How does the anchor system prevent semantic drift?

It uses three layers: Immutable anchors (verbatim phrases), Semantic anchors (preserved meanings), and Structural anchors (organization patterns). These must be verified after every edit.

What happens if an iteration breaks an anchor?

If an immutable anchor is lost or violated, the skill will halt and ask the user to either revert the change, update the anchor, or stop the iteration process.

When should I use the IAQI skill instead of a normal edit?

Use IAQI for complex tasks needing measurable quality improvement, such as refining prompts, agents, or skills. Avoid it for simple fixes or one-shot edits where a review loop is overkill.

What does IAQI stand for?

IAQI stands for Intent-Anchored Quality Iteration, a method of improving artifacts through repeated reviews while 'anchoring' the original meaning to prevent the AI from changing the intent.

Can I customize the scoring criteria?

Yes, during Phase 3 (Scoring System), you can customize dimensions, set specific target scores, and designate non-negotiable requirements that the artifact must meet.

Intent-Anchored Quality Iteration (IAQI)

Name: Intent-Anchored Quality Iteration (IAQI)
Author: bengous

bybengous

•

생산성 및 워크플로

Refines complex artifacts like prompts and code through parallel reviewer loops while preventing semantic drift using a multi-layered anchor system.

소개

IAQI is a specialized Claude Code skill designed to iteratively improve high-stakes artifacts—such as prompts, AI agent definitions, and technical documentation—without losing their original intent. It employs a rigorous four-phase workflow (Understand, Anchor, Score, Iterate) that uses multiple parallel AI reviewers to evaluate outputs against custom rubrics. The core innovation is the 'anchor' system, which utilizes immutable phrases and semantic checksums to ensure that as an artifact's quality improves, its fundamental purpose and identity remain intact, effectively preventing the common drift often caused by iterative AI editing.

주요 기능

Parallel Reviewer System: Spawns multiple agents simultaneously to provide diverse feedback and objective scoring.
Autonomous Iteration Loop: Automatically applies prioritized fixes and re-evaluates until target quality thresholds are reached.
Multi-Layered Anchoring: Protects core identity using immutable, semantic, and structural checksums to prevent drift.
Comprehensive State Management: Records every iteration, score, and anchor check in a localized state file for full auditability.
Custom Scoring Rubrics: Automatically generates and customizes evaluation dimensions based on the specific artifact type.
2 GitHub stars

사용 사례

Refining custom Claude Code skills or commands to ensure they follow best practices and maintain technical accuracy.
Improving technical documentation while strictly preserving specific terminology and organizational structures.
Optimizing complex AI system prompts to reach specific performance targets like a 9.0+ quality score.

소개

주요 기능

Parallel Reviewer System: Spawns multiple agents simultaneously to provide diverse feedback and objective scoring.
Autonomous Iteration Loop: Automatically applies prioritized fixes and re-evaluates until target quality thresholds are reached.
Multi-Layered Anchoring: Protects core identity using immutable, semantic, and structural checksums to prevent drift.
Comprehensive State Management: Records every iteration, score, and anchor check in a localized state file for full auditability.
Custom Scoring Rubrics: Automatically generates and customizes evaluation dimensions based on the specific artifact type.
2 GitHub stars

사용 사례

Refining custom Claude Code skills or commands to ensure they follow best practices and maintain technical accuracy.
Improving technical documentation while strictly preserving specific terminology and organizational structures.
Optimizing complex AI system prompts to reach specific performance targets like a 9.0+ quality score.