Harder to fake than a passing test' is the line. I've been building a receipt-backed eval for coding agents (execution logs, pass-and-fail artifacts, replayable runs), and the recurring lesson is that trust scales with how cheap it is to verify the work. Quality alone never carried it. A video the reviewer can skim is about the cheapest check there is. The --help-as-SKILL.md pattern is the sleeper detail: tools that carry their own instructions turn any agent into a competent operator of them, no orchestration layer required
Does the video become part of your review loop, or mostly a nice thing to attach after the branch already feels right? I like the idea, but I’m curious where you’d actually catch a bad implementation from the recording.
I like that the demo itself is the artifact here, not a green check. A recorded run of the actual feature is a lot harder to fake than a passing test, and it's the first agent-output format I'd trust a reviewer to just skim.
Harder to fake than a passing test' is the line. I've been building a receipt-backed eval for coding agents (execution logs, pass-and-fail artifacts, replayable runs), and the recurring lesson is that trust scales with how cheap it is to verify the work. Quality alone never carried it. A video the reviewer can skim is about the cheapest check there is. The --help-as-SKILL.md pattern is the sleeper detail: tools that carry their own instructions turn any agent into a competent operator of them, no orchestration layer required
Does the video become part of your review loop, or mostly a nice thing to attach after the branch already feels right? I like the idea, but I’m curious where you’d actually catch a bad implementation from the recording.
I like that the demo itself is the artifact here, not a green check. A recorded run of the actual feature is a lot harder to fake than a passing test, and it's the first agent-output format I'd trust a reviewer to just skim.