it generates entire code instead of targeted changes

[AlejandroSanchez90]

sometimes, when working with buffers with 200+ lines, and I ask a simple question, it generates the entire buffer code again plus the answer in it, is there a way to just render the answer ?

[deathbeam]

Are you using /COPILOT_GENERATE? it tries to usually do smaller changes, but ofc there is only so much we can do if AI starts halucinating

[syaghoubi00]

The system_prompt used by several of the default prompts use /COPILOT_GENERATE, as @deathbeam mentioned.

The prompts using /COPILOT_GENERATE are: Fix, Optimize, Docs, and Tests.

CopilotChat.nvim/lua/CopilotChat/config.lua

Lines 335 to 346 in ade6c93

	Fix = {
	prompt = '> /COPILOT_GENERATE\n\nThere is a problem in this code. Rewrite the code to show it with the bug fixed.',
	},
	Optimize = {
	prompt = '> /COPILOT_GENERATE\n\nOptimize the selected code to improve performance and readability.',
	},
	Docs = {
	prompt = '> /COPILOT_GENERATE\n\nPlease add documentation comments to the selected code.',
	},
	Tests = {
	prompt = '> /COPILOT_GENERATE\n\nPlease generate tests for my code.',
	},

You can see in rules 1, 3, 5 of COPILOT_GENERATE prompt specifies to return all the code.

CopilotChat.nvim/lua/CopilotChat/prompts.lua

Lines 66 to 91 in ade6c93

	M.COPILOT_GENERATE = M.COPILOT_INSTRUCTIONS
	.. [[
	Your task is to modify the provided code according to the user's request. Follow these instructions precisely:
	1. Return *ONLY* the complete modified code.
	2. *DO NOT* include any explanations, comments, or line numbers in your response.
	3. Ensure the returned code is complete and can be directly used as a replacement for the original code.
	4. Preserve the original structure, indentation, and formatting of the code as much as possible.
	5. *DO NOT* omit any parts of the code, even if they are unchanged.
	6. Maintain the *SAME INDENTATION* in the returned code as in the source code
	7. *ONLY* return the new code snippets to be updated, *DO NOT* return the entire file content.
	8. If the response do not fits in a single message, split the response into multiple messages.
	9. Directly above every returned code snippet, add `[file:<file_name>](<file_path>) line:<start_line>-<end_line>`. Example: `[file:copilot.lua](nvim/.config/nvim/lua/config/copilot.lua) line:1-98`. This is markdown link syntax, so make sure to follow it.
	10. When fixing code pay close attention to diagnostics as well. When fixing diagnostics, include diagnostic content in your response.
	Remember that Your response SHOULD CONTAIN ONLY THE MODIFIED CODE to be used as DIRECT REPLACEMENT to the original file.
	]]

---

The fix is likely to generate your own system_prompt and use that instead.

This could be accomplished with something like this:

  prompts = {
    ShortReply = {
      system_prompt = 'You are an expert coder who provides short concise answers to questions.',
    },
    ShortFix = {
      prompt = '/ShortReply There is a bug in this code, find it and fix it',
    }
  }

[deathbeam]

I improved the default /COPILOT_GENERATE prompt to try to output code blocks instead of whole text now that we support proper headers (and hopefully AI stops forgetting to add them constantly) it should be better

[syaghoubi00]

I've had a pretty good experience following Anthropic's recommendation of using XML tags to structure my prompts. They seem to work well with other models too. You can see an example of one that I've been using for coding in my dotfiles.

[syaghoubi00]

Another thing I've noticed is that when doing a diff on code blocks it wants to delete all the other code, but that might just be a bug with using markdown renderer for the output. So, outputting the whole code can be a useful workaround for that, but it's probably better to output only changed blocks by default.