Temperature is the most visible knob in every AI tool and the least deliberately used. Most people pick one value per app and never think again. Then they wonder why their classifier hallucinates and their brainstormer is boring.
Temperature reshapes the probability distribution the model samples from. Low values sharpen it — the model almost always picks its top choice. High values flatten it — unlikely words become reachable. Neither setting is "better"; they trade consistency for variety.
| Task | Temp | Why |
|---|---|---|
| Classification / triage / routing | 0.2–0.4 | You want the same input to always yield the same label |
| Data extraction, JSON generation | 0.3–0.5 | Structure integrity beats creativity; combine with JSON mode |
| SEO outlines, technical docs | 0.5–0.6 | Structured but not frozen |
| Social posts, hooks, titles | 0.8–1.0 | Variety IS the value; you curate from many options |
| Localization | 0.6 | Natural phrasing without drift from meaning |
The meta-skill is matching the knob to the failure mode you fear: fear of randomness → lower it; fear of blandness → raise it. Every workflow in my packs ships with these values already tiered per node — generation nodes hot, analysis nodes cold.