The black box
Each dot is one AI session. The dot size shows how many messages it had. Dots group by topic. The moving sparks show the same week jumping from one topic to another. Type in the search box, or click a topic, to zoom in.
Where the line splits
Some prompts belong to two topics at once. These counts come from a first keyword pass over all prompts. A local model will replace the keywords before this page is final.
Care and evidence
prompts. Care problems turned into records: recordings, custody logs, transcripts.
Care and the law
prompts asked what the rules say: statutes, ADA, appeal rights.
Faith and conspiracy
prompts. This is where the search for meaning turned into theories. It is a small number, and it is shown on purpose.
Calling out the tools
The caregiver checks the AI. When the AI is wrong, the error gets a name and a number: a CVE-FAUX record.
Call-outs
prompts corrected the AI ("that's not right", "you said…"), in sessions.
Glitches
AI replies had stray Chinese characters in English text.
Caught mid-thought
The CVE wall
The missing answers
Tools and cost
What the agents used. MCP tools connect the AI to the family's own data. The local ones never leave the house.
Tool calls
calls.
MCP calls:
Tokens
The workbench
What the agents did with their hands: the commands they ran, the folders they worked in, and the files they made or changed.
Most used commands
Most used folders
Files
Topics by prompt count
Where the sessions came from
Build the tools · What tools
The public repos that came out of the sessions. The goal behind all of them: prepare for the care when we cannot be there.