AI Dev Steward — AI skill for Claude Code

AI Dev Steward

**English** | [简体中文](README.zh-CN.md)

**Less clutter. Traceable experiments. Reviewable optimizations.**

AI Dev Steward is an Agent Skill for keeping AI-assisted algorithm development lean, reproducible, and reviewable. It helps consolidate redundant planning files, audit stale experimental artifacts, and organize complex optimization branches into changes that reviewers can trace to tests, measurements, ablations, and rollback steps.

The skill combines a development workflow with four read-only Python tools. It favors existing issues, pull requests, and experiment trackers over new administrative files, and treats cleanup candidates and performance claims as evidence to review—not permission to delete or merge.

**v1.1.0 · Python 3.10+ · Git · No third-party Python dependencies · Read-only tools**

Why this skill exists

AI-assisted development can leave behind overlapping plans, forgotten logs, intermediate datasets, and experimental outputs. Meanwhile, a single optimization branch may mix refactoring, algorithm changes, precision adjustments, and caching, making it difficult to tell what changed, which experiment measured it, or how to undo it.

AI Dev Steward addresses both problems without creating another layer of planning clutter.

Problem	Approach
Duplicate planning, planing, and versioned plan files	Keep one active plan per task. Preserve unique decisions and acceptance criteria before consolidating old documents. Filename matches are review hints, not proof of duplication.
Stale or unowned experiment artifacts	Inventory files and check lifecycle records, ownership, expiry, reproducibility, content hashes, and declared dependencies before identifying quarantine candidates.
Cleanup that could damage reproducibility	Retain unknown or protected artifacts and their transitive dependencies. Require explicit, path-level approval and a verified recovery procedure for manual quarantine.
Large, mixe