Skip to main content

​
Plan first, then implement

Asking Bezi to jump straight into implementation can produce the wrong outcomes and waste credits. Use Plan Mode to find the best approach before execution.

​
Specific prompts = better outcomes

Bezi cannot read minds; specific prompts take fewer prompts to deliver accurate results. Tag Pages, set constraints, and define what you consider “success”.
  1. Define the current state
  2. Define what needs to be done
  3. Define the desired end-state
  4. Define how Bezi should measure successful completion of the task
  5. Point Bezi to resources: tag Pages, pin related parts of the project, reference existing patterns (in any projects), share suggestions or ideas, etc.
Before submitting a prompt, read it in isolation. Is it very clear what the prompt refers to, what Bezi needs to do, and the resources and success criteria to use?

​
Give Bezi rich data to inform decisions

Use @ to reference in-project context, pages, and external Connections
  1. Add in-line pins to define project references (assets, scripts, scenes, etc.)
  2. Pin Pages to give Bezi guidelines or documentation to follow
  3. Attach images (UI mockups, screenshots, etc.) directly to the prompt
  4. Set Rules to standardize how Bezi navigates and interacts with the Workspace

​
Low context window = high performance

Performance degrades and prompts take more credits as the context window fills. Always check a thread’s context window tracker (bottom right of the prompt field) before prompting. Try these thread context modifications if the context window is full:
  1. The thread’s own prompt and response history; start new threads every ~10 prompts
  2. All active Workspace Connections (Unity projects and local data sources); disable Connections that are unrelated to a given task

​
If you get responses you don’t like, start a new thread

Bezi ingests the active thread’s prompt and response history as context. If Bezi gives you a response you don’t like and you prompt on top of it, Bezi may reference the wrong response as “relecant context” for its next attempt.