temper-skillslisted

# Temper-Skills (library mode) Turn the decision logic of an agent skill or prompt into deterministic Python — zero LLM calls at inference — by **driving the installed `temper-skills` CLI**. You orchestrate the CLI and relay its output; the library does the tempering. You never write the decision tree yourself. > Use this skill when you do **not** have a native subagent/Task primitive to run the loop > yourself. Inside Claude Code, prefer the subagent-mode skill (it runs keyless, with a live > panel). This mode shells out to the CLI instead. ## When to use - A `.md` skill, system prompt, or playbook holds routing/decision logic (if-this-then-that over structured inputs) that should be testable, diffable in a PR, or resilient to model updates. - Good fit: ticket routing, classification rules, eligibility checks — categorical decisions over pre-computable features. - **Not** a fit: continuous scoring, text generation, or logic with no documentation the model can reach. ## Non-negotiable rules 1. **Always call the library through its CLI. Never reimplement the decision logic in natural language** as a substitute for running it. 2. **Never skip schema ratification.** The schema you approve is the *naive seed* the loop then co-evolves (the `schema_critic` grows it, the earn-a-branch guard prunes it), so a silent extraction error in the seed still poisons everything. Use the file-based gate below (`--propose-schema`), even when no terminal is attached. Rev