Notes
The parts I would tell someone else
Running models and agents on a laptop is mostly bookkeeping: how much memory is free, how many tokens are left, which of the five things that can fail actually failed. These are the rules I now follow, collected while building the projects on the work page.
Showing all 8 notes.
-
Check free memory before you download, not after
A model that needs 20 GB of memory will not run on a card with 8 GB, and no amount of patience changes that. I read the free-memory figure first and pick the model to fit it — the reverse order cost me an afternoon and 14 GB of downloads.
-
An empty answer is usually a budget problem
When a model returns nothing, the first suspect is the token cap, not the network. A reasoning model spends part of that budget thinking, so a cap of 20 can produce a perfect request and an empty reply. Raise the cap and read the usage numbers before you touch the prompt.
-
Retry once before you debug
A single 500 from a hosted API is not a bug in my code. I now run the same request again; if it passes, I write the first failure in the log and move on. Debugging a transient error teaches you nothing and costs an hour.
-
One clear tool beats five vague ones
An agent picks tools by name and description alone. When two tools sound alike, it picks the wrong one and then explains the choice confidently. Fewer tools with sharper descriptions produced better behaviour than any prompt I tried.
-
A screenshot is evidence; a description is an opinion
For anything visual — a layout, a chart, a page that "looks wrong" — I send the image instead of describing it. Descriptions lose exactly the detail that matters, and they lose it silently.
-
For diagrams, ask for code, not a picture
A generated image of a diagram cannot be corrected, only regenerated. Hand written SVG can be edited line by line, stays sharp at any size, and weighs a few kilobytes. Every figure on this site is SVG for that reason.
-
Write down what broke, on the day it breaks
Six weeks later the only detail I remember is that it worked in the end. One line per failure — what I expected, what happened — is enough to recognise the same problem when it returns, which it always does.
-
Local models buy privacy, not quality
A small local model is worse at reasoning than a hosted one and better at keeping data on the machine. That trade is worth it for personal notes and not worth it for real work, and it is a decision to make per task rather than once.