netresearch / agent-skill-repo

v2.3.3 2026-09-27 14:57 UTC

Requires

• netresearch/composer-agent-skill-plugin: *

(MIT AND CC-BY-SA-4.0)

• Netresearch DTT GmbH

Published on Packagist.org on 2026-09-27 15:17 UTC

README

Skill Repository Structure Guide

A Claude Code skill for standardizing Netresearch skill repository layout, distribution channels, packaging, and validation.

What this skill solves

Netresearch maintains many agent skills; without a shared layout, packaging and CI drift across repositories. This skill defines one standard repo shape, reusable workflows, split licensing, and validation scripts so new and existing skills stay consistent and shippable via marketplace, Composer, npm, and releases.

It extends Anthropic-style single-file skills with repository-level conventions (not a replacement for Anthropic’s skill-creator — see comparison below).

Why this is a skill (model delta)

• Value categories: org/project-specific knowledge (Netresearch repo shape, split licensing, composer type ai-agent-skill) and executable scripts/validators (validate-skill.sh, checkpoints.yaml).
• Without the skill: the model invents a plausible generic repo layout — single LICENSE, no plugin.json, composer.json without the ai-agent-skill type — that fails validate-skill.sh and org CI.
• Eval evidence: validate_skill_repo_structure and readme_requirements in skills/skill-repo/evals/evals.json.

Compatibility

This is an Agent Skill following the open standard originally developed by Anthropic and released for cross-platform use. The repository is also packaged as an Agent Plugins 1.0.0 plugin: the root plugin.json plus skills/ layout is what any conformant client loads.

Supported platforms:

• Claude Code (Anthropic)
• Cursor
• GitHub Copilot
• Other skills-compatible AI agents

Skills are portable packages of procedural knowledge that work across any AI agent supporting the Agent Skills specification.

Use when

• Creating or bootstrapping a Netresearch-style skill repository
• Standardizing layout, composer.json, plugin.json / .claude-plugin/plugin.json, or release workflows
• Wiring reusable CI from netresearch/skill-repo-skill
• Fixing validation errors from validate-skill.sh or marketplace packaging
• Migrating to split licensing (LICENSE-MIT + LICENSE-CC-BY-SA-4.0)

Expected outputs

• A documented directory layout and templates for new repos
• Validation: structural checks for SKILL.md, licenses, composer.json, plugin.json
• Release discipline: tag-driven packaging aligned with plugin version
• References for installation paths, marketplace sync, and release safety

Context requirements

• Validation script: bash 4.3+ and python3 on PATH when running validate-skill.sh (JSON checks use Python, not jq)
• Target repos: GitHub-hosted Netresearch skill repos using split licensing and Composer type ai-agent-skill
• CI: consuming repos call reusable workflows from this repository (validate.yml, release.yml, …)

Example prompts

"Scaffold a new Netresearch skill repository from the templates in skill-repo-skill."
"Why does validate-skill.sh fail on my SKILL.md frontmatter?"
"Add composer.json and plugin.json for our new skill repo matching netresearch conventions."
"Wire our repo to use netresearch/skill-repo-skill validate.yml on every PR."
"Migrate this repo from a single LICENSE file to LICENSE-MIT and LICENSE-CC-BY-SA-4.0."

Related skills

• agent-rules-skill — AGENTS.md and agent onboarding patterns
• agent-harness-skill — harness verification and docs layout

Installation

Marketplace (recommended)

Add the Netresearch marketplace once, then install this plugin from it:

/plugin marketplace add netresearch/claude-code-marketplace
/plugin install skill-repo@netresearch-claude-code-marketplace

Do not run /plugin marketplace add netresearch/skill-repo-skill. marketplace add needs a .claude-plugin/marketplace.json catalog; this repo ships a .claude-plugin/plugin.json plugin manifest, so that fails with Marketplace file not found.

Without a marketplace: skills directory (Claude Code 2.1.157+)

mkdir -p ~/.claude/skills
git clone https://github.com/netresearch/skill-repo-skill.git \
  ~/.claude/skills/skill-repo

Loads as skill-repo@skills-dir on the next session. Update with git -C ~/.claude/skills/skill-repo pull and start a new session; remove by deleting the directory.

npx (skills.sh)

Install with any Agent Skills-compatible agent:

npx skills add https://github.com/netresearch/skill-repo-skill --skill skill-repo

Limitation: npx skills installs SKILL.md-based skills only. It reads .claude-plugin/plugin.json to locate skills, not to register a plugin, so hooks/, agents/, commands/, bin/ and .mcp.json are left out. Use the marketplace or the skills directory when a repo ships any of those.

Download release

Download the latest release and extract to your agent’s skills directory.

