Handbook
Wizard 301 — Refine
Steps like Understanding may offer Refine: LLM-assisted iteration while you stay in control.
What it is
Steps like Understanding may offer Refine: LLM-assisted iteration while you stay in control.
Parent: Wizard 301 — Advanced usage.
How to use it
Read output, edit notes, refine again. Paste constraints and non-goals explicitly.
Tie-in: In the worked example session on Wizard 301 — Advanced usage, Refine almost drops an SLO until the team edits notes and runs Refine again — same pattern you should use when the model smooths away facts you need.
Step-by-step usage (typical)
- When Understanding (or a similar step) offers Refine, run it on a short paragraph you already agree is directionally right — not on empty text.
- Read the model output; edit the notes field with facts the team insists on (dates, names, boundaries).
- Refine again only when the delta is worth the cycle; otherwise move forward and fix detail in Review & generate.
- If Refine errors appear, see Troubleshooting — often policy or local API loopback.
UI affordances (plain language)
- Refine is assistive: you remain the editor of record; nothing ships without your Review step.
- Constraints and non-goals belong in your notes so Refine does not “invent” organization policy.
Verify
After Refine, the updated text still reflects facts the team named (dates, owners, boundaries). If the model dropped a constraint you need, put it back in notes and run Refine again or fix it in Review & generate.
Recheck loop
After Refine-heavy steps, expect Recheck / repair to matter more — see Review and recheck.