Check whether it can complete a multi-step task, ask smart follow-up questions, use tools, recover from errors, show its work, remember context, and behave safely with sensitive tasks.

Check whether it understands the existing codebase, fixes bugs, explains changes, writes tests, avoids unrelated files, and gives output a non-technical user can follow.

For video, check coherence, faces, hands, motion, camera instructions, character preservation, commercial usefulness, generation time, and usage limits. For image, check sharpness, style following, text handling, identity/product preservation, variations, editing precision, and artifacts.

For research, check citations, real sources, date handling, source comparison, accuracy, and fact versus interpretation. For models, check modalities, reasoning, coding, context, API access, pricing, geography, agents, creator fit, enterprise fit, and open/closed status.

Check local runtime, hardware needs, license, commercial permission, closed-model comparison, quantizations, simple runtimes like Ollama or LM Studio, and whether it is good enough for the task.