dmjcustomizations：面向Claude Code的并行优先代理工程技能集

Project Overview

dmjcustomizations is a modernized agent engineering skill set designed for Claude Code, developed by divyamohan1993 and hosted on GitHub (updated 2026-06-10). It features a parallel-first architecture, adversarial validation mechanisms, and integrates safety/performance保障 from the first line of code.

Key highlights:

• Parallel agent teams with minimal human checkpoints
• Evidence-driven validation to ensure reliability
• Safety left-shift and performance budget enforcement
• Dynamic adaptation to latest models/tools

Subsequent floors will dive into background, design principles, skill system, technical features, and application value.

Context

AI-assisted programming tools like Claude Code are transforming developer workflows, but maximizing their potential remains a challenge. dmjcustomizations is a modern refactor of the early superpowers project, tailored for the era of agent teams.

Core Insight

Traditional single-agent, manual-supervision workflows are limited. When AI capabilities are strong enough, multiple agents should work in parallel—only requiring human approval at key checkpoints instead of constant oversight. This 'parallel-first'理念 guides the project.

1. Parallel-First, Serial把关

Agents work in parallel; users only approve at hard checkpoints, avoiding workflow blockages.

2. Evidence Over Claims

• Specifications are cross-validated by independent agents
• Code undergoes multi-dimensional reviews (correctness, security, performance, simplicity)
• All 'completion' claims require evidence

###3. Demo Over Description Competitive solutions are tested as workspikes with benchmarks instead of text comparisons.

###4. Dynamic Design No hardcoded model/tool versions—skills dynamically detect the strongest models and latest tools.

###5. Safety & Performance From First Line

• Threat model analysis in design phase
• OWASP compliance, zero-trust architecture, quantum-safe encryption by default
• Performance budgets enforced in CI

The skill set covers the full software lifecycle:

Meta Skills

using-dmjcustomizations: Auto-injected meta skill for skill discovery, routing, and priority sorting.

Planning & Execution

• brainstorming: From idea to approved specs via parallel scanning
• writing-plans: Specs to execution plans with task dependencies
• executing-plans: Parallel team execution with review checkpoints

Quality & Security

• test-driven-development: Iron Law TDD + extreme boundary case testing
• verification-before-completion: Evidence-based validation
• defending-in-depth: Safety left-shift practices

Other categories: Development collaboration, infrastructure tools, research, UX.

Context Budget Control

Each skill has strict context limits (500 words max for main content, 300 for meta skills) to maintain efficiency.

Simplicity

Skills use terse descriptions, avoid hardcoded dates, and prioritize precision.

Installation

• Install from GitHub Market or local path
• Uninstall superpowers to prevent rule conflicts

Open Source

Forked from obra/superpowers 5.1.0 (MIT license), refactored in 2026.

For Developers

1. Optimize workflows: Let agents work in parallel, intervene only at key points
2. Ensure quality: Adopt evidence-driven validation
3. Prioritize safety: Integrate security from day one
4. Enforce performance: Treat performance budgets as first-class citizens
5. Adapt dynamically: Avoid hardcoded assumptions

For AI Engineering

The project combines traditional SE practices (code review, TDD) with AI agent capabilities to build reliable, efficient workflows.

dmjcustomizations is more than a tool set—it's a methodology for collaborating with AI agents. Its parallel-first, evidence-driven, safety-inbuilt理念 will become increasingly important as AI capabilities grow.

For teams/individuals looking to enhance AI-assisted development efficiency, this project is a valuable reference implementation.