Is there a sample agent instructions

[warm river]

Officially, no but we're working on it. You can try pointing an agent to the llms.txt or llms-full.txt from the docs site. Unoffically, I have some examples that work well enough for my use cases here https://github.com/kpenfound/dagger-programmer/blob/main/moduleWorkspace/knowledge/go_sdk.md (and others in that directory)

[zinc ember]

Cool! Let me try those. I agree, documentation around this would be really helpful!

[zinc ember]

Hey @warm river any update on this? It would be really nice to have something official. LLM code gen for dagger is still hit or miss. It keeps trying to read the generated code and directly using Go tooling.

[warm river]

Yes will get back to you on this tomorrow 👍 What i have is the daggerizer agent that can essentially bootstrap dagger modules for you and write good dagger code in any SDK. The project itself doesn't do anything (yet) to assist in modifying dagger code, but the primitives are there to help piece something together

[zinc ember]

Awesome! To clarify, I am looking for something akin to AGENTS.md or copilot-instructions.md or a prompt that can generate those files with Dagger specific instructions. I had some success manually creating that file, adding https://github.com/kpenfound/dagger-programmer/blob/main/moduleWorkspace/knowledge/go_sdk.md verbatim, and then asking VSCode to generate the instructions file.

It looks at the existing file and then the project and updates it.

At the moment, I won't be able to use Daggerizer or the dagger llm primitives as we don't have access to API based LLM models. Everything is through Github copilot.

sry my copy/paste got messed up

[warm river]

yeah, the daggerizer has a bunch of better prompts similar to that one, but it also has sub-agents that do smart things. For example, one is a sub-agent specialized in the dagger docs and can answer questions about "how do I create a Service with php" with code snippets. So giving that agent to the main agent saves a ton of context and tokens in the main agent. It's a bit tricker if you don't have api based LLMs but it could probably run on local models too

[zinc ember]

We can't do local models either sadly. I don't think copilot supports multi agents yet. Hopefully soon.

I am hoping all of the details can be condensed into a single instructions file (or a few based on filetypes).

[warm river]

can copilot use mcp servers? thats an easy way of doing multi agent

one thing thats tricky with trying to get it all in a reference prompt is if you have any dependency modules then copilot has to either have a working lsp integration, some way to know how to use the schema of the dependency, or spend a bunch of tokens guessing

We can't do local models either sadly

🙀 not even with docker desktop?

[zinc ember]

Not even with that. I know it's silly.