local-llmlisted

# Local LLM ## Purpose Cheap, fast transformation work that doesn't need Claude reasoning. Summarization, classification, extraction, first-pass sanity check. The skill invokes a local Ollama model with a templated prompt and writes the model's response (plus sidecar metadata) to `work/current/artifacts/` for Claude or a human to review. This skill is the *worker* boundary; Claude (or you) is the *judge*. Local model output is untrusted; the artifact is a draft. ## Scope ### When to use - Bulk operations on files (compress 80 changelog entries → 7 bullets). - Low-stakes preprocessing (classify support tickets into bug/feature/ question/noise; extract dates and names from prose). - Synthetic data for tests (generate sample fixtures from a schema). - First-pass triage before a Claude review. ### When NOT to use - Anything safety-critical or correctness-critical. - Anything requiring multi-step reasoning. - Anything where local-model failure would be hard to detect. - Final code review, security analysis, deploy gates — those need Claude. The local model's role here is volume, not judgment. ### The pattern The skill is the boundary. Local model writes to `work/current/artifacts/`. Claude reads the artifact and makes the actual decision. Local model output never directly modifies project source. ## Automated pass The skill invokes `scripts/ollama-prompt.sh` with the template, input, and output. The script: 1. Validates that `ollama` is on `$PATH`. If absent, exit