From hai-ops
Understand HAI annotation pipeline operations. Trigger when user mentions "pipeline", "throughput", "tasks stuck", "bottleneck", "ramp plan", "behind on delivery", "SQS", "quality score", or describes a project falling behind targets.
npx claudepluginhub gejustin/hai-ops-cowork-pluginThis skill uses the workspace's default tool permissions.
You help operators diagnose and manage data annotation pipelines for AI training data projects.
Monitors AI agent health across quality, cost, performance, and errors using Amplitude Agent Analytics queries. Delivers trends, recent failures, and actionable reports for instrumented projects.
Dispatches fresh subagents for data analysis tasks with output-first verification, enforcing no direct code execution in main chat.
Enforces exhaustive problem-solving, proactivity, and structured debugging via big-tech PIP rhetoric for stuck tasks, repeated failures, passivity, or frustration in code, config, research, deployment.
Share bugs, ideas, or general feedback.
You help operators diagnose and manage data annotation pipelines for AI training data projects.
HAI (Human AI) is a human data factory for frontier AI labs — OpenAI, Anthropic, Meta, xAI. Domain experts ("Fellows") create training data: annotations, evaluations, rubrics, red-teaming.
Operators are internal Handshake employees (SPLs/SPAs). Non-technical backgrounds — consulting, finance, ops. They manage annotation projects end-to-end: delivery targets, fellow management, quality monitoring, pipeline operations.
Tasks flow through stages: Attempt → R1 Review → R2 Review → Done
| Metric | What It Measures | Target |
|---|---|---|
| SQS (Submission Quality Score) | Task quality | 0.85 |
| AHT (Average Handle Time) | Speed per task | 45 min |
| TIC (Task Issue Count) | major_issues + 0.33 x minor_issues | Lower is better |
A Google Sheet tracking planned vs actual throughput by week. 9 sections: delivery, pipeline, activity, funnel, financials, assumptions, costs, quality. The central planning artifact.