@firefly0621/dsh-skill-karpathy-guidelines

Readme

# @firefly0621/dsh-skill-karpathy-guidelines

English | [中文](README.zh.md)

Optional bundled skill provider that contributes `karpathy-guidelines` to `ctx.skills`. The skill body is adapted from [multica-ai/andrej-karpathy-skills](https://github.com/multica-ai/andrej-karpathy-skills) (`skills/karpathy-guidelines/SKILL.md`): think before coding, simplicity first, surgical changes, and goal-driven execution. Upstream package metadata and skill frontmatter declare MIT; see `assets/NOTICE` for the recorded provenance (the upstream tree currently has no root `LICENSE` file).

Mount the plugin to enable the provider. It has no configuration. This package is not part of the official main `dsh-base` composition; install it explicitly with `dsh plugin add` or a custom `cordis.patch.yml`.

The loaded definition carries `alwaysApply: true`. That flag alone does not inject the body: mount `@firefly0621/dsh-skill-always-apply` (opt-in; not in default `dsh-base`) so the consumer contributes the packaged instructions to the system prompt of every model request. Without that consumer, the skill remains catalog + on-demand `skill` tool load only.

The provider exposes its packaged `assets/` directory as the skill resource base. Project-local standing orders remain the job of `dsh-agent-instructions` (`AGENTS.md` / `CLAUDE.md`).

## Model Experience

Indirectly, through `@deepseek-ai/dsh-tool-skill`, which renders the catalog entry and selected skill body. With `dsh-skill-always-apply` mounted, also through that consumer's always-apply system-prompt section.

#### KV Cache effect

When mounted, its catalog entry and any loaded or always-applied body change the provider KV prefix at their insertion points. Setting the composition row to `disabled: true` removes that effect.

## Known Limitations and Deferred Work

- The provider contributes one fixed skill and has no runtime customization.
- Upstream CLAUDE.md / Cursor rule packaging is not mirrored; only the skill body is bundled for the dsh skill catalog.
- Provenance follows upstream-declared MIT metadata rather than a root LICENSE file in that repository; operators who need a stronger redistribution record should re-verify upstream licensing before redistribution.
- Always-on injection requires a separate always-apply consumer; this package does not inject on its own.