Concise observations, transparent breakdowns, and material updates from the work—without manufacturing a daily content treadmill.
Evidence-conscious notes
Published Field Notes
Complete answers, not a content treadmill.
New articles appear here only after their evidence, claims, sources, and practical value have been reviewed.
AI and Workflow Automation
How to Use Computer Use in AI Workflows: A Control Check
A click counts only when your application checked the action and you can see the screen it changed. A model request is not permission. A closing sentence is not the result.
A mismatch between the outcome a reader expected and what the page did. A chosen installer can stay. A virus, a silent homepage switch, or a trapped back button is the pattern to retire.
How to Use Background Mode in AI Workflows: A Status Guide
A background result counts only when a later retrieve shows a terminal status and the output. A queued create is not that result. A webhook id is not the output.
Spam added through a channel you opened for readers. A real comment can stay. A promo post, a spam account, or an unmonitored upload is the pattern to retire.
What Is Misleading Functionality? A Claimed-Service Test
A page that claims a service it does not provide. A working calculator can stay. A fake credit generator, or a merge tool that only opens ads, is the pattern to retire.
Responsible AI systems for repeatable work, with human review retained where accuracy, trust, and risk matter.
Duff signs his name to the work.
First-person claims, timelines, screenshots, and results appear only when the supporting evidence is available. Research and interpretation are labeled plainly.