Iteratively improve artifacts via parallel reviewers with anchor-based drift prevention. NOT for: one-shot edits, simple fixes, tasks without measurable quality. USE for: skills, prompts, commands, agents needing iterative refinement with score targets. Triggers: "iterate to 9.0", "quality loop", "IAQI", "reviewer loop", "improve until threshold", "iterative quality".
View on GitHubbengous/claude-code-plugins
software-craft
January 24, 2026
Select agents to install to:
npx add-skill https://github.com/bengous/claude-code-plugins/blob/main/software-craft/skills/iaqi/SKILL.md -a claude-code --skill iaqiInstallation paths:
.claude/skills/iaqi/Intent-Anchored Quality Iteration (IAQI) prevents semantic drift during iterative improvement. The problem: each iteration risks optimizing for reviewer preferences rather than original intent. IAQI solves this through anchors—semantic checksums verified after every change. <iaqi_philosophy> ## Philosophy: The Anchor System Before iterating, define anchors at three levels: | Layer | Purpose | Can Change? | Example | |-------|---------|-------------|---------| | **Immutable** | Core identity phrases | Never | "Skills encode judgment. References encode facts." | | **Semantic** | Meaning to preserve | Wording only | "Force intentional choices" (can rephrase, not remove) | | **Structural** | Organization patterns | Flexible | 7-section template (order can change, sections preserved) | **Why anchors matter**: Without them, iteration 5 optimizes for reviewer 5's preferences, not iteration 0's intent. Anchors are checksums against drift. A good anchor set has: - 3-5 immutable phrases (exact strings that must appear verbatim) - 4-6 semantic anchors (meanings that must be preserved) - 3-5 structural anchors (patterns/organization to maintain) - A list of forbidden changes (what would violate the intent) </iaqi_philosophy> <iaqi_workflow> ## Workflow: 4-Phase Structure Parse arguments from invocation: - `artifact-path` (required): file to improve - `--reviewers N` (default: 5): parallel reviewer count - `--target X.X` (default: 9.0): score threshold for success - `--max-iterations N` (default: 7): iteration limit ### Phase 1: UNDERSTAND (human required) 1. Read the artifact thoroughly 2. Propose an intent statement (1-2 sentences: why this artifact exists, its distinctive quality) 3. Create a workflow diagram showing the artifact's structure or flow: - Skills: phase/step workflow - Commands: decision tree or execution flow - Agents: interaction pattern - Prompts: section structure 4. Present to human via AskUserQuestion: - Show intent statement and dia