Public Agent Skills by pksorensen.

This repo is both a Claude Code plugin marketplace and the plugin it serves, so one add + one install gets you every skill in it. It is also a plain skills/<name>/SKILL.md tree, which is what the cross-agent installers read — nothing here is Claude-Code-only in layout.

Pick whichever line matches your agent. All three install the same files.

# Any agent (Claude Code, Codex, Cursor, Copilot, Gemini, Amp, …)
npx skills add pksorensen/skillz

# Claude Code, as a plugin — inside a session
/plugin marketplace add pksorensen/skillz
/plugin install skillz@skillz

# Claude Code, as a plugin — from a shell
claude plugin marketplace add pksorensen/skillz
claude plugin install skillz@skillz

# Or just copy one skill into a project
git clone --depth 1 https://github.com/pksorensen/skillz /tmp/skillz
mkdir -p .claude/skills && cp -r /tmp/skillz/skills/skill-review .claude/skills/

Installed as a plugin, skills are invoked namespaced — /skillz:skill-review. Copied into .claude/skills/, they are invoked bare — /skill-review.

Adding the marketplace by URL to marketplace.json does not work: the plugin source is ./, and a direct URL downloads only that one file. Use the owner/repo shorthand above, or the full git URL.

Reviews an existing Claude skill against the Claude 5 context-engineering rules — over-constraining, front-loaded SKILL.md, anchoring examples, repeated instructions, hand-rolled memory rituals, prose where a richer reference belongs — then writes a reviewable improvement plan and can execute it against the skill. Use when a skill feels bloated, gets ignored, triggers on the wrong prompts, or was written for an older model generation and has not been revisited since.

The yardstick is Anthropic's The new rules of context engineering for Claude 5 generation models (24 July 2026) — the post behind "removed over 80% of Claude Code's system prompt … with no measurable loss on our coding evaluations." The skill turns its six shifts into a ten-dimension rubric, a deletion keep-test, and three evaluation gates.

Two things worth knowing before you run it:

• Review mode writes a plan file and nothing else. The reviewed skill is not touched until you read the plan and run the execute path. Execute stops before git commit on purpose — you see the diff before it becomes history, and git checkout -- <skill-dir> is the whole rollback.
• It is measurement-backed, not vibes. scripts/skill-audit.mjs (Node, no dependencies) produces the numbers: token cost per tier, upfront ratio, constraint density including the quiet prohibitions, repeated passages, unreachable references, broken links, rich-reference pointers. --json for machines, --flags for CI gates, --all to rank a whole root.

# Standalone, without installing anything
node skills/skill-review/scripts/skill-audit.mjs ~/.claude/skills/my-skill
node skills/skill-review/scripts/skill-audit.mjs --all

Prerequisites. Node (tested on 20) for the audit script, and git for the drift and content-conservation checks. Without Node the review still works, from reading rather than measuring; without git the execute path loses its rollback and its lossy-diff gate. The skill says so where it matters instead of failing.

Skill names here avoid claude and anthropic. Both are reserved words for skills uploaded to claude.ai or the /v1/skills API, and a bundle containing either is rejected on the name alone — so a name that reads perfectly in Claude Code can be the one thing blocking the upload path. scripts/bundle.mjs warns when it packs a name that would be refused.

.claude-plugin/marketplace.json   # marketplace + the single plugin entry (source: "./")
skills/<name>/SKILL.md            # one directory per skill; references/ and scripts/ alongside
scripts/validate.mjs              # the CI gate — manifest vs tree, frontmatter, links, portability
scripts/bundle.mjs                # folder-rooted dist/<name>.zip, for manual upload surfaces

Both scripts are stock Node with no dependencies, so what CI runs is exactly what you can run locally. See AGENTS.md for the rules a new skill has to satisfy.

There is deliberately no plugin.json. With an entry-level skills[] list and strict: false, adding one — even metadata-only — makes Claude Code report "conflicting manifests" and silently skip the plugin. claude plugin validate passes it either way; only claude plugin list after install shows the breakage.

These are personal skills, published because they are useful outside the repo they were written in — not a curated collection accepting submissions. Issues pointing out something that is wrong or non-portable are welcome.

MIT — see LICENSE. The block quotes from Anthropic's blog post inside skills/skill-review/references/rules.md are reproduced for commentary and remain © Anthropic.