anthropics
Include a built-in default context loader skill, or let us add to the system prompt
I've put a lot of effort into building my overall contact system for my Claude Code environment. That means providing lots of context around preferences, contacts, the personality of my agent, etc.
I've fully migrated to the skills system, but skills only invoke based on that particular task.
I tried creating a general skill that always executes, but that did not work.
So what I did is I built a context skill and had my session start hook call that. And that seems to work.
But I feel like this is still not ideal compared to you having it built in natively.
You can see the system here.
https://github.com/danielmiessler/Personal_AI_Infrastructure/tree/main
Modifying the system prompt with claude code is already possible, here's a snippet from a writeup I did on the methods I found, including the not yet documented --system-prompt.
- Output Styles: Replaces the "Tone and style" and "Doing tasks" sections. Gets placed near the start of the prompt, above tool definitions.
- --append-system-prompt: Adds your instructions right above the tool definitions. Stacks with output styles (output style goes first). Good for adding specific behaviors without replacing existing instructions.
- --system-prompt (NEW in v2.0.14): Replaces the ENTIRE system prompt except tool definitions and one line about being a Claude agent. This is the nuclear option - gives you almost full control but you're responsible for everything.
I've also been told its possible to remove tool definitions from the system prompt by adding them to the denied tools list, but I have not confirmed it, and CC will still assume it has some tools like the Task (Subagent) tool and try to use them. If you disable them, you would need to design your own replacements, or develop some sort of proxy tool to access the tools it should have, with customized, leaner tool definitions.
I've put a lot of effort into building my overall contact system for my Claude Code environment. That means providing lots of context around preferences, contacts, the personality of my agent, etc.
I've fully migrated to the skills system, but skills only invoke based on that particular task.
I tried creating a general skill that always executes, but that did not work.
So what I did is I built a context skill and had my session start hook call that. And that seems to work.
But I feel like this is still not ideal compared to you having it built in natively.
You can see the system here.
https://github.com/danielmiessler/Personal_AI_Infrastructure/tree/main
