dsh-skills
Generalized development-specification skills distilled from DeepSeek Harness's engineering conventions — defensive patterns, code conventions, testing tiers, prose standards, decision records, review and PR discipline.
- Stars
- 4
- Language
- Python
- Created
- Aug 18, 2026
- Updated
- Aug 18, 2026
Introduction
DeepSeek-Harness-Skills
A set of coding-agent skills distilled from the engineering conventions of DeepSeek Harness — executable standards your coding agent applies while coding, reviewing, testing, or documenting.
English · 简体中文
Get started
Install the set with the skills.sh installer, then pick the skills and agents you want:
npx skills@latest add arch3rPro/dsh-skills
Or just read one — each SKILL.md is a self-contained executable standard. Pick the discipline that matches what you're doing: code-conventions, testing-tiers, prose-standard, or any in the set.
The set
Ten skills across four categories. Every one is model-invoked — the agent reaches for it automatically when a trigger fires.
| Skill | Category | What it does |
|---|---|---|
| defensive-patterns | architecture | bug-class rules for lifecycle, concurrency, teardown, untrusted I/O |
| code-conventions | architecture | reversible side effects, explicit-over-implicit, config-over-hardcoding, branded ids |
| extension-points-and-seams | architecture | build a seam only when a capability has multiple implementations or consumers |
| testing-tiers | testing | pick the right tier, test the real entry path, verify the world not the self-report |
| prose-standard | documentation | contract-first prose: preserve propositions, drop reasoning transcripts |
| documentation-placement | documentation | one home per fact; tutorials vs references; budgets as guardrails |
| decision-records | process | ADRs with classification, lifecycle, mandatory alternatives |
| minimal-evidence-checks | process | the smallest check set that covers the diff; fix or explain failures |
| code-review | process | two axes (standards + spec); substantiated blockers over nits |
| pr-history-hygiene | process | lease-protected rewrites, deliberate labels, native stack merges |
Beta (in development)
Five further disciplines are grounded in DeepSeek Harness and under development. They are npx skills-installable but not yet part of the official plugin bundle.
| Skill | Category | What it does |
|---|---|---|
| postmortem | process | backward-looking failure record: executive summary, root cause, guardrails |
| runtime-invariants | testing | assert only authoritative state, never service/method presence; fail loud |
| responding-to-review-on-a-stack | process | fix on the introducing PR, propagate up-stack, re-audit after rewrites |
| event-contract-matrix | documentation | who produces and who listens to each event, bypass sites included |
| module-layering | architecture | derive and enforce the module dependency graph |
Where they come from
Every skill is distilled from a real convention in DeepSeek Harness and adapted so it works in any codebase — not only inside the harness it came from.
How it's packaged
Each skill is one <category>/<name> directory carrying three files in a single change:
SKILL.md— the executable standard (Claude Code / any agent harness)agents/openai.yaml— Codex metadata and invocation policydocs/<category>/<name>.md— a human-facing docs page
skills/
architecture/ how code is shaped and how failure is handled
testing/ how behavior is verified
documentation/ how prose and docs are written
process/ the workflow around code — decisions, review, checks, git
Invocation design
Skills are either model-invoked (the agent reaches for them automatically when a trigger fires) or user-invoked (only a human typing the skill's name can fire them — zero context load, but the human is the index). All ten skills here are model-invoked.
Installation
The two routes below are alternatives, not add-ons — pick one.
npx skills — any agent (Claude Code, Codex, …)
Install one skill at a time (or the whole set from Get started):
npx skills@latest add arch3rPro/dsh-skills --skill=code-review
npx skillspulls from GitHub — publish the repo first, then run the command above.
Claude Code — the plugin
Add this repo as a Claude Code marketplace, then install the plugin:
claude plugins marketplace add arch3rPro/dsh-skills
claude plugins install dsh-skills@dsh-skills
The plugin manifest lives in .claude-plugin/ (marketplace.json + plugin.json).
Acknowledgments
- DeepSeek Harness — an open-source agent harness developed by DeepSeek AI.
- mattpocock/skills — Matt Pocock's agent skills for real engineering.