Name the four layers
Record the model label, the product you use, the connected tools, and the final artifact. A chat product may expose text, image and voice features through different systems. A coding agent can write a script that another application runs. “Made with Claude” does not tell you whether the result came from text generation, Blender execution, imported assets or human edits. Keep unknown fields unknown; do not infer a precise model version from the date of a post.
Compare capabilities as questions you can test
For text: does the result preserve the intended meaning and cite inspectable evidence? For images: can it revise one requested element while preserving the others, and can you export the dimensions you need? For voice: are you testing transcription, spoken output or a live interruption-and-response loop? For tools: can the agent reach the necessary project and stop for sensitive actions? Use your own task-specific questions rather than putting all four into one unearned score.
A 3D clip makes the distinction concrete
In the Blender MCP source project, an assistant communicates with Blender through an integration. A rendered scene therefore demonstrates a tool-assisted route, not an isolated text model’s ability to generate every pixel. The useful follow-up is to inspect the native scene, modify one object and reopen the export. If only the clip is available, it remains an author demonstration. This publication has not reproduced every scene or established a controlled model winner.
Supporting documentation (opens in a new tab)Keep disagreement without averaging it away
One author may praise surprising ideas while another complains about cautious refusals. Before calling these conflicting measurements, compare the task, prompt, plan, settings and success criterion. A designer evaluating visual initiative and a developer evaluating predictable edits may both describe real experiences. Save both observations with their context. Do not average subjective praise into a precision-looking numerical rating, and do not present comments on different tasks as a head-to-head experiment.
Make a small evidence card
For the next demo you save, fill in: required output; author and original link; product and model if disclosed; tool dependencies; visible result; what is missing; one change request worth trying. Choose a conclusion: inspiration, worth a bounded test, or verified for my particular use. The final label requires your own checks. This card helps you choose the next action without pretending to know every capability of a model or every feature of a subscription.
Put this method beside a concrete example.
Inspect the related case →Our evaluation framework, not a controlled test or a guarantee of results.
Explore the free Jev collection →