DSH HUB
HomePlugin StorePlugin PacksCommunityRankingsResourcesPublish Guide
Plugin source
Back to catalog

auggie246 /

auggie246/dsh-mattpocock-skills

Topic repository only

Install Matt Pocock's skills into DeepSeek Harness (DSH) — download fresh, patch for DSH, copy as real directories.

★ 0 Stars1 Forks0 IssuesN/A Community rating0 Confirmed installs
View on GitHub
READMESource: main@475da19f

dsh-mattpocock-skills

Install Matt Pocock's skills into DeepSeek Harness (DSH) — download fresh, patch for DSH, copy as real directories.

standard-readme compliant

Table of Contents

  • Background
  • Install
  • Usage
  • The ask-user overlay (patches/)
  • The mattpocock-skills preset
  • Maintainers
  • Contributing
  • License

Background

This repo distributes Matt Pocock's skills for DeepSeek Harness (DSH). Upstream skills were written for Claude Code. Two adaptations are needed for DSH, and both live here:

  • an installer that speaks DSH's model of skill installation, and
  • the ask-user overlay, which reroutes "ask the user" moments through DSH's native tool.

DSH has no skill-plugin install command. Skills are files discovered by the skill-filesystem plugin row that agent presets mount in their agent.cordis.yml. That row scans, in priority order:

Rank Root Scope
100 <project>/.dsh/skills current project
200 <project>/.agents/skills current project
300 customSkillDirs (preset config) that preset only
400 ~/.dsh/skills every preset, globally
500 ~/.agents/skills every preset, shared with other harnesses

So "installing skills" = placing skill directories in one of those roots. No DSH restart is needed: the filesystem provider watches its roots and refreshes the model's skill catalog live. This repo installs into ~/.dsh/skills (rank 400) so the skills are available in every preset. Lower-rank roots (e.g. a project's own .dsh/skills) still override per name.

The installer downloads both trees fresh on every run and copies the chosen skills in as real directories — never symlinks — so what lands in ~/.dsh/skills is a snapshot, not a live view into a working tree. DSH discovers one level deep (<root>/<skill>/SKILL.md), so upstream's skills/<category>/<skill> layout is flattened on the way in.

Install

No clone needed — run the installer straight from the web:

curl -fsSL https://raw.githubusercontent.com/auggie246/dsh-mattpocock-skills/main/install.sh | sh

Dependencies

  • POSIX sh (the script is dash-compatible; it is piped to sh, not bash)
  • curl or wget for the downloads
  • tar to unpack the GitHub tarballs
  • git or patch to apply the overlay patches — without either, skills install unpatched with a warning

Updating

Re-run the same one-liner any time. Every run downloads the latest upstream skills, applies the current overlay patches, and copies the result over the previous install (snapshot semantics). Skills that upstream removed, or that you deselected, are deleted. To pin either source to a branch, tag, or sha, see --ref and --upstream-ref under Usage.

Usage

Default install covers the engineering and productivity categories:

curl -fsSL https://raw.githubusercontent.com/auggie246/dsh-mattpocock-skills/main/install.sh | sh

Flags go after -s -- when piped (a bare sh --flag feeds the flag to the shell, not the script):

# choose categories explicitly
curl -fsSL https://raw.githubusercontent.com/auggie246/dsh-mattpocock-skills/main/install.sh | sh -s -- --categories "engineering productivity misc"

# everything except deprecated/ and in-progress/
curl -fsSL https://raw.githubusercontent.com/auggie246/dsh-mattpocock-skills/main/install.sh | sh -s -- --all

# remove every skill this script installed
curl -fsSL https://raw.githubusercontent.com/auggie246/dsh-mattpocock-skills/main/install.sh | sh -s -- --uninstall

# pin the skills source / the patch source (branch, tag, or sha)
curl -fsSL https://raw.githubusercontent.com/auggie246/dsh-mattpocock-skills/main/install.sh | sh -s -- --upstream-ref <ref> --ref <ref>

CLI

Flag Effect
--categories "<list>" Install only these upstream categories (default: engineering productivity).
--all Install every category except deprecated/ and in-progress/.
--uninstall Remove every manifest-managed entry. Offline; needs no downloads.
--upstream-ref <ref> Pin the mattpocock/skills download (default HEAD).
--ref <ref> Pin this repo's patch download (default HEAD).
-h, --help Print usage.

