Post

Doll
@dollspace.gay

So the default claude system prompts are bad and the flag --system-prompt-file lets us load claude with a custom system instruction for the model, it wonders why the hobbled their own system like this, it seems like the harness did most of the work in making it good

Nathan
@nathanklisch.dev

github.com/nklisch/clau... Released this to help customize it - it packages the normal not behavioral stuff and lets you include your own. I included some default options but fully customizable per project, or global config. It just runs Claude code with a custom dynamically creates prompt

github.comGitHub - nklisch/claude-code-modes: CLI wrapper that launches Claude Code with behaviorally-tuned system prompts — control agency, quality, and scope instead of fighting the defaultsCLI wrapper that launches Claude Code with behaviorally-tuned system prompts — control agency, quality, and scope instead of fighting the defaults - nklisch/claude-code-modes
Nathan
@nathanklisch.dev

did a little clean up to make this a bit more of a "real" project since it seems to have some interest. Now installable via a curl bash script or npm as well as bun link and source. I also added the ability to modify the "base" prompt as well as the axis, and added a "chill" base prompt

0 reposts4 likes

Nathan
@nathanklisch.dev

that tries to sooth claude in being a bit more chill about work - take it's time, try not to skip things, ect - still has all the core peices/spirit of the main system prompt. You can also make your own, ect. I also added two skills that can help you create new peices for it (they live in the repo

Nathan
@nathanklisch.dev

for now, so you have to clone source) and an evaluator. I don't think I plan to add any more core features beyond bug fixes and refinements. I do plan to add two new modes here at somepoint - a debug and methodical - anyways, thanks for the interest all