OmniDev Kit — Summary

OmniDev Kit is a Claude Code skill (/od) that orchestrates a 5-phase AI development workflow (Assess → Blueprint → Plan → Dev → Test) with extensive safety guardrails, bilingual support (Chinese/English), dynamic skill composition, T-shirt sizing for scope-adaptive phase skipping, and a self-evolution engine. It integrates concepts from superpowers, OpenSpec, planning-with-files, and everything-claude-code into a single unified skill. Commands are triggered by the /od prefix (22 commands with short aliases). Key features include: requirement alignment protocol that blocks ambiguous execution, lazy context loading per phase, cross-session memory via YAML frontmatter + Markdown files, interactive mode with numbered choice menus, and a self-learning system that proposes rule improvements from session retrospectives. With 1 star and PowerShell scripts for install, it targets Chinese enterprise teams.

Compared to seeds: closest to superpowers (skill-based behavioral framework) but with far more command surface (22 vs. 14 skills), phased workflow structure borrowed from openspec, and session state persistence similar to taskmaster-ai. The self-evolution engine has no direct seed equivalent.

OmniDev Kit — Overview

Origin

Created by GitHub user zy-eagle (published as zayeagle). Active development in May 2026. PowerShell install scripts suggest targeting Windows enterprise environments. README is bilingual (Chinese primary, English secondary).

Philosophy

From README:

"OmniDev Kit is an AI-driven development workflow toolkit that transforms the AI from a 'typist who only writes code on command' into a 'senior R&D engineer who understands cost control, architecture design, writes their own tests, and never forgets.'"

First Principles Safety Rules

Rule B.3 — Requirement Alignment:

"When requirements are vague, ambiguous, or missing key info, the AI is prohibited from guessing or self-interpreting. It must proactively confirm the core problem, final goal, delivery criteria, and root cause."

Rule B.4 — Problem Fix Protocol:

"For any bug fix, security patch, or behavior correction, the AI is prohibited from shipping a quick patch. It must first produce a complete solution plan with root cause analysis, impact scope, and regression risk."

Self-Description

From the README:

"OmniDev Kit is an AI-driven development workflow toolkit that integrates the core essence of top open-source AI frameworks (such as obra/superpowers, OpenSpec, planning-with-files, and everything-claude-code)"

This is the only framework in the batch that explicitly claims to synthesize multiple seed frameworks.

Bilingual Design

Chinese and English locale support:

• /od cfg -l zh → Chinese mode (loads phases/zh/ files)
• /od cfg -l en → English mode (loads phases/en/ files)
• Only one locale loaded at a time — saves tokens

Activation Pattern

The skill activates ONLY when user message starts with /od. Without that prefix, behaves as normal conversation.

OmniDev Kit — Architecture

Distribution

• Type: standalone-repo (skill files + rules + install scripts)
• Install: Remote URL method (preferred):

  /od install https://github.com/zayeagle/omnidev-kit.git
  

  Or manual copy of skills/ and rules/ directories.

Runtime Requirements

• Claude Code
• PowerShell (for install scripts on Windows)

Directory Structure

omnidev-kit/
├── INSTALL.md
├── README.md
├── README.zh-CN.md
├── rules/
│   └── 01-omnidev-workflow.mdc    # Lightweight trigger (alwaysApply: false)
└── skills/
    └── od/
        ├── SKILL.md               # Main spec — single source of truth
        ├── phases/
        │   ├── 00-assessment.md   # Root-level fallback
        │   ├── 01-02-planning.md
        │   ├── 03-development.md
        │   ├── 04-testing.md
        │   ├── zh/                # Chinese locale
        │   │   ├── 00-assessment.md
        │   │   ├── 01-02-planning.md
        │   │   ├── 03-development.md
        │   │   └── 04-testing.md
        │   └── en/                # English locale
        │       ├── 00-assessment.md
        │       ├── ...
        └── engine/
            ├── evolution.md        # Self-evolution engine
            ├── special-flows.md
            ├── zh/
            └── en/

Lazy Loading Architecture

Each phase declares context_requires listing exactly which files to load. The AI loads ONLY the listed files — no pre-reading downstream phases. This keeps token usage proportional to the current phase.

State Files

docs/omnidev-state/
├── context.yaml       # YAML frontmatter + Markdown session state
├── current-task.md
└── evolution-log.md   # Self-learning retrospective log

OmniDev Kit — Components

Main Skill: od (SKILL.md)

Activates on /od prefix only.

Commands (22)

Phase Files (5 per locale = 10 total)

Engine Files (2 per locale = 4 total)

Rules

OmniDev Kit — Uniqueness

differs_from_seeds

Closest seeds: superpowers (skill-based behavioral framework, prefix-triggered activation) and openspec (phased workflow with explicit artifact generation). OmniDev differs from superpowers in: (1) explicit 22-command surface with phase navigation (vs. superpowers' 14 autonomous skills with no commands); (2) session state persistence via YAML frontmatter in docs/omnidev-state/ (superpowers has no persistent state); (3) self-evolution engine that can propose modifications to its own rules. OmniDev differs from openspec in: OpenSpec focuses on planning artifacts before implementation; OmniDev covers the full dev lifecycle (assess → blueprint → plan → dev → test) with security, TDD, and enterprise reporting included in the same skill. The self-evolution engine — where the agent logs its own corrections and proposes rule changes subject to human approval — has no equivalent in any seed framework. The bilingual locale system (loading different phase files per language at runtime) is also unique.

Positioning

"Everything-in-one" senior developer assistant for Chinese enterprise teams. The self-description "integrates the core essence of obra/superpowers, OpenSpec, planning-with-files, and everything-claude-code" is accurate — this is the most feature-integrated framework in the batch, at the cost of complexity. The lazy loading architecture is a genuine performance contribution: other frameworks load all their prompts into context at session start; OmniDev loads only the files for the current phase.

Observable Failure Modes

1. Activation guard is easy to bypass: the skill only activates on /od prefix — any conversation that starts without it proceeds without OmniDev's safety guardrails
2. Self-evolution instability: proposals that pass /od ln y approval modify the skill's own rules. A series of poorly-reviewed proposals could degrade the skill's behavior in non-obvious ways.
3. Chinese-first UX: while bilingual, the default state file contents and many error messages appear to be Chinese, which may confuse English teams using the tool in en mode
4. No artifact schema: unlike OpenSpec's explicit artifact graph, OmniDev's phases produce blueprint.md and plan.md in unspecified formats — the schema is defined only in the phase prompt files, not machine-validated
5. External skill discovery is fragile: scanning for skills in 4 local directories will produce different results on different machines, making the Dynamic Skill Composition behavior unpredictable in teams