The advisor: a second opinion
A stronger model that reads the whole conversation and reviews the plan before it is carried out, and again before the work is declared done.
The idea
The advisor is a second, stronger model that the working model can consult. It automatically sees the whole conversation, every command that was run and every result, and answers with advice: “you are about to build on an assumption nobody checked”, “you forgot this case”.
How I have it set up
It is a single line in my settings, which names the model to use, and a command to change it at any time (or to switch it off). I keep it on the most capable model available.
When it is worth it
- Before the work starts, once the first orientation is done (files read, state understood) and before an approach is chosen: that is when advice is cheapest and changes the most.
- When stuck, or when results do not fit what was expected.
- Before declaring done, after the result is saved or committed, so a mistake is caught while it is still cheap.
For small, obvious tasks it adds nothing, and it is not called.
What it does not replace
It does not know my house. Most of what went wrong in my sessions came from facts about my own devices that nobody had written down, and a second model reading the same conversation does not have them either. That is the job of Memory and of The rules I set myself. The advisor is good at reasoning, not at remembering what is plugged into what.