My Claude Code

March 23, 2026 · View on GitHub

Personal collection of workflows, skills, plugins, and configurations for Claude Code.

Structure

.
├── workflow/             # Five-stage development workflow system
│   ├── CLAUDE.md         # Entry point — workflow stages + principles reference
│   ├── principles.md     # Global principles governing all decisions
│   ├── agents/           # 8 specialized agents (critics, synthesizers, verifiers)
│   └── skills/           # 5 workflow skills (problem-def, design, planning, impl, debug)
├── skills/               # Standalone skills (not part of the workflow)
├── plugins/              # Full plugins with hooks, scripts, and commands
├── commands/             # Slash commands for common workflows
└── settings.json         # Claude Code configuration

Workflow System

A five-stage, multi-model development pipeline. Problems are defined before solutions are proposed. Proposals are generated competitively and verified against principles, not documents.

StageSkillWhat It Does
1. Problem Definition/problem-definitionFrame the problem. Three review agents attack the framing in parallel.
2. Design/designExplore solution space. Codex X-High evaluates approaches. User picks direction.
3. Planning/planningTwo Codex High agents propose independently. Opus synthesizes. User approves.
4. Implementation/implementationSub-agents build in parallel. Opus verifies each against problem + principles.
5. Debugging/debuggingRCA → dual proposal (Opus + Codex) → synthesis → alignment → implement.

Model Routing

RoleModelWhy
Research validationCodex X-High (via pi)Analytical rigor, gap detection
Approach evaluationCodex X-High (via pi)Tradeoff analysis
Proposal generationCodex High x2 (via pi)Strategy across problem space
Synthesis & alignmentOpusIntent understanding, judgment
VerificationOpusDivergence evaluation
ImplementationClaude sub-agentsTool access, codebase interaction

Agents

AgentModelRole
definition-criticOpusAttacks problem definitions
context-analystOpusMaps landscape around a problem
assumptions-auditorOpusSurfaces invisible assumptions
proposal-synthesizerOpusSelects best proposal, justifies choices
research-validatorCodex X-HighEvaluates research quality, directs follow-up
approach-evaluatorCodex X-HighCompares solution approaches rigorously
build-verifierOpusRuns tests, generates adversarial cases, gates on execution
integration-checkerOpusVerifies parallel-built pieces fit together

Key Design Decisions

  • Proposal as compass, not contract. Implementation agents have latitude. Verify against the problem, not the document.
  • Select, don't merge. The synthesizer picks the better proposal as a base. Merging is the exception.
  • Verify by executing. Build verification runs tests and generates adversarial cases. Code review alone is rubber-stamping.
  • Scope through context, not personas. Agent quality comes from controlling what context they receive, not from elaborate role descriptions.

Standalone Skills

SkillDescription
agent-browserBrowser automation for web testing, form filling, screenshots
deep-researchDeep research using Parallel.ai API
pi-agentDispatch tasks to pi coding agent for cross-model work
skill-creatorGuided skill generation with validation
skill-from-mastersCreate skills based on expert methodologies
learningsKnowledge base with maturity pipeline: [DRAFT][CONFIRMED][REGRESSION]
yt-transcribeYouTube transcription with smart model routing

Plugins

PluginDescription
voice-mlxVoice notifications, iTerm2 status, macOS alerts

Installation

# Copy the workflow system
cp workflow/CLAUDE.md ~/.claude/CLAUDE.md
cp workflow/principles.md ~/.claude/principles.md
cp -r workflow/agents/* ~/.claude/agents/
cp -r workflow/skills/* ~/.claude/skills/

# Copy a standalone skill
cp -r skills/agent-browser ~/.claude/skills/

# Install a plugin
cp -r plugins/voice-mlx ~/.claude/plugins/custom/voice-mlx

Requirements

  • Workflow system: Requires pi CLI (npm install -g @mariozechner/pi-coding-agent) for cross-model dispatch
  • deep-research: Requires PARALLEL_API_KEY
  • agent-browser: Requires agent-browser CLI
  • voice-mlx: Requires edge-tts, ffplay, terminal-notifier

License

CC BY-NC 4.0 - Share and adapt with attribution. Commercial use requires prior consent.