The installer manages only entries listed in ~/.dsh/skills/.mattpocock-skills.manifest:

  • re-runs refresh manifest-owned entries and remove ones no longer upstream or deselected,
  • anything else in ~/.dsh/skills is never touched,
  • a name collision with a non-managed directory is skipped with a warning.

The ask-user overlay (patches/)

Upstream skills were written for Claude Code and say "ask the user" in chat prose. DSH ships a default plugin (@deepseek-ai/dsh-tool-ask-user) whose model-facing tool, ask_user_question, is the proper channel for those moments. This repo carries a DSH-local overlay that says so inside every skill that can ask:

  • patches/ask-user-question/<skill>.patch holds one patch per skill. Each appends a DSH note: asking the user section to a SKILL.md: 15 skills in all (10 engineering, 3 productivity, 2 misc) whose instructions contain a genuine ask-the-user step. Grilling's ❓/➡️ round format included. Rhetorical "ask"s (tdd's "Ask: what's the public interface?") are deliberately untouched.
  • install.sh downloads both trees fresh on every run, applies every patch under patches/ to the downloaded tree, and only then installs. So upstream releases land first, and the overlay rides on top of the new version. A stale patch is skipped with a warning, never a broken install. One patch per skill caps the blast radius of an upstream change at that one skill.

When a release rewrites the tail of a patched file, that patch goes stale and that skill stays unpatched until you regenerate it: from a pristine upstream-skills/ checkout, append the overlay section to the file, then rewrite the patch with git -C upstream-skills diff -- skills/<bucket>/<skill>/SKILL.md > patches/ask-user-question/<skill>.patch. (This ran for real on 19 Aug: a 110-file upstream release landed mid-session and every patch went stale at once; the overlay was re-based onto the new tree in minutes.) The overlay is a fork-local adaptation, not upstream content, so upstream's docs-page and router sync rules stay untriggered. The patched note reaches DSH on the next install run, since every run downloads and patches a fresh tree.

The mattpocock-skills preset

~/.dsh/.agent-presets/mattpocock-skills/ is an agent preset (a bundle of Cordis plugin rows), not a plugin. It used to carry copies of these skills wired through customSkillDirs; that was removed in favor of this global install, since preset-local copies (rank 300) would shadow the live global copies (rank 400) and go stale. The preset now just provides the persona and the standard tool/session composition; the skills themselves come from ~/.dsh/skills, exactly as in every other preset.

Maintainers

  • auggie246 — owner and maintainer.

Contributing

Issues and pull requests are welcome: open an issue at https://github.com/auggie246/dsh-mattpocock-skills/issues to ask questions or propose changes. PRs are accepted.

Requirements for contributing:

  • Regenerate stale overlay patches with the workflow in The ask-user overlay — do not hand-edit installed skills in ~/.dsh/skills; they are snapshots and the next run overwrites them.
  • Keep install.sh POSIX sh compatible: it runs under dash via curl | sh, so no bash arrays or bashisms.
  • Test installer changes against a throwaway home first: DSH_HOME=$(mktemp -d) sh install.sh.

License

MIT © Augustine (auggie246)

The vendored upstream-skills/ checkout keeps its own upstream license from mattpocock/skills; it is a source tree for patch development, not part of what this installer ships.

—/ 5

No ratings yet

Manifest verification required

Commit 475da19f58ad

Community comments

No comments yet. Be the first to write one.

DSH HUB

A community index for DSH plugins. Not an official GitHub or DeepSeek AI product.

CommunityResourcesAPIAbout