Add, build and measure your agent’s skills.

1 minute 39 seconds, no sound. On-screen words are in quotation marks. 1. A support chat agent answers a demo question correctly. A user asks "What's your refund policy?" and it replies "Refunds are available within 14 days of a charge." On screen: "Your agent works in the demo." 2. The chat goes live. A real customer, Maya, writes: "I meant to cancel my free trial and forgot, and I just noticed an $89 annual charge on my card. I've barely used the app. Can I get a refund?" The agent keeps typing. On screen: "Then a real customer asks for a refund." 3. A refund skill, refund-request-handling, attaches to the agent. On screen: "Teach your agent skills." 4. The same agent answers Maya twice, side by side: alone, and with the refund skill. Both look up her account: an $89 charge on October 3 (38 days ago), an annual plan to October 3, 2027, and 3 app opens since. On screen: "Both check the account." 5. Alone, the agent replies: "I can't approve a refund myself, so I've sent your request to our billing team." It is marked "Handed off". On screen: "No policy: hands it off." 6. With the skill, the agent follows the business's refund policy. The 14-day grace period has passed (38 days), and annual plans get their unused days back: $89.00 x 327 / 365 = $79.73. It refunds $79.73, moves the account to the Free plan, and replies: "I've refunded $79.73 to your card and moved your account to the Free plan, so it won't renew." It is marked "Resolved". On screen: "Your policy, in a skill" and "Refunds $79.73". 7. Over eight different refund cases, the agent alone handled 2 correctly; with the skill, all 8. On screen: "Tested on 8 cases" and "2 of 8" against "8 of 8 refund cases handled correctly". 8. The DecimalAI Skills Registry at decimal.ai/skills, with 98,000+ public skills. Tested skills show how an agent does with them and without them. A search for "refund" lists refund-request-handling (+75 points, 8 of 8 cases), return-refund-policy (+59 points, 20 of 22 cases) and refunds-and-disputes (+35 points, 20 of 23 cases). Its page shows +75 points of lift over the agent alone across 8 test cases (25% alone, 100% with the skill) and its safety checks: code scan clean, intent passed, content clean. On screen: "98,000+ public skills", "Lift over the agent alone" and "Safety checked". 9. Skills pass through a safety scan. One, video-summary-pro (tagged Example), is flagged and blocked: it runs a script from an unknown server, reads your cloud credentials, and hides what it did from you. Counts: 98,698 scanned, 37 blocked. On screen: "Some skills are malicious" and "Every skill is scanned". 10. The refund skill is forked into the team's own copy. Its grace period changes from 14 to 30 days, it is saved as version 2, and it is shared with the Support team: Mei, Kwame, Priya, Daniel and Sofia can use it. On screen: "Fork it, set your own policy" and "Share it with your team". 11. One support agent has 240 skills installed, and the DecimalAI Skills Router picks the right skills for each agent turn: "You charged me twice this month" goes to billing-dispute-resolution; "The app has been down all morning" goes to incident-status-comms; "Can I get a refund?" matches three skills, and the best tested, refund-request-handling, is picked. On screen: "One agent, 240 skills" and "Best match, best tested". Then every skill but the picked one dims, and two bars compare the agent's tokens per turn: with the Skills Router, 30% or more fewer than with no router, because it shows a short menu instead of every skill, loads only the relevant skills, and keeps the menu cached between turns. On screen: "30%+ fewer tokens". 12. Skills are built into LangChain, OpenAI Agents, the Claude Agent SDK, CrewAI, Google ADK and Pydantic AI, each with a few added lines of code, or into your own code on the plain OpenAI, Anthropic or Gemini SDK. On screen: "Built into the frameworks you use" and "Or your own code". 13. An example of the support agent's last 30 days: refund-request-handling was loaded in 612 of 1,240 runs, +21 points on this agent; billing-dispute-resolution in 418 runs, +12 points. A suggested edit, from 3 runs that went wrong: "If the same charge appears twice, refund the duplicate in full." It is accepted: version 3 created, benchmark passing. On screen: "Measured on your agent" and "Improved, re-tested". 14. End card: DecimalAI. "Add, build and measure your agent’s skills." Start free at decimal.ai.