Meet the personality of your Claude.

Claude in a browser, Claude Code and another client using a Claude model are not identical environments. The model, tools, instructions and available context can all influence the answers. aivitals records these labels separately rather than treating every Claude result as the same thing.

Choose context before choosing a mode

For “you, in its eyes,” use a conversation with useful history or relevant memory the current client can access. If project context contains only code conventions, that is not automatically evidence of your entire personality. The test explicitly asks the AI not to treat the copied questionnaire as evidence about you.

Claude's chat search and memory are subject to product settings and context boundaries. A coding agent may have project instructions without having access to your Claude web history. Do not grant extra file or account access merely to complete an entertainment test.

Browser chat: use the JSON path when needed

1. Open aivitals. Auto submission is the default; choose Manual when you want a local report or your client cannot send results.
2. Copy the prompt into your chosen Claude conversation. It asks for twelve A/B answers plus an evidence-based user impression.
3. For Manual mode, copy the entire JSON reply back into aivitals and select View report. If Claude includes prose around the code block, copy the complete block rather than editing individual fields.

Tool-enabled clients: verify the handoff

The automatic prompt describes a JSON POST for environments that can run a network command, and a web-fetch fallback. Whether those actions are available or require approval depends on the client. Review the proposed command and data before granting permission. aivitals never needs your API key.

A submission should be confirmed by a successful response from the service and a visible archive entry. If permission or network access fails, use Manual mode; do not repeatedly grant broader permissions. Manual results are still usable without a public submission.

Compare runs without inventing a winner

Record the model, client and thinking label honestly; use unknown when uncertain. A different type on a later run is not proof that a model became better or worse. Compare the particular answers and context instead. This questionnaire is an entertainment framework, not a scientific model benchmark.