A workstation assistant that points, never types

Level IntermediateEstimated time 1 hRead first Which local model for which job: a bench that measures real work, A RAG on your own journal: never solve the same problem twiceA word you don't know? Glossary

"Turn the volume down." "What state is the server in?" "What did we decide about the GPU during games?" A small local assistant can handle these, as long as it is never allowed to type a command. Ours receives a closed table of functions, points at one, and the script does the rest.

What you get. An assistant on your workstation that acts in under a second when the server is up, still answers (slower, never mute) when it sleeps, and says "I don't know" instead of inventing.

1. The model points, the code acts

Measured: the model still got one logic question in seven wrong. So it never generates a command. It returns one call from a closed table; the arguments are read as Python literals, never evaluated; anything outside the table becomes NONE rather than a near miss.

set_volume(percent: int)
machine_state(host: str)
search_memory(query: str)
escalate(question: str)

Declare the table as Python signatures with examples. Declared as JSON Schema, the model picked the right function but did not fill the values; as Python it returned 6 exact calls out of 7.

2. Sources before conversation

Asked about "the GPU mode", a model lectured confidently on Docker's network modes: plausible, entirely wrong. So a question is first answered from your notes and your infrastructure journal, with the model allowed to use only those excerpts. Each model in the cascade must answer exactly INSUFFICIENT when it does not know, and the code checks that word, letter for letter, before passing on.

Keep that marker in the code, not in the editable persona file: one edit to the persona, and the safety disappears.

3. Fast when it can, never mute

The server's general model designates first, at about 140 tokens per second; a small function-calling specialist on the laptop's processor takes over, at 5.5, when the server sleeps. The specialist is good on explicit orders; on an indirect question it gave up where the general model found the right search.

4. Escalate with the case file

The last step is a frontier model. Our first escalation answered from the sources alone, unable to look at the machine; it now receives a snapshot of the machine's state and the journal excerpts, and the screen says so, because it costs a remote call.

Found this useful? Tell us what to measure next.