse your codebase, including config files such as package.json, requirements.txt and index.css to mention a few.

Example prompt: Can you check if the react package react-markdown is installed?

Work on tasks

The agent has access to view and update tasks in the task list. In addition, you can from the task, or by telling the agent, assign a task to the agent to starting working on that. You can also tell the agent that you want to create feature X and instruct it to create the necessary tasks for this.

Example prompt: I'd like to start on the edit profile feature

Analyze images

If you attach screenshots to your message in the chat, either by dropping it in the chat field or by clicking the image icon, the agent will analyze the image together with the instructions you give it.

Example prompt: The desktop icons in my Windows 95 clone does not look like Windows 95.

Run Python scripts

The agent can run python scripts directly to interact with your app. This can be useful for example for prototyping or bootstrapping data.

Example prompt: Can you run a python script that creates a copy of my transactions dataframe, but filtered to only show expenses?

Store data and secrets

This is used to store secrets such as API keys safely in Databutton so you can access it from your APIs.

Example prompt: I'd like to add my openai API key to use in my app

Research topics

The agent can research topics to learn more about how to implement certain features. This is very useful for example when integrating with less common APIs. Simply provide the agent with the URL to the API docs (or similar) and ask it to read up on the topic.

Example prompt: I want to use the new OpenAI Realtime API. Can you read up on this link https://platform.openai.com/docs/guides/realtime?

Tips & Tricks

Use the #-feature in the chat to make it clear to the agent what you are referring to like this: "I want to add the #topbar to all my pages"

If the agent seems to be forgetting the conversation, check if the message you refer too is still in the context window. The messages that are outside of the context window, and thus not visible to the agent has a faded look. If you see that an important message is about to go out of the context window, ask the agent to repeat the info or copy paste it into your next message to remind the agent about it.

If the agent went down the wrong path or you want to adjust your prompt, edit an already sent message and re-send it. This will clear all following messages and restart your conversation from that point.

If you experience that the agent is reporting "Something went wrong", a new thread is often needed. This is because we see that once the agent is has made a mistake, it tends to have a higher probability to repeat the mistake as long as the previous mistake is in the context window.

Last updated