Feature Request: Support for Custom Prompts

[aa956]

Currently, the system prompts are hardcoded in the following locations:

• internal/assets/prompts/instruction-planning.md
• internal/assets/prompts/instruction-coding.md

As a result, any modifications to these prompts require building a custom version of late-cli.

Proposal:
Add a way to specify custom prompt paths at runtime via parameters or environment variables. This would allow the prompts to reside within the user's project directory or any other arbitrary path.

Reasoning/Use Case:

I am currently investigating the Andrej Karpathy LLM Wiki pattern using Obsidian. I want late-cli to have access to my Obsidian vault; while the default coding prompts already work reasonably well, I would like to customize them to further emphasize data processing.

Everything else is working great - the bash tool is a perfect match for this workflow.

[mlhher]

I am not sure I understand fully. You can set the orchestrator prompt yourself using the flags and you can inject things into it using --append-system-prompt "...". If you mean something else please explain it.

I am currently investigating the Andrej Karpathy LLM Wiki pattern using Obsidian

I was also looking at this recently it is quite interesting but couldn't get to more yet.

[aa956]

I mean, instruction-coding.md is never updated? All three options, -system-prompt, --system-prompt-file and LATE_SYSTEM_PROMPT only modify planner's prompt?

"You are typically asked to write code, refactor functions, or fix bugs in specific files." - basically the only thing that may need updates.

Although I'm not sure if this will help, maybe Qwen3.6 is trained on coding tasks too much, when asked to rewrite documents it often tries to write script for that instead of just editing the file.

Edit: Actually good idea, maybe just changing the model to some of the Gemmas or Qwen 3.5 will be better idea for non-code tasks.

[mlhher]

This is not as straightforward as changing the orchestrator system prompt due to the architecture.

Can you give an example what you'd want the subagent to do here explicitly (e.g. a specific task where handing it off would help)? For coding tasks the payoff is huge due to it being code, where editing text is rarely straightforward. For a simple natural language edit I am not sure there is much payoff (though I am always open if there is obvious benefit).

[aa956]

Sorry for the late response, had other things keeping me away from personal projects this week (late response :) ).

Looks like model change was the right decision, Gemma 4 properly writes to/edits markdown files without trying to script the changes (qwen3.6 as coding agent tried to do that in documentation project).

So no need to change coder system prompt for now, custom planner system prompt is enough.

Getting Gemma 4 planner to read AGENTS.md is the hard part, I've gone through 3 revisions of planner system prompt and it still tries to just ls -R and then immediately start the task, without ever checking project-specific instructions (things like "never use system python, always use 'uv run python'" or "this is a Sphinx based project, use reStructuredText instead of Markdown" are in AGENTS.md).

Well, there's nothing to do here apart from model adjustments (I'll try to adjust temperature next I think and get back to the system prompt next weekend).

[mlhher]

No worries. From my experience Qwen is significantly better for coding; Gemma seems significantly better for well anything else. The AGENTS.md part is interesting because the prompt explicitly instructs it to check for one. Potentially adding it to a CRITICAL INSTRUCTIONS heading might help.

In my experience DeepSeek and Google are the only entities that do not seem to train on distills from competitors. Might be related.