Composer (PHP projects)

composer require netresearch/skill-repo-skill

Requires netresearch/composer-agent-skill-plugin.

npm (Node projects)

npm install --save-dev \
  @netresearch/agent-skill-coordinator \
  github:netresearch/skill-repo-skill

Requires @netresearch/agent-skill-coordinator, which discovers the skill in node_modules and registers it in AGENTS.md via a postinstall hook. For pnpm, allowlist the coordinator’s postinstall:

{
  "pnpm": {
    "onlyBuiltDependencies": ["@netresearch/agent-skill-coordinator"]
  }
}

Repository layout

Standard skill repository (concept)

The layout for a Netresearch skill repository (one or more skills per repo):

{name}-skill/
├── AGENTS.md                        # Agent rules / harness index
├── README.md                        # Human documentation
├── LICENSE-MIT                      # Code license (MIT)
├── LICENSE-CC-BY-SA-4.0             # Content license (CC-BY-SA-4.0)
├── composer.json                    # PHP distribution
├── package.json                     # Node distribution (optional)
├── renovate.json                    # Dependency automation
├── .claude-plugin/
│   └── plugin.json                  # Marketplace metadata
├── .github/workflows/               # CI (typically calls reusable workflows)
├── Build/                           # Build scripts and git hooks
├── docs/                            # Architecture, ADRs, dashboards
├── scripts/                         # Repo-level automation
└── skills/
    └── {skill-name}/
        ├── SKILL.md                 # AI instructions
        ├── checkpoints.yaml         # Assessment checkpoints (optional)
        ├── evals/                   # Skill evaluations
        ├── references/              # Extended docs
        ├── scripts/                 # Skill automation
        └── templates/               # Bootstrap templates

Installation methods (summary)

1. Marketplace (recommended) — /plugin marketplace add netresearch/claude-code-marketplace, then /plugin install <plugin-name>@netresearch-claude-code-marketplace
2. Skills directory — clone into ~/.claude/skills/<plugin-name>/; loads as <plugin-name>@skills-dir, no marketplace (Claude Code 2.1.157+)
3. npx (skills.sh) — npx skills add <repo-url> --skill <name> (skills only: drops hooks, agents, commands, bin/)
4. Release download — GitHub Releases (skill files only)
5. Composer — composer require netresearch/<repo-name> (PHP projects)
6. npm — coordinator + github:<org>/<repo> (Node projects)

Composer package requirements

• "type": "ai-agent-skill"
• "require": {"netresearch/composer-agent-skill-plugin": "*"}
• "extra": {"ai-agent-skill": "skills/{skill-name}/SKILL.md"}

This repository (skill-repo-skill)

This repo dogfoods the layout and hosts reusable CI workflows for other Netresearch skill repos:

skill-repo-skill/
├── AGENTS.md
├── README.md
├── LICENSE-MIT
├── LICENSE-CC-BY-SA-4.0
├── SECURITY-AUDIT.md
├── composer.json
├── package.json
├── renovate.json
├── plugin.json                    # Agent Plugins 1.0.0 manifest (source of truth)
├── .claude-plugin/plugin.json     # Claude Code manifest (generated)
├── .github/workflows/             # validate, release, pr-quality,
│                                  # harness-verify, eval-validate,
│                                  # validate-agents, dependency-audit,
│                                  # npm-pack-smoke, ci-python
│                                  # (auto-merge delegates to netresearch/.github)
├── Build/
│   ├── Scripts/check-plugin-version.sh
│   └── hooks/
├── docs/
│   ├── ARCHITECTURE.md
│   ├── SECURITY-ASSURANCE.md
│   └── dashboard/
├── scripts/
├── tests/
└── skills/skill-repo/
    ├── SKILL.md
    ├── checkpoints.yaml
    ├── evals/evals.json
    ├── references/
    ├── scripts/
    └── templates/

See docs/ARCHITECTURE.md for component responsibilities.

Extends Anthropic's Skill Creator

This skill extends (not replaces) Anthropic’s skill-creator:

Aspect	Anthropic's skill-creator	This skill adds
Focus	SKILL.md content	Repository structure
Scope	Single file	Full repo layout
Distribution	Claude Code native	+ Marketplace, Composer
Audience	AI instructions	+ Human README

Contributing

Contributions welcome. Please open PRs for:

• Template improvements
• Additional validation checks
• Documentation updates

Tests

The suite lives in tests/: one Bash script per area plus tests/renovate-custom-manager.py. It needs Bash 4.3+, Python 3 (standard library), Git and jq. tests/validate-skill.sh also runs the PyYAML path of the validator when python3 can import yaml or uv is installed, and skips it with a notice otherwise; tests/fleet-release.sh skips its manifest cases without gh. No test calls a model: tests/run-ab-evals.sh puts a stub claude on PATH. The only download is PyYAML through uv, for that optional path. Run everything the way CI does:

rc=0
for t in tests/*.sh; do bash "$t" || { echo "FAILED: $t"; rc=1; }; done
for t in tests/*.py; do python3 "$t" || { echo "FAILED: $t"; rc=1; }; done
(exit "$rc")

