Summary

A step-by-step build: installing Claude Code and UiPath's CLI, connecting them, generating a support-ticket classifier agent from a single plain-English prompt, testing it, and publishing it to UiPath's cloud. The presenter states at the outset that the video is made in partnership with UiPath, and it closes on a sign-up pitch for the free community plan, so it should be read as sponsored vendor documentation rather than independent assessment — no competing platform is considered and no limitation of the approach is discussed. Its value here is not the verdict, which is foregone, but the mechanism it documents: how a vendor packages its own conventions so a general coding agent can build correctly for a platform it was never trained on. Note the presenter introduces himself as Garrick although the channel is Kevin Stratvert's.

• The mechanism is the substance of this entry. UiPath ships a set of agent skills — instruction files installed into the coding agent — that teach Claude Code the platform's patterns, project structures and conventions before it writes anything. The presenter's analogy is 'giving Claude Code a specialized manual before it starts building'.
• The consequence claimed is that output is not generic code but a project following the vendor's specific structure, using their SDK, in a format their platform can run, evaluate and deploy. The skill loads automatically when the agent is opened inside a project of that type, and the presenter points on screen to the line showing it firing.
• The same approach is claimed across the vendor's other product areas, with the relevant skill loading per project type: UI automation, reusable API workflows, application testing, document processing and orchestration. Only the agent case is demonstrated; the rest is asserted.
• The stated problem it addresses is a real one for enterprise tooling: building on such a platform has traditionally meant learning its visual editors and whole ecosystem, which is a barrier for a developer who needs to ship one automation and return to other work.
• Two-way editing is claimed — work can be moved into the vendor's visual editor and continued there at any point. Not demonstrated in the video.
• The agent itself was produced from one plain-English prompt describing inputs and the categories and priority levels wanted. The presenter reports the whole build taking a couple of minutes, including the agent registering itself in the project configuration and running its own test unprompted.
• The self-correction shown is genuine and worth recording precisely because it was a failure first: the generated agent used a field named 'message' while the test used 'body', which threw an error. The agent identified the mismatch, fixed it, re-ran the test, and then asked the user to confirm the rename. This is one observed instance, not a measured reliability rate.
• Two test results are shown. A ticket reading 'charged twice' with 'duplicate charge on my card' was classified billing, high priority. One reading 'cannot log in' with 'my password reset is not working' was classified account, medium priority, with reasoning given that the issue affects access but indicates no immediate business-critical impact. Two tickets is not an evaluation.
• The setup is not frictionless, and the video is unusually honest about this by leaving the failures in. A Python version conflict surfaced because the project initialised expecting 3.9 while the required package needed 3.11, fixed by pinning the version and editing the project file. The first publish attempt failed outright with 'project authors cannot be empty', fixed by adding an author line. Neither is presented as a problem, but both are real friction in a demonstration otherwise built to look effortless.
• Platform-specific constraints reported: on Windows the workflow requires Git Bash for most steps, with PowerShell and Command Prompt explicitly stated not to work, while the vendor's own CLI installer must be run in PowerShell rather than Git Bash. Git for Windows is a prerequisite.
• Stated prerequisites: a UiPath account on the free community plan, Claude Code with either a Claude Pro or Max subscription or an Anthropic API account, Node.js 22 or higher with 24 recommended, and Python 3.11 or higher with the uv package manager.
• Deployment is described but its value is asserted rather than shown: the published agent gets a version number in the vendor's web studio and can then be scheduled, triggered by API, or dropped into a larger workflow. The presenter describes connecting it to a ticketing system such as Zendesk or Freshdesk as what would happen 'in a real deployment' — so the demonstrated agent is not connected to anything.

Why it matters

This is a sponsored tutorial and its conclusions carry no weight, but the pattern it documents does, and it is the reason to keep the entry. An established enterprise vendor is not building its own AI assistant here — it is shipping instruction files that make somebody else's coding agent competent on its platform. That is a different commercial posture from the one usually reported: the platform accepts that the user will arrive with a general-purpose agent already in hand, and competes by making itself the easiest thing that agent can target. If that becomes the norm, the interesting question shifts from which vendor has the best assistant to whose conventions the agents know, which is a distribution question rather than a model-quality one. Against that, everything specific here is single-run vendor demonstration: two test tickets, one observed self-correction, no error-rate figures, no comparison, and a verdict decided before filming. The version numbers and install commands will date within months; the packaging strategy is the part worth watching.

TSLQyhYUSLE-transcript.txt