The Knowledge Graph Knows. The Skills Registry Does.
A custom AI agent is software that takes real action on your business. Action requires two things: knowing the state of the world (the knowledge graph tells the agent that) and being able to change the state of the world (the skills registry lets the agent do that).
A chatbot has neither. A vector-search RAG product has the knowing but not the doing. A workflow tool like Zapier has the doing but not the knowing. A real custom AI agent has both, and the skills registry is the part that handles the doing.
Every action your agent takes routes through a skill: send email via Gmail, create deal in HubSpot, file ticket in Zendesk, update inventory in NetSuite, dispatch technician in ServiceTitan, charge customer via Stripe. Each one is a discrete capability the agent can call.
What Is Actually in a Single Skill
A skill is a small spec file, usually markdown, that the agent reads at runtime. Three required components:
**1. The SDK reference.** A link to the public SDK or API documentation for the underlying tool. This is the non-negotiable part. Without an SDK reference, the agent writes plausible-looking code that fails silently because the model invents API calls that do not exist. With the SDK reference, the skill works first try because the model is grounded in actual API shapes.
**2. The input contract.** What does the skill need to be invoked? A URL? A customer ID? A structured object? Specify the input shape exactly. If you skip this, the agent will hallucinate an input contract and pass wrong data to the API.
**3. The success condition.** What does success look like? "Email delivered with 200 status" is a success condition. "Email composed" is not. The agent needs to know when it is done.
More advanced skills also include: error handling rules, retry policy, escalation triggers, idempotency requirements, audit logging hooks. But the core three (SDK reference, input contract, success condition) are what separate skills that work in production from skills that look plausible but fail under load.
This exact pattern is what makes Karpathy's CLAUDE.md repo go viral, what powers Anthropic's Skills product, what Eric Siu open-sourced in his 17-category marketing playbook, and what Sentie's substrate uses for our 248+ tool integrations. Convergent design.
Why One File Per Action
The skills-as-files pattern beats other architectures for three reasons:
**Foundation models read prose better than code.** A markdown skill spec explains what the action does in language the model can reason about. A buried YAML config or a function signature in a codebase gives the model less context.
**Plain text is human-reviewable.** Your team can read a skill file and understand exactly what the agent will do when it invokes that skill. Your security team can audit it. Your operations lead can edit it. No engineering team required to interpret it.
**One file per action is composable.** New skill needed? Add a file. Skill no longer used? Delete the file. Skill needs to be tightened? Edit the file. The unit of change is the unit of capability.
This is why the .md file format is becoming the default for agent specs. CLAUDE.md, AGENTS.md, SKILL.md, DESIGN.md - same pattern under different names.
How Many Skills Does an Agent Actually Need
More than you think.
A real custom AI agent for a service business typically uses 30-80 skills in production. A complex enterprise agent can run on 200+.
For an HVAC contractor as an example:
- **CRM skills**: create customer, update customer, search customers, log communication, attach quote, mark deal won/lost (6-10 skills) - **Field service skills**: create job, assign technician, reschedule job, mark job complete, attach photos, log parts used (8-12 skills) - **Communication skills**: send email, send SMS, place outbound call, log inbound call, send branded message via Mailchimp (6-8 skills) - **Accounting skills**: create invoice, charge payment via Stripe, send receipt, mark invoice paid, refund (5-7 skills) - **Scheduling skills**: check technician availability, propose appointment, confirm appointment, send reminder (4-6 skills) - **Inventory skills**: check part availability, order from supplier, transfer between trucks (3-5 skills) - **Industry-specific skills**: lookup IRA tax credit eligibility, calculate rebate amount, file rebate paperwork (3-5 skills)
Total: 35-55 skills for a reasonably complete HVAC agent. Each is a small spec file. The Success Manager configures them during the assessment.
Vendors who claim their agent can do "hundreds of things" via a single big system prompt are not building skills registries; they are hoping the foundation model improvises. That works in demos and fails in production.
Sentie's 248+ Skills Catalog
Sentie's substrate ships with 248+ pre-built skills covering the major tool categories every customer needs:
- **CRMs**: HubSpot, Salesforce, Pipedrive, Close, Copper, Zoho, Insightly, Freshsales - **Helpdesks**: Zendesk, Intercom, Freshdesk, Help Scout, Front, Crisp, Tidio - **Field service**: ServiceTitan, Housecall Pro, Jobber, FieldEdge, ServiceFusion, Workiz, Successware - **Accounting**: QuickBooks, Xero, NetSuite, Wave, FreshBooks, Sage - **Communication**: Gmail, Outlook, SMS via Twilio, voice via various providers, Slack, Discord, Microsoft Teams - **Marketing**: Mailchimp, ActiveCampaign, Klaviyo, Customer.io, ConvertKit, Constant Contact - **Scheduling**: Calendly, Acuity, ScheduleOnce, Google Calendar, Outlook Calendar, Calendly - **Payments**: Stripe, Square, PayPal, Authorize.net, the new Stripe Link agent wallet for [agentic commerce](/blog/agentic-commerce-and-the-stripe-link-wallet) - **Documents**: Google Drive, Dropbox, Box, DocuSign, PandaDoc, OneDrive - **Project management**: Asana, Linear, Notion, ClickUp, Jira, Monday.com
The 248+ count is the catalog as of 2026-05-21. The Success Manager adds custom skills for tools outside the catalog as part of the standard configuration during assessment.
New skills can be added in days, not weeks, because each skill is a single spec file (SDK reference + input contract + success condition) rather than a code project. If your tool has an API or Sentie can wire it into the substrate.
How to Evaluate a Vendor's Skills Registry
Five questions to test depth:
**1. Show me the skill catalog.** A real registry has a browsable list. If the vendor cannot show you, the skills do not exist as discrete units.
**2. Pick a skill and show me the spec.** What does the input look like? Where is the SDK reference? What is the success condition? A real skill is documented; a marketing claim is not.
**3. How fast can you add a new skill for a tool I use that is not in the catalog?** Days, not weeks, is the right answer. If the vendor says months or quotes you for a 'custom development project,' the skill model is not the unit of capability.
**4. How are skills tested?** Real skills get exercised against a representative input set before going live. "We run the model and hope" is not testing.
**5. Can my engineering team author skills?** Even if most customers will not, the option matters. If skills are sealed inside the vendor's code, you have a black box. If they are spec files anyone with API knowledge can write, you have a real platform.
The Skills Registry as a Moat
Custom AI agent vendors that publish their skill specs (or open-source them, like Eric Siu / Single Grain did with their 17-category marketing playbook) build a flywheel. Every customer benefits from every skill ever shipped. Every new tool integration improves the whole substrate.
Sentie's 248+ skills compound across the customer base. An HVAC contractor benefits from the same Stripe integration a dental practice does. The dental practice benefits from the same Mailchimp skill an ecommerce store does. The skill is configured per tenant (your HubSpot account, your Mailchimp API key) but the skill spec is shared.
This is part of why managed substrate beats DIY: the skill library is the cumulative work of every customer's deployment, not the work of any one customer's engineering team. Sentie's customers ride on 248+ integrations someone else's deployment paid the engineering cost to build.
[Start the free assessment](/onboarding) to see which skills your Business Brain would use for your specific business. Your Success Manager walks through the skill list during discovery.