Run a small model locally for one task. Not ten. That is the whole trick people miss when they jump on this trend. You load up Ollama or LM Studio, start testing fifty things, and end up with a dozen unfinished scripts that each kind of work but none reliably.
I keep two local models running. A 7B parameter model handles code snippets and regex patterns. Another smaller model reformats raw text into structured notes from my voice recordings. Both are single-purpose. Both stay loaded on a machine that does nothing else.
The automation part is where people overthink. You do not need a fancy agent framework. A cron job that pipes text into stdin and captures stdout is enough. I have a script that watches a folder for new meeting transcripts. When one lands, the local model strips timestamps, removes filler words, and writes a clean summary to another folder. Takes fifteen seconds per hour of recording.
Local models shine when you control every variable. No rate limits. No unexpected API changes. No data leaving your network. But they fail hard at complex reasoning or creative writing. Do not ask a 7B model to draft a contract or write poetry. It will embarrass you.
Pick one repetitive text task you do daily. Find a model that fits in half your available VRAM. Wire it up with a simple file watcher or scheduler. Test it for a week. If it works, keep it. If not, delete everything and try a different model or task. Do not hoard model files.