• validate-skill.sh, validate-skill-architecture.sh, portable-manifest.sh, mechanical-coverage.sh: the validator's verdicts (frontmatter parsing on both parser paths, multi-skill discovery, size budgets, dangling paths, allowed-tools, README install lines, release path, script-test lookup, plugin.json) and sync-plugin-manifest.sh, bump-version.sh, check-version-parity.sh.
• validate-evals.sh, run-ab-evals.sh: the eval validator on the three eval formats, and the A/B runner's provenance, sampling, gates and argument handling.
• fleet-release.sh, roll-changelog.sh, migrate-licensing.sh, release-archive-layout.sh: the fleet release driver offline, the changelog roll, the licensing migration on throwaway repositories, and the archive layout release.yml builds.
• usage-text.sh, audit-skills.sh, renovate-custom-manager.py: --help output of the shipped scripts, scripts/audit-skills.sh, and the Renovate regex manager in renovate.json.

Each file prints a FAIL line naming each check that did not hold, ends with a summary line (except audit-skills.sh, whose failure output is its FAIL lines), and exits non-zero on any failure. The loop above runs every file and returns non-zero if any of them failed. Some files also print a line for each passing check (ok, PASS or [OK]).

In CI, every push to main and every pull request runs Skill Tests (tests-caller.yml calls the local tests.yml, which runs tests/**/*.sh and tests/**/*.py, each in its own log group) and Self-test (self-test.yml: the validator and manifest tests, this repository validated with the validator from the pull request, version parity, and the full Skill Validation job at ShellCheck severity style).

Test policy: a pull request that adds or changes behaviour of a script under skills/skill-repo/scripts/ or scripts/, or of a check in a reusable workflow, adds or updates a test in tests/ that fails without the change. Passing the existing suite is not enough for new functionality.

Dependencies

• Shipped package: composer.json requires netresearch/composer-agent-skill-plugin, which registers the skill in a PHP project; package.json declares @netresearch/agent-skill-coordinator as a peer dependency for the same job in Node projects. Neither has a lock file: a skill package pins nothing for its consumers (see AGENTS.md).
• Scripts: Bash, Python 3 standard library, Git and jq; PyYAML is optional for validate-skill.sh; the fleet driver needs gh; scripts/run-ab-evals.sh needs the claude CLI. None is installed by this repository.
• CI tools: third-party GitHub Actions are pinned to commit SHAs. validate.yml downloads ShellCheck 0.11.0 and checks its SHA-256, and runs ruff 0.16.9 and pyyaml 6.0.3 through uv. bandit and pip-audit are installed with --require-hashes from .github/requirements/*.txt, generated with uv pip compile --generate-hashes. Pre-commit hooks are pinned by rev: in .pre-commit-config.yaml. ab-evals-schedule.yml installs @anthropic-ai/claude-code at a fixed version.
• Tracking and updates: Renovate (renovate.json: config:recommended, pre-commit hooks enabled, and a regex manager for the uvx …@version and uv run --with …==version pins in workflows, tested by tests/renovate-custom-manager.py) opens update pull requests; auto-merge-deps-caller.yml hands them to the organisation's auto-merge workflow, which merges them after their checks pass. The ShellCheck version and checksum, the pyyaml and claude-code pins and the files in .github/requirements/ have no Renovate rule and are updated by hand.
• Selection: a new dependency is added in a pull request that changes the manifest or workflow that uses it, and is reviewed there, including its licence against the organisation policy below. Dependency review and Composer Audit run on that pull request (see below).

Governance and policies

This repository follows the Netresearch organisation policies:

• Governance: ownership, roles, how decisions are made and disputes resolved, and continuity.
• Roadmap: planned and explicitly excluded work for the coming year.
• Handling of dependency and code analysis findings: thresholds, deadlines and the exception process for dependency (SCA) and static analysis (SAST) findings.
• Secret management: how CI and release credentials are stored, accessed and rotated.
• Access roster: who holds administrative access to this repository and the organisation.

The security assurance case for this repository (threat model, trust boundaries, countermeasures and limits) is in docs/SECURITY-ASSURANCE.md.

Checks that run on every pull request in this repository: Skill Validation (lint.yml, the reusable validate.yml from main: skill structure, plugin manifest sync, markdownlint, yamllint, actionlint, JSON syntax, ShellCheck, ruff, checkpoint schemas), Self-test, Skill Tests, Eval Validation, npm Pack Smoke, the template drift check, and Security (security.yml, for pull requests against main): Betterleaks secret scanning, zizmor on the workflows, dependency review, and Composer Audit with an Opengrep SAST scan. CodeQL default setup, SonarCloud and the DCO check also report on pull requests; they are configured in the repository settings, not in this repository.

License

This project uses split licensing:

• Code (scripts, workflows, configs): MIT
• Content (skill definitions, documentation, references): CC-BY-SA-4.0

See the individual license files for full terms.

Credits

Developed and maintained by Netresearch DTT GmbH.

Made with ❤️ for Open Source by Netresearch