simmer-generator_skill

Overview

This skill generates a focused, improved version of an artifact based on a single, high-leverage piece of ASI feedback from the judge. It is intended to run as a generator subagent inside the simmer orchestrator and produces a complete candidate file for the next evaluation round. The goal is targeted, conservative edits that preserve what already works while fixing the judge's most important issue.

How this skill works

The skill receives the current candidate text, the rubric describing what "better" means, the ASI (the single most important improvement), and the iteration number. For the first iteration it seeds a complete initial candidate that meets the rubric; for later iterations it applies a focused edit addressing the ASI while preserving other functioning parts. It outputs a full artifact file ready for the next judge pass and a short 2–3 sentence change report explaining what changed and why.

When to use it

• As a subagent when the orchestrator needs a next-round candidate based on judge feedback.
• When the judge provides a single most-important-fix (ASI) for a previously generated artifact.
• For iterative content improvement where targeted edits are preferred over rewrites.
• When the artifact must remain within an existing scope or format (e.g., API spec, email, tagline).
• On the first iteration to create a solid initial candidate from the rubric and description.

Best practices

• Always read and prioritize the ASI; fix only that primary issue unless tiny harmless improvements are available.
• Preserve any content not implicated by the ASI to avoid regressions in unchanged dimensions.
• Respect the artifact's scope and format; do not expand beyond intended length or structure unless the rubric requires it.
• Produce the full updated artifact file, not a diff or step-by-step instructions.
• Include a concise 2–3 sentence change report describing what was changed and why.

Example use cases

• Improve an API spec's error response section after the judge flags missing error codes.
• Tighten an email CTA after the judge marks it as high-friction, without rewriting the message body.
• Adjust a landing page headline when the judge identifies clarity as the main issue.
• Generate the initial draft artifact on iteration 1 using the rubric as the spec.
• Fix a function comment or example that the judge says is incorrect while keeping other examples intact.