Matt Pocock Agent Skills — for DeepSeek Harness (DSH)

English | 中文

⚠️ Unofficial community adaptation. This package is an adaptation, not the upstream project. The skill bodies were written by Matt Pocock (mattpocock/skills, MIT, © Matt Pocock) — we did not author them. We only re-arranged them into DSH's skill layout and wrapped them in a DSH bundle plugin. If you find them useful, star the upstream repo. License & attribution: LICENSE.

A DeepSeek Harness bundle plugin that ports Matt Pocock's "Skills for Real Engineers" set into native, discoverable DSH skills.

This is a skill-set adaptation, not a fork of the Claude Code plugin harness: the SKILL.md bodies are preserved as-is (same prompts, same process), re-arranged so DSH's one-level skill discovery (<root>/<name>/SKILL.md) finds them. Upstream keeps its skills in four buckets (engineering/, productivity/, misc/, in-progress/); this package flattens them into skills/ and records each skill's upstream bucket in upstream.json.

Synced with upstream mattpocock-skills 1.3.2 (commit 4ae2317, 2026-10-31). Adapted and verified against DSH 0.2.2-rc.1.

This package ships a curated set: 9 enabled skills (grill-with-docs, research, and the seven productivity skills). The other 30 upstream skills are disabled and can be switched back on through enabled-skills.json.

Contents — 9 enabled skills

enabled-skills.json lists what this package ships. Everything else upstream provides is disabled; see Disabled skills to restore one.

Engineering (2 of 20)

Skill	What it does
grill-with-docs	Grilling session that also builds the domain model and docs as it goes
research	Research a question against primary sources into a cited Markdown file

Productivity (7 of 7)

Skill	What it does
grill-me	Relentless interview to sharpen a plan or design
grilling	Grill a plan, decision, or idea until the design tree is resolved
handoff	Compact the conversation into a handoff document for another agent
teach	Teach a skill or concept across sessions in a stateful workspace
to-questionnaire	Turn a decision into a questionnaire for someone else to fill in
wait-what	Re-pitch a message that did not land, in plain language
writing-for-agents	Write documents (skills, AGENTS.md) that agents can actually follow

Each enabled skill is a single-level skills/<name>/ directory with its resources (agents/openai.yaml, scripts/*, reference .md) kept alongside. DSH discovers them exactly like any local skill.

Disabled skills (30)

Not shipped today, kept one edit away:

• Engineering (18): ask-matt, code-review, codebase-design, diagnosing-bugs, domain-modeling, implement, implement-spec, improve-codebase-architecture, pr, prototype, retro, setup-matt-pocock-skills, tdd, to-spec, to-tickets, triage, wayfinder, wizard
• Misc (4): git-guardrails-claude-code, migrate-to-shoehorn, scaffold-exercises, setup-pre-commit
• In-progress (8): chief-of-staff, claude-handoff, loop-me, setup-ts-deep-modules, software-factory, writing-beats, writing-fragments, writing-shape

To ship one, add its name to enabled-skills.json and run npm run sync:upstream -- --prune. node scripts/sync-upstream.mjs --all ignores the file and ships everything.

How to invoke each skill

DSH classifies skills by who may load them. The 9 enabled skills fall into two groups (verified against the real skill-filesystem provider):

Group	Skills	How to use
Model + user (3)	grilling, research, writing-for-agents	The model may load them on its own when a task matches; they also appear in the / slash menu for you to force-invoke.
User only (6)	grill-with-docs, grill-me, handoff, teach, to-questionnaire, wait-what	Only you can start them — pick them in the / slash menu, or type /name <args> in the composer. Most carry disable-model-invocation: true upstream.

Why most are user-only: upstream marks interactive coaching flows and work-intake flows as user-gesture-only so the model does not start a multi-session process unprompted. To let the model pick one up by itself, remove the disable-model-invocation: true line from that skill's SKILL.md.

Install

Requires DSH 0.2.2-rc.1 (the bundle pins @deepseek-ai/dsh-skill-filesystem@0.2.2-rc.1; bump the pin when you upgrade the runtime).

# from the GitHub release (latest published tarball)
dsh plugin --profile web add https://github.com/xiaoxiaosrm/dsh-mattpocock-skills/releases/download/v0.5.2/mattpocock-community-dsh-engineering-skills-0.5.2.tgz

# or from a local checkout / tarball
dsh plugin --profile web add file:./mattpocock-dsh-engineering

In DSH Desktop, install the same package from the Plugins view (or add it to the Desktop profile's package.json dependencies and dsh.profile.bundles) and restart the app.

Then restart the profile. Verify the bundle loaded and its skills are visible:

dsh --profile web --dump-config | grep -A5 'skill-filesystem-mattpocock'

The skills appear in the model's catalog alongside your normal local skills — on every DSH surface (web, TUI, headless). User-only skills additionally appear in the / slash menu after the profile restarts and a session (re)opens.

Note on agents/*.yaml resource files. The agents/*.yaml companions are Claude Code interface descriptors and are not interpreted by DSH. They ship only for fidelity to the source set and are inert. Several skills also reference docs/agents/issue-tracker.md and similar per-repo files created by /setup-matt-pocock-skills; run that skill (or provide an issue tracker) once in your project for the tracker-dependent skills (code-review, triage, to-tickets, to-spec) to resolve a spec source.

Updating from upstream

Upstream changes often; this repo keeps the sync mechanical and reproducible:

npm run check:upstream   # download upstream, validate all skills, report adds/updates/removes (no writes)
npm run sync:upstream    # write the enabled skills, refresh upstream.json
node scripts/sync-upstream.mjs --prune   # also delete skills that are not enabled
node scripts/sync-upstream.mjs --all     # ignore enabled-skills.json, ship all 39

The script downloads the upstream tarball, validates every SKILL.md against the DSH frontmatter contract (name equal to the directory and kebab-case, description present, kebab-case invocation keys), asserts the upstream plugin manifest's promoted set is covered, and only then writes. It ships only the names listed in enabled-skills.json. upstream.json records the synced commit, plugin version, and the enabled/disabled split.

Why a plugin vs. just copying to ~/.dsh/skills/

• Distributes and version-skills from one deliverable across machines/repos.
• Lives under your profile's node_modules, so dsh plugin remove cleans it.
• A single cordis.patch.yml registers its own dedicated host skill-filesystem instance (skill-filesystem-mattpocock) into DSH's global skill layer — so the set loads in every surface (web, TUI, headless). The bundle does not patch the shared skill-filesystem row: the web profile ships that row disabled ("presets own local discovery"), so appending customSkillDirs to it made the whole set invisible on web. Instead this bundle inserts a distinct instance with providerName: mattpocock and includeDefaultRoots: false, isolating its discovery to exactly this bundle's skills/ directory. Verified against DSH 0.2.2-rc.1's bundle patch baseUrl (the profile directory).

Note: the npm tarball intentionally ships only the runtime files (skills/, enabled-skills.json, upstream.json, cordis.patch.yml, LICENSE, README.md, README.zh.md). The full change history lives in the repository's CHANGELOG.md on GitHub.

License & attribution

MIT. The skill bodies are adapted from mattpocock/skills (MIT, © Matt Pocock). See LICENSE. If you find the originals useful, star the upstream repo; this package is a community distribution adaptation only.