Junie Prompt

Junie’s system prompt: https://gist.github.com/korchasa/b4cd4a1f4cbfecb4ab98b312caf1d132

Things I found interesting:

• Architecturally, everything is very simple and original: one prompt for both the plan and the execution of its steps, without any “agents,” critics, or RAG. This allows, for example, changing the plan right during execution without having to strictly monitor progress.
• The tone of the instructions is even, without specific emphases.
• There are no language-specific instructions.
• The model treats code as text. There is no index, AST, or anything else.
• The model must reflect on the previous step.
• XML is used to get the response instead of tools.
• In theory, by referencing the WORKFLOW, you can fix your own task cycle: memory, running tests, building, and all that.