simmer-reflect_skill

Overview

This skill records iteration results for the simmer orchestrator, maintains a running trajectory table, and forwards the judge's ASI unchanged to the next generator. It identifies and tracks the best candidate across rounds and flags regressions when an iteration performs worse than the best-so-far. Do not invoke this subskill directly — it runs automatically after each judge round.

How this skill works

After each judge round, the skill appends the iteration's criterion scores, composite, and a short key-change summary to the trajectory table. It compares the current composite to the best-so-far, updates the Best candidate line if improved, and detects regressions when the current composite is lower. Finally, it returns the judge's ASI untouched, indicates which iteration file is the best input for the next generator, and reports whether iterations remain.

When to use it

• Automatically after every judge round in the simmer pipeline
• When you need a single authoritative record of score history and changes
• When the orchestrator must know which candidate to feed to the next generator
• When you must detect and handle regressions between iterations

Best practices

• Write the trajectory table with only the required columns and no candidate content
• Condense the generator report into a short Key Change (under 60 characters)
• Never modify or summarize the judge's ASI — pass it forward unchanged
• Always compare composite scores to the current best, not to the last iteration
• If regression occurs, mark REGRESSION and recommend using the best-so-far as input

Example use cases

• Recording scores and changes across iterative refinement of a prompt or model output
• Choosing the correct candidate to supply to the next generator after a poor iteration
• Maintaining a compact, machine- and human-readable trajectory.md for audits
• Signaling the orchestrator that iterations remain and which file to load next