What is an optimization target in this context?

An optimization target is a collection of metrics, dimensions, and signals that define the desired output quality, mapped to a 0-1 scale for objective evaluation.

What inputs are required to start an optimization target?

The skill requires a valid evaluation contract snapshot (typically from an analyzer tool) and a clear goal statement for the feature being optimized.

What is a 'lever' in an agentic loop?

Levers are specific components—such as prompts, graders, or configuration files—that the agent is authorized to modify to achieve the optimization goal.

Can I prevent the AI from changing specific files?

Yes, the skill supports 'frozen' paths and hard constraints that define immutable boundaries the agent cannot cross during an optimization run.

How are optimization settings persisted?

All targets, constraints, and allowed levers are stored directly in a journal.yaml file within the .claude/optimization-loops/ directory.

Agentic Optimization Target

Name: Agentic Optimization Target
Author: mberto10

bymberto10

0•

データサイエンスとML

Defines precise optimization goals and regression guardrails for autonomous agent loops using normalized 0-1 scoring.

The Agentic Optimization Target skill provides a structured framework for Claude to define exactly what needs to be improved and what must remain stable during autonomous optimization loops. By establishing a journal-embedded model of metrics, dimensions, and levers, it allows for data-driven iterations on prompts, configurations, and code. This skill ensures that every optimization attempt is governed by strict pass/fail criteria and explicit guardrails, preventing regressions while seeking peak performance based on evaluation signals.

主な機能

01Integration with evaluation contracts for automated, signal-driven improvement

02Strict 0-1 scoring normalization for objective performance measurement

030 GitHub stars

04Journal-based persistence in .claude/optimization-loops for historical tracking

05Granular lever management to control which files or configs can be tuned

06Explicit regression guardrails to maintain baseline performance standards

ユースケース

01Iteratively tuning complex prompts based on specific scoring dimensions

02Systematically improving code performance using Langfuse evaluation signals

03Optimizing application configurations while enforcing immutable safety boundaries

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add mberto10/mberto-compound optimization-target

For use in Claude.ai and ChatGPT

Download Skill