Open Agent Skills · install in any compatible agent
Thirteen open-source skills turn experienced ways of researching, deciding, publishing, and maintaining systems into executable workflows: explicit steps, inspectable checkpoints, and defined outputs. Claude Code works; so does any client that supports the Agent Skills specification.
Every flagship makes the path between request and result visible: what it reads, when it asks, what it checks, and where it stops.
Research
Scope the question, gather multiple sources, maintain a registry, check citations, and deliver a report with conclusions, trade-offs, and explicit limitations.
13Use deep-research to compare three RAG architectures for our support teamDecision interview
grill-me-to-doc reads available evidence first, asks one decision per turn, explains its recommendation, resumes from saved state, and delivers PRODUCT-DOC before a hard approval stop.
Use product-manager in grill-me-to-doc mode for an AI hiring productPresentation
Confirm the story before polishing slides, then move through page briefs, registered layouts, visual drafts, and explicit visual QA.
Use deck-studio to turn this quarterly review into a consulting deckContent pipeline
Run article, image-prompts, layout, or full-pipeline to produce an article, provider-neutral image manifest, and WeChat-safe inline HTML. Prompts are not generated images, and nothing is auto-published.
Use wechat-article-writer full-pipeline for these notesDo not load all 13 by default. Pick one workflow you will use this week; the installer copies only that skill to the cross-client .agents/skills directory.
git clone --depth 1 https://github.com/staruhub/ClaudeSkills.git
cd ClaudeSkills
python3 scripts/install_skill.py deep-researchInstalled at ~/.agents/skills/deep-research
Use deep-research to compare three options and write a cited decision memo
Choose another workflow: replace deep-research with product-manager, deck-studio, or wechat-article-writer.
Project-only install: add --project. List every short name with python3 scripts/install_skill.py --list.
Claude Code native path: add --client claude-code. Invocation and discovery behavior otherwise depend on the host agent.
Real repository artifacts make the promise concrete. Verification stays labeled by what it actually proves.
A real Deep Research template with findings, citations, trade-offs, and explicit limitations.
A real Product Manager deliverable that preserves decisions, open questions, and the approval boundary.
Checked Directory and structure contracts for 13 curated skills.
Limit Not live business-task E2E for all 13.
Checked 91 schema and conflict cases across 10 skills.
Limit Not model-executed routing accuracy.
Checked Generators, rendered pages, rubrics, scores, and known defects.
Limit Not independent external certification.
Checked Per-skill reads, writes, network, shell, credentials, and deletion.
Limit Not a substitute for your own policy review.
python3 scripts/validate.py && python3 scripts/run_routing_evals.py && python3 scripts/validate_site.pyFour flagships plus focused workflows for coding, security, architecture, Three.js, PDF, sales, and AI-native product methods. Experimental work stays in the separate lab.
Inspect before installing
Skills can read files and run scripts. The security guide separates bundled script behavior from permissions the host agent may request, with a capability matrix for every curated skill.
deep-researchREAD · WRITE · NET · SHELLsecurity-auditREAD · WRITE · SHELLc-drive-cleanerREAD · DELETE** Deletion is guarded: every destructive action requires explicit confirmation.
A prompt disappears into a conversation. Each curated skill is a versioned folder with a SKILL.md and, when needed, references, scripts, assets, schemas, and negative cases.
No. The four flagships have deterministic contract tests, and Deck also uses real Chrome rendering plus PPTX assembly. Subjective future-model quality, live image providers, and WeChat publishing still require separate verification.
Confirm that the client scans .agents/skills, or use its native directory. The installer also strips the repository prefix so Geek-skills-deck-studio becomes the stable identity deck-studio.
Start with an issue. New skills incubate in lab/ and graduate after they meet the repository's quality gates and capability disclosure rules.
Build in public
Open an issue with a redacted input, output, and reproduction steps. New skills start in lab/ and graduate after meeting the repository gates.