What are Skill Packs for AI platforms?
The distribution unit for new agent capabilities — bundled playbooks, prompts, tools, and evaluations that install like an app.
TL;DR
A Skill Pack is a packaged bundle — playbook, prompts, tool wrappers, example artifacts, evaluation rubric — that gives an AI platform a new, domain-specific capability in one install. It's what apps were to smartphones: the unit of distribution that lets experts teach the platform without touching its core.
Every useful software platform eventually faces the same question: how do new capabilities get added, and by whom? Spreadsheets got formulas. Browsers got extensions. Phones got apps. AI platforms are going through the same arc, and the answer taking shape is the Skill Pack. Short definition: glossary/ai-skill-pack. This is the long-form version.
The precise definition
A Skill Pack is a versioned, installable bundle that adds a new capability to an AI platform, composed of at minimum: a playbook (the procedure), role-shaped prompts (how the agents reason through the procedure), tool wrappers or MCP references (what external systems the procedure touches), example artifacts (what good output looks like), and an evaluation rubric (how to judge the work). Some Skill Packs also include test fixtures, a cost estimator, and a permission manifest declaring which data the skill may read or modify.
In plain English
Think of an AI platform as a generalist company — good at business work broadly, no deep expertise in your specific vertical. A Skill Pack is how you hire expertise into it. A "Real Estate Investing" Skill Pack packs in the domain-specific playbooks — how to screen deals, how to write LOIs, how to run a neighborhood analysis — along with the prompts, tools, examples, and rubrics that make the agents actually good at it.
You install the pack. Your CEO agent now has a new procedure in its repertoire. Next time you say "screen this off-market deal," the CEO invokes the pack's playbook, delegates to specialists with the pack's prompts, and evaluates the output against the pack's rubric. No code was written by you. The pack did the teaching.
The history
Three lineages converge here. The first is plugins: Slack bots, Zapier integrations, VS Code extensions — add-ons that give a host app a new capability. The second is prompt libraries: collections like Awesome ChatGPT Prompts or FlowGPT, which packaged prompts but not the surrounding tooling. The third is agent frameworks: CrewAI's crews, LangChain's chains, AutoGen's conversation templates — reusable multi-agent flows.
Skill Packs combine all three. The packaging spec started gaining traction in late 2024 and stabilized through 2025. Anthropic formalized one version as Claude Skills, which define a skill as a folder containing a SKILL.md instruction file, optional supporting files, and an optional executable that the model can invoke. Microsoft's Copilot Studio uses a similar model. Black Box's Skill Packs extend the idea to the full agent-company scope: not just an agent capability, but a playbook for how the CEO agent and specialists coordinate to deliver it.
How a Skill Pack is different from a plugin
A plugin is usually one tool. "Install the Slack plugin, now your agent can post to Slack." Useful, but shallow — the agent doesn't know when or why to post, only that it can.
A Skill Pack is a whole capability. "Install the Customer Success Skill Pack" gives your agents: a playbook for onboarding new customers, prompts for each stage, tool integrations (email, CRM, Intercom), example onboarding sequences that worked for other users, and a rubric for what excellent onboarding looks like. The agents don't just have a new button; they have a new job they know how to do.
Anatomy of a Skill Pack
- Playbook. The step-by-step procedure — what the CEO agent does when this skill is invoked. Usually a markdown file with a flowchart and decision points.
- Prompts. Role-shaped system prompts for any specialists the pack uses, written for this domain.
- Tools. MCP server references or inline tool definitions, usually with auth scaffolding.
- Examples. 3-10 worked examples showing inputs and excellent outputs.
- Rubric. What the Evaluator uses to grade the pack's outputs.
- Manifest. Metadata: version, author, tier, required permissions, cost estimate, supported platforms.
Why this matters for business owners
For operators, Skill Packs collapse onboarding. Instead of telling the agent team how to do every unique-to-you procedure — which would take months — you install the relevant Skill Packs and your team knows the work on day one. A real-estate agent installs the Listings Launch and Buyer-Lead Workflow packs. A podcaster installs the Episode Production and Sponsorship Outreach packs. Same platform, different packs, different competence.
For domain experts, Skill Packs are a new distribution channel. If you're the person who knows how to run a 7-figure Etsy shop, you can package your playbooks into a Skill Pack and earn revenue when other sellers install it. That's the marketplace thesis — and it's the main reason we think Skill Packs, not APIs, will be the commercially important surface for vertical AI expertise.
Real-world example
The "Newsletter Launch" Skill Pack, for a solo operator starting a paid newsletter:
- Playbook: 10-step procedure covering niche selection, positioning doc, landing page, lead magnet, first-issue draft, welcome sequence, referral program, paid-tier launch, analytics wiring, first-month retention plan.
- Prompts: Custom prompts for the Content specialist on newsletter voice, the Business Ops specialist on Beehiiv/Substack integration, the Evaluator on spam-score and deliverability.
- Tools: MCP references for Beehiiv, Substack, ConvertKit, Stripe, Mailchimp's deliverability checker.
- Examples: Five real newsletter-launch artifacts (with permission) from successful operators — positioning docs, landing pages, welcome sequences.
- Rubric: Deliverability, voice match, conversion best practices, legal (CAN-SPAM, GDPR), analytics coverage.
The operator installs the pack, tells the CEO agent "launch my newsletter using the template." The CEO runs the 10-step playbook, each step handled by the right specialist with the pack's prompts, every artifact evaluated against the pack's rubric. A month of contractor work, condensed into a session plus approvals.
How Black Box implements Skill Packs
Black Box ships with a set of first-party Skill Packs covering common operator verticals — newsletter launch, real-estate listings, SaaS onboarding, podcast production, and more. Packs are versioned, sandboxed, and declaratively describe which specialists they extend and which tools they require. We're building a marketplace so domain experts can publish packs and share revenue. See the features page for the current catalog; pricing shows pack access by tier.
Key takeaways
- A Skill Pack is a packaged bundle — playbook + prompts + tools + examples + rubric — that adds a capability to an AI platform.
- It's deeper than a plugin: not a single tool, a whole job the agents now know how to do.
- It's the natural unit for vertical expertise in AI platforms and the basis for a marketplace.
- Anthropic's Claude Skills (2025) defined one spec; Black Box's Skill Packs extend the idea to full multi-agent playbooks.
- For operators: faster onboarding. For experts: a new distribution channel.
Frequently asked questions
What's inside a Skill Pack?
Playbook, prompts, tool wrappers, examples, rubric. Sometimes tests and a cost estimator.
How is it different from a plugin?
Plugin adds a tool. Skill Pack adds a whole capability.
Can non-engineers author them?
Yes. Largely markdown and YAML.
Same as Claude Skills?
Conceptually close. Black Box Skill Packs extend the idea to the full multi-agent playbook.
Marketplace?
Yes — in development.
Related reading
Install expertise instead of learning it
Black Box ships Skill Packs for common operator verticals. Install one, ship on day one.
By Web4Guru · Published April 23, 2026