Skip to main content
We strongly recommend you follow the below 4 recommendations to make Hercules AI as accurate as possible.
  1. Create new chats for each new task: AI agents perform worse in longer conversations
  2. Use Plan Mode / Debug Mode: Have Hercules interview you to better understand your intent before making changes
  3. Include screenshots & videos with capture tools: Share visual context like screenshots, annotations, and video recordings to improves accuracy
  4. Tell Hercules the end result you want, not how to get there: Tell Hercules your requirements. Then let it work out how to get there
Hercules app editor with AI chat

​
1. Create new chats for each new task

AI agents perform worse in longer conversations. Longer conversations accumulate context that can confuse the Agent and cause it to make mistakes. Start a new conversation for new tasks like new features, bug fixes, or making changes to separate parts of the app Use the cmd+shift+O hotkey to quickly create a new chat. Learn more in multi-agent
New chat button showing cmd+shift+O hotkey

​
2. Use Plan Mode / Debug Mode

Hercules offers different agent modes to optimize the agent for specific tasks.
  • Plan mode is best for building new features and large changes. It interviews you to understand your intent, confirms requirements, then creates a detailed implementation plan.
  • Debug mode is best for fixing bugs and when build mode isn’t solving your issue. It interviews you to understand the issue, researches the codebase, forms hypotheses, then prioritizes the fix.
Plan and Debug mode are cheaper, faster, and more accurate than the default mode for accomplishing your goal. Learn more in agent modes
Use Plan or Debug mode for more accurate results

​
3. Include screenshots & videos with Capture tools

Hercules sees code, not the visual user interface that you see. It may think it implemented something correctly when in reality it can’t perfectly see what it’s done. Use Capture tools to give the Agent precise visual context of what you’re seeing
  • Screenshot: show the exact UI you want to change
  • Annotate: annotate a screenshot with comments and drawings
  • Screen record: create a video of you reproducing a bug or interaction that doesn’t work the way you expect
  • Select element: select a specific HTML element that isn’t working as you expect
Capture tools menu showing Screenshot, Annotate, Screen record, and Select element options

​
4. Tell Hercules the end result you want, not how to get there

The Agent performs best when you tell it the end result you want. Don’t tell it how to get there, tell it your end goal. If you over-specify implementation, you fight against better solutions it would choose. We recommend you give Hercules instructions in the format of 
users should be able to ...
e.g.
  • “Users should be able to complete checkout in 2 clicks.”
  • “Users should be shown a short, simple, seamless onboarding experience after login”
  • “Users should be able to easily add or remove new contacts”

​
Summary

  1. Create new chats for each new task
  2. Use Plan Mode / Debug Mode
  3. Include screenshots & videos with Capture Tools
  4. Tell Hercules the end result you want, not how to get there

​
Additional FAQ

Don’t think of talking to AI like talking to a human. AI is not a human. Think of AI as a statistical model. It takes your input, follows a mathematical pattern, and produces an output. The output seems human, but it’s not. Because AI is not human, you need to use the tricks above to get it to output the results you want.
Old context from earlier requests piles up. This context may no longer be relevant but still influences the Agent’s decisions. Starting fresh removes stale context.
No. AI is a novel technology. It will always make mistakes. The tricks above minimize errors but cannot eliminate them.
The four above are the main ones. You can find other tricks online. We strongly recommend you don’t threaten the AI (it can actually make it perform worse!).