2 posts tagged with "prompt engineering"

An agent decides which tool to call almost entirely from one thing: the tool's description. Not its name, not its parameters — the sentence or two that says "Read the contents of a file in the project" is what the model weighs when it picks read_file over run_command.

Until now those descriptions were string literals compiled into the plugin. You could switch a tool off, but you couldn't tell the model anything about the tools that remained. So "use ./gradlew, never a global gradle" or "prefer semantic search over grep in this repo" meant repeating yourself in every prompt, or editing the plugin.

You can now rewrite any built-in tool's description from Settings.

The system prompt is the most powerful setting in DevoxxGenie and the one nobody touches twice. You write a good one, it works, and then it quietly becomes wrong for half of what you do. The prompt that makes a model a great pair programmer — write the code, be concrete, show me the diff — is exactly the prompt that makes it a mediocre reviewer, because a reviewer that starts rewriting your method has stopped reviewing it.

The usual workaround is to argue with the model in the prompt itself: "don't implement anything, just review". It works about as often as it doesn't.

Personas, new in DevoxxGenie 1.12.0, fix this the boring way: give each role its own system prompt, name it, and put the names in a dropdown.