Open: 01_live-dashboards/audit-kit-human-feedback-workbench.html (self-contained)
Problem. Agents need fast, structured human judgment on batches of proposed items (images, rules, claims) — chat is slow, spreadsheets lose the visuals.
What I built. A zero-dependency single HTML file: the operator opens it, reviews each card with Yes / No / IDK + a feedback box, click-to-zoom on images, 44px+ tap targets, localStorage persistence, then exports a typed JSON payload (audit-kit-v1: section, id, decision, feedback) that the agent reads back.
Design rationale (from the repo). Every visual item must carry its image; "IDK + text" is a real answer; "No + text" is a specific rejection reason; locked decisions become read-only on later passes.
Why it matters. Designing the human feedback channel into an agent loop — deterministic schema in, deterministic schema out — is a core applied-AI workflow skill.
When an agent proposes a batch of things, a human still has to judge them, and chat is too slow while spreadsheets lose the images. So this is one self-contained HTML file: the operator goes card by card, Yes, No, or I-don't-know, with a feedback box and click-to-zoom, then exports one typed JSON payload the agent reads back. Note the schema, IDK-plus-text is a real signal, and No-plus-text is a specific rejection reason. I'm designing the human feedback channel of an agent loop as a clean contract, structured in, structured out.
GG · Applied AI Engineering portfolio · synced to youdontneedmy.help design