From harness-toolkit
Use when the user wants to CREATE, AUDIT, or IMPROVE a Claude Agent Skill. Create scaffolds a new SKILL.md. Audit runs the R01-R20 rubric and emits severity-tagged findings. Improve applies fixes from an audit report. Mode is the first argument.
How this skill is triggered — by the user, by Claude, or both
Slash command
/harness-toolkit:skill-toolkitThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
A thin router that dispatches to one of three modes for managing Claude Agent Skills. This file stays small on purpose — each mode's real instructions live in `references/mode-*.md` and are loaded only when that mode runs (progressive disclosure).
A thin router that dispatches to one of three modes for managing Claude Agent Skills. This file stays small on purpose — each mode's real instructions live in references/mode-*.md and are loaded only when that mode runs (progressive disclosure).
In scope: creating new SKILL.md files, auditing existing skills against a 20-rule rubric, applying audit findings as fixes.
Out of scope: building plugins end-to-end (this only handles skills), installing/distributing skills (use /reload-plugins yourself), managing subagents or slash commands (different primitives).
Bash is in allowed-tools)The delegated mode files need Bash for two narrow purposes — nothing else:
date -u +"%Y-%m-%dT%H:%M:%SZ" — UTC timestamp at the top of every audit report (see references/report-template.md).basename/dirname — deriving a skill's directory name for the R01 check ("name must equal containing directory name") in mode-audit.md.The router itself does not invoke Bash. It is declared at the router level so modes loaded via Read references/mode-*.md inherit the permission. No other shell usage is permitted.
Before running any mode, read the shared rubric so the same validation rules apply everywhere:
references/rubric.md — the 20 rules (always load)references/frontmatter-schema.md — per-field spec (load when needed)Parse the first whitespace-separated token of $ARGUMENTS as the mode. If $ARGUMENTS is empty or the mode is unknown, ask the user which mode they want (create / audit / improve) and a target.
create [name]Load references/mode-create.md and follow it. Default scaffold: .claude/skills/<name>/.
audit <path>Load references/mode-audit.md and follow it. Read-only — MUST NOT Write or Edit. <path> can be a SKILL.md, a skill directory, or a plugin root.
improve <path> [--report=<file>]Load references/mode-improve.md and follow it. If no report given, run audit inline first.
create: suggest skill-toolkit audit <path-to-new-skill> as an independent sanity check.audit with CRITICAL/HIGH findings: suggest skill-toolkit improve <path> --report=<report>.improve: always re-audit and report the residual.Not yet validated — update as real failures are discovered.
allowed-tools includes Write/Edit because create/improve need them. In audit mode, you must self-enforce the constraint.references/anti-patterns.md.Provides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub jason-hchsieh/marketplace --plugin harness-toolkit