AI Harness public skills

A curated public catalog of portable Agent Skills for planning, documentation, repository work, presentations, meetings, and day-to-day agent workflows. Install one skill at a time with GitHub's Agent Skills CLI, or browse the source before deciding.

Provenance and maturity: This repository is a generated public projection of a privately maintained canonical catalog. Only explicitly approved skills and their resolved dependencies are copied here. Publishing is active, but the catalog and preview installation tooling may continue to evolve. Each projection records its source revision and file digests in PUBLIC-PROJECTION-PROVENANCE.json.

Choose and install one skill

The primary supported path is GitHub's Agent Skills CLI, which discovers skills directly from this repository's skills/*/SKILL.md layout — no marketplace registration step required:

gh skill install sergeiwallace/skills spec --agent claude-code --scope user

Invoke the installed skill literally in Claude Code:

/spec describe how to add a health-check endpoint

/spec should return an implementation plan and wait for your approval before changing anything. Replace spec with any name in the catalog below, and --agent claude-code with your own host from gh skill install --help (Cursor, Codex, Gemini CLI, and many others are supported).

Preview a skill before installing it:

gh skill preview sergeiwallace/skills spec

A Claude Code plugin-marketplace distribution is planned but not yet published for this repository; gh skill install is the currently supported path. For other hosts, follow the host's own Agent Skills installation instructions and review the skill's requirements first. Some skills use repository-level shared packages or assets, so copying only a skills/<name>/ directory is not a universal installation method.

Included skills

The descriptions below come from each skill's SKILL.md frontmatter. Every public skill is licensed under Apache-2.0; requirements are shown only when they materially affect use.

Skill	Description	Requirements	Install
clean-task-panel	Remove all completed tasks from the CC TUI task panel.	Claude Code task-panel tools	gh skill install sergeiwallace/skills clean-task-panel --agent claude-code --scope user
explain	Build a grounded mental model of a design or plan document, issue, code target, or technical concept. Use for explanation, refresher, rationale, comparison, trace, consequence, or learning-mode requests where faithful relationships matter more than compression.	—	gh skill install sergeiwallace/skills explain --agent claude-code --scope user
graphic	Provider-neutral pipeline for small graphic assets (icons, favicons, simple marks) — inspect/brief, route to reuse or constrained SVG primitives, sanitize+normalize through an SVG allowlist and SVGO, render a contact sheet, and gate on human selection before packaging. Phase one only; vector generation, tracing, and raster generation are deliberately out of scope.	Node.js and npm	gh skill install sergeiwallace/skills graphic --agent claude-code --scope user
jira	Jira operations — create issues, search, update, link, comment, sprint overview	Jira MCP connector and access	gh skill install sergeiwallace/skills jira --agent claude-code --scope user
report-feedback	Prepare privacy-scoped bug reports or feature requests for public GitHub and authenticated Claude Code feedback, with an explicit human gate before any GitHub write.	uv; gh for the optional GitHub route	gh skill install sergeiwallace/skills report-feedback --agent claude-code --scope user
presentation	Author a local, review-gated Slidev presentation from a normalized event brief and optional research.	Node.js, npm, and Python 3	gh skill install sergeiwallace/skills presentation --agent claude-code --scope user
meeting-analysis	Turn a zip or directory of meeting sources (recording, vendor transcript, notes export) into a template-conformant analysis doc. Cross-checks the vendor transcript against an independent local ASR pass to recover mangled product names, reads screen-share frames natively, and records what did not work. Use for "analyze this meeting" or "write up the standup".	ffmpeg, ffprobe, and pdftotext	gh skill install sergeiwallace/skills meeting-analysis --agent claude-code --scope user
research-doc	Write a self-contained, template-conformant research document with Chain-of-Verification grounding, adversarial review, live citation validation, a provenance ledger, and claim-tier tags. Use for a scoped "research X and write it up" request, inline or through an installation-provided research adapter.	uv and a live research/fetch capability	gh skill install sergeiwallace/skills research-doc --agent claude-code --scope user
spec	Present an implementation plan for review before making changes	—	gh skill install sergeiwallace/skills spec --agent claude-code --scope user
summarize	Faithfully compress caller-selected documents, issues, code, topic packets, or explicit session batches for a declared reader and purpose.	—	gh skill install sergeiwallace/skills summarize --agent claude-code --scope user
sync-git	Safely synchronize safe uncommitted work and local/remote branch state across selected repositories and worktrees; discard and gitignore recommendations always require human approval.	Git and uv	gh skill install sergeiwallace/skills sync-git --agent claude-code --scope user

How skills work

Each skill is a directory centered on a SKILL.md file, with optional scripts, templates, references, and assets beside it. The Agent Skills specification defines the portable structure and metadata that gh skill install and other Agent Skills-compatible tooling discover automatically. A skill with repository-level shared package dependencies (see packages/) resolves them as part of its own closure — install the skill itself rather than copying files by hand.

Skill instructions do not override your responsibility to review proposed commands, file access, network access, credentials, or external writes. Host support also does not imply that every optional external tool used by every skill is installed.

Trust and safety

Curated does not mean guaranteed safe. Skills can contain instructions and executable scripts, and dependencies can change risk. Before installation, inspect the linked skill directory and its generated plugin, confirm that you trust this repository and the recorded source revision, and check requested tools and permissions. Use least-privilege credentials and keep secrets out of prompts, fixtures, issue bodies, logs, and generated artifacts.

Public candidates are reviewed for portability, private references, dependency closure, license consistency, executable content, secrets, compatibility, and relevant evals before projection. Those gates reduce known risks; they are not a security certification. The authoritative projection record is PUBLIC-PROJECTION-PROVENANCE.json, and each generated plugin carries its own RELEASE-PROVENANCE.json.

For a suspected vulnerability or malicious skill, use GitHub's private security advisory flow. Do not disclose sensitive details in a public issue.

Contributing

This public repository is a generated projection, so a direct edit to projected skill content may be overwritten. For a new skill, behavioral change, or catalog correction, open a public proposal describing the use case, portable behavior, requirements, and license/provenance.

Maintainers review accepted changes in the private canonical source, add eligible skills to the private public-candidates.toml allowlist, run the portability, licensing, security, secret, dependency, and eval gates, and then regenerate this repository. Public eligibility is an explicit review decision; adding a folder or opening a proposal does not add it to the catalog automatically.

Use the issue tracker for ordinary bugs and questions. Security reports follow the private route above.

License

Apache License 2.0 for the projected skill content, code, and generated plugins. See LICENSE and NOTICE. Source revisions, dependency closure, and file digests are recorded in PUBLIC-PROJECTION-PROVENANCE.json.