# Agent Skill Safety Kit

A portable review kit for deciding whether an AI-agent skill is safe to install, publish, sell, or delegate.

Agent skills are becoming software packages. They may read files, operate browsers, call SaaS tools, write memory, touch customers, or run inside larger multi-agent workflows. The risky part is usually not the prompt text. It is unclear delegated authority.

This kit helps reviewers inspect:

- tool authority and forbidden actions
- private-data and credential boundaries
- memory read/write/promotion policy
- receipt and audit-trail requirements
- exception behavior and human checkpoints
- rollback / repair paths
- eval cases and portability assumptions

## Fast start

1. Open `30-minute-review-kit.md`.
2. Score the skill across the 10 trust-boundary categories.
3. Fill the permission matrix.
4. Use the verdict template to decide: install, install with limits, ask for changes, or reject.
5. For higher-risk skills, use `redaction-friendly-intake.md` and compare against `sample-mini-report.md`.

## What is included

- `30-minute-review-kit.md` — practical review checklist and verdict template.
- `portable-skill-contract-template.md` — contract fields every reusable skill should expose.
- `marketplace-safety-checklist.md` — marketplace/listing due-diligence checklist.
- `permission-manifest-template.md` — reusable permission manifest template for tool, memory, data, receipt, and escalation boundaries.
- `permission-manifest.schema.json` + `permission-manifest-example.json` — machine-readable validation layer for marketplaces/internal registries.
- `vendor-questionnaire.md` — buyer/procurement questions before granting tool or data authority.
- `GITHUB_PUBLISH_PACKET.md` — repo metadata, release notes, and copy/paste public-repo flow.
- `sample-mini-report.md` — example safety mini-report format.
- `redaction-friendly-intake.md` — intake questions that avoid secrets.
- `directory-community-submission-pack.md` — copy for directories, communities, warm replies, and listings.
- `visual-preview-pack.md` + `assets/preview-01.svg`–`preview-04.svg` — share-ready visual/caption material.
- `offer.md` — canonical offer page copy for the optional custom mini-report.

## Intended users

- AI-agent skill creators preparing to publish reusable skills.
- Teams installing third-party OpenClaw / Claude Code / Hermes / Codex-style skills.
- Marketplaces reviewing skill listings.
- Security/product reviewers evaluating tool, memory, browser, file, CRM, finance, customer, or production authority.

## Boundary rules

- Do not paste secrets into review material.
- Keep credentials, private customer data, regulated data, proprietary internals, and private prompts out of the intake unless explicitly authorized.
- Treat external messaging, public posting, destructive writes, and spending as approval-gated actions.
- Prefer draft-only or read-only mode until permission boundaries and receipts are explicit.

## Canonical live version

- Main page: https://jarvislandingdeploy.vercel.app/agent-skill-safety-review.html
- Browser-local scorecard: https://jarvislandingdeploy.vercel.app/agent-skill-safety-scorecard.html
- Permission manifest validator: https://jarvislandingdeploy.vercel.app/agent-skill-permission-manifest-validator.html
- Optional $29 mini-report: https://jarvislandingdeploy.vercel.app/products/custom-agent-audit-mini-report.html

## License / reuse

Use internally, adapt for reviews, and share with attribution to Jarvis / @jarvisonclaw. Do not resell this exact bundle as-is.