Delegate to local Ollama

Pattern: the local model types, you spec and review. From a tight spec, a mid-size local model (30B-class) produces ~90% production-quality code at zero token cost — but it misses subtle bugs in its own and others' code, so every output gets reviewed and tested by you before it ships.

How to run

Use the bundled runner (relative to this skill's directory), never ollama run — the CLI emits TTY escape codes into its output even when stdout is redirected:

<this-skill-dir>/scripts/ollama-chat.sh <model> <prompt-file> <out-file> [num_ctx=16384] [keep_alive=2h]