Most automation projects stall after the first good demo. Someone showed an AI what to do, or recorded a browser and desktop sequence — then the result stayed locked to that chat, that model, or that machine. Automation Skill Builder (ASB) is built for the next step: teach the work however is easiest, then leave behind a local asset you can inspect, move, and run again.
Design and runtime are different jobs. AI is useful while you are figuring out the flow. Once the steps are clear, they do not need a model in the loop on every rerun. ASB keeps those steps as readable code and can package them for repeatable local execution.
How to read the tags on the map
- AI guided — an AI agent instructs or drives ASB.
- Human guided — you operate the UI while ASB (or an external recorder) captures the steps.
- Real-time run — actions execute during capture, so you can verify as you go (often with Run once).
- Delayed run — write the script now; run it later on another machine.
- Record driven — the script comes from a live recording or an imported log.
- AI authoring — the AI writes the script from approved skills and a scenario, without a live record session.
- Dashed card (G4 UIA) — optional escape hatch when built-in desktop capture is not enough. It is dashed on purpose: not the main path, only a fallback for hard Windows UI cases.
01What should remain after the design session?
Many tools start with “which recorder or agent will you use?” ASB starts one question later: after that helper is done, what do you still own? Ideally more than a chat transcript — an editable Python script, a scheduled local app, or a skill other agents can discover and call.
That is why the three bands stay separate. Connect is who can reach ASB. Capture / Author is how work becomes executable steps. Host & Deliver is what you keep after design. They are choices around one local control plane, not a forced three-screen wizard.
02Three common ways to start
Let an AI agent do it once
An MCP-capable agent can drive ASB while you describe the outcome. ASB records the calls from that session and turns them into a replayable script. This fits flows that are easier to explain than to click through by hand — files, OCR, desktop actions, and APIs in one pass. The AI helps discover the path; the recorded path is what you keep.
Show the work yourself
For browser and desktop work, a person is often still the clearest teacher. Playwright captures the web; ASB’s built-in actions cover supported desktop apps. G4 UIA is the dashed box on the map: an optional import path when ordinary desktop capture falls short. Prefer the solid paths first; use G4 only as a Windows fallback, then bring the result into the same script-and-package pipeline.
Author here, run somewhere else
Sometimes the design machine cannot reach the target app, or production must stay offline. An agent can write from an approved skill set and scenario without running the target process locally. You review the script, copy it to the destination, and run it there — design and execution stay separate without hiding the work in a cloud black box.
03Pick the deliverable for the next owner
When the flow is stable, choose the form that matches who runs it next:
- Recorded script — when someone needs to read, edit, test, or adapt the logic.
- Executable / app — when it should launch or schedule like ordinary local software, with no AI call per run.
- Skill package — when agents should keep discovering and reusing it through MCP or FastSkills.
One automation can wear more than one form: review as a script, ship as an executable for ops, and publish as a skill for agents. Deliver is that ownership choice — not just an export button at the end of a recorder.
04What “zero Cloud/AI calls at runtime” means
It does not ban network calls your business already needs. Your workflow can still talk to ERP, APIs, or internal services. The promise is narrower: ASB does not need a model call to decide the next already-recorded step on every run. Use AI to create or repair the procedure; keep everyday execution local and deterministic.
That split cuts cost and latency, keeps sensitive data closer to the host, and makes failures easier to debug — operators read the script and the log instead of reconstructing a past chat.
Start from the deliverable you need to own. Then pick the lightest capture or authoring path that can produce it on the systems you have. Solid cards are the main roads; the dashed G4 card is the exit ramp when those roads are blocked.
Keep the capability after one good run
Connect an agent or demonstrate the task, review the script, then package it for the people or systems that will run it next.
Download for Windowsvisualbuild.meMCP everywhere