We also recommend our and the rest of this section for a deeper understanding of effective prompting strategies and examples. Prompting in Databutton
Starting prompts and strategies
Outline Before Execution:
I need to create/modify XYZ. Don’t write code. Outline how you will complete the task.
Step-by-Step Execution:
Proceed. Take this one step at a time. Do not deviate. Confirm each step before continuing.
Memory Jogging:
Let’s work on #xyz. Remind me why this functionality was important in the first place.
Saving Workflows: ( Optional and if necessary )
Save this workflow as a .txt file within internal storage for future reference.
Debugging Prompts:
Basic Debugging:
Read code in the console. Come up with a plan to fix it. Do not write code yet.
Addressing Errors:
Initial Analysis
Take a step back and tell me what errors you are experiencing. Don’t write code. Respond.
Follow-up
Proceed with corrections. Do not deviate. One step at a time. Confirm before continuing.
Post-Debugging Analysis:
Recall all errors and analyze their causes. Was it due to my prompt or the rules?
Was there a simpler fix that could have been included in the original prompt?
Handling Agent Loops
Prompts to Resolve Loop Issues:
Breaking the Loop:
It still isn’t working. Write a report about the issue: what you have tried, how the behavior works in the app, and your proposed fix. DO NOT WRITE CODE YET.
Optimal Approach:
Can you tell me if this specific update is necessary or if there’s a more optimal approach to solve the same problem?
Subtasks for Clarity
Create and save subtasks after analyzing the problem. Perform each subtask step by step.
Specialized Prompts for Debugging
Log Analysis
Scan the error log and provide an outline of the issues. Do not write code yet.
Log Visualization
Color-code logs for easier identification:
- 🔵 Real-time updates received
- 🟢 State updates completed
- 🟡 Initialization events
- 🔴 Errors
Refactoring Workflow
Review the workflow and suggest improvements. Respond with a plan before implementation.
"Super" prompts ( termed by pro Databutton Builders)
"It still is not working. Take a step back and write a report about the issue I have detected, what you have tried, how the behaviour works in the app and what you are going to do to fix it. DO NOT WRITE CODE YET."
Can you confirm where we are with the current issue? Identify what has been completed, what still needs to be done, and provide a clear summary of the current status. DO NOT WRITE CODE YET.
Enhancing Collaboration with AI Agents
Improving Efficiency
Please scan the codebase for unnecessary logs and suggest optimizations. Do not write code yet.
Testing Updates
Based on the updates made, provide step-by-step user interaction instructions to verify everything works as expected.
General Prompting Tips by Community experts
General Tips:
Use the # function to accurately refer to what you are talking about.
Include/upload references to your prompt;
UI: go find examples of what you want and take a screenshot (styles you like, designs you like, draw a layout on a piece of paper and snap a photo). Upload the sample to your prompt. The agent will get pretty close most times with minor revisions.
Backend/frontend code: Include a link to a website, reddit post, knowledge base, article, python documentation. This information helps the agent immensely.
Knowledge is power and the more knowledge or insight you can give the agent, the better the outcome/design/functionality. But don’t confuse knowledge for prompt length. Keep your prompt reasonable.
When a complex, multistage workflow is being used, always create rules for the workflow. Place the rules in the agent instructions located in settings. I label the instructions with the same # naming convention as the api. Example: “#supabase_models rules”.
Don’t be afraid to ask the agent questions before you proceed. Learn from the agent. Ask it why it is doing something a certain way. You will be surprised how many times you learn something really cool or the agent responds with “Thanks for mentioning that, here is a better way to do this”.
If you feel like the agent may be wrong or you need another sounding board, use ChatGPT to give a comparative response.
If your prompt gets too long or you feel like you aren’t efficiently describing the request, use either the agent or ChatGPT to revise the prompt for you. “Please review my prompt. Please reformat the prompt so it is optimized for you (or an AI agent) to understand. Don’t write code. Respond.”
