claude-workflow

A portable Claude Code Standard Workflow template. Drop CLAUDE.md into any repo and Claude follows the full development process — on CLI, Desktop, IDE, and Claude Cloud.

Integrates superpowers + Everything Claude Code (ECC) skills into one coherent workflow.

What's Inside

File	Purpose
CLAUDE.md	The workflow itself — Claude reads and follows this
setup.sh	One-time plugin installer — run after cloning
docs/CLAUDE.template.md	Same workflow with fillable project sections
docs/claude-setup.sh	Auto-setup for existing repos (creates CLAUDE.md + settings)
docs/settings.local.template.json	Open permissions (skip-permissions equivalent)
.claude/settings.local.json	Pre-configured open permissions

The Workflow (8 Steps)

Step	Skills (superpowers + ECC)	What Happens
Session Start	MemPalace wake-up + git log + active plan	Load project memory, brief user on state
0. Research	search-first → docs → deep-research → exa-search	Find existing solutions before building
1. Brainstorm & Spec	superpowers:brainstorming → spec-expander → spec-reviewer	Generate ideas, write spec, adversarial review (KILL/FIX/SHIP)
2. Plan	superpowers:writing-plans + planner agent	Structured plan with phases and risks
3. Implement	superpowers:executing-plans + tdd + quality-gate	TDD: RED → GREEN → IMPROVE, auto quality checks
4. Review	superpowers:requesting-code-review + language reviewer + security-reviewer + database-reviewer	Quality + security + DB review
5. Verify	superpowers:verification-before-completion + context-budget	Verify completion + check token usage
6. Git	superpowers:finishing-a-development-branch	Conventional commits, worktrees
7. Learn	learn / learn-eval → instinct-status → promote	Extract patterns, build project instincts
Session End	MemPalace store decisions + state + handoff note	Persist everything for next session

Also Included

• 3-Agent Harness — planner → implementer → evaluator for complex features
• Autonomous Execution — Ralph Loop pattern for hours of autonomous work (Cloud: /schedule, CLI: claude -p, Remote: --remote)
• Agent Teams — Multi-agent parallel execution with shared task lists and direct communication
• 6 Agent Definitions — planner, implementer, evaluator, reviewer, researcher, security-reviewer (in .claude/agents/)
• Continuous Learning — MemPalace taxonomy: halls (decisions/patterns/anti-patterns/fixes), rooms (topics), wings (scope)
• Cloud-Ready Config — .claude/settings.json + .mcp.json committed to repo → 100% Cloud-Funktionalität
• Model-Routing — Haiku / Sonnet / Opus per task type with cost table
• Design-Routing — brainstorming → frontend-design → ui-ux-pro-max → frontend-patterns
• Context Management — strategic-compact, context-budget, save-session, resume-session + MemPalace
• Debugging — systematic-debugging, build-fix, codex:rescue, security-scan

Quick Start

Option 1: Claude Cloud (zero setup)

1. Click "Use this template" on GitHub → new repo
2. Open on claude.ai/code → connect repo
3. Start working. Everything activates automatically:
   • MemPalace auto-installs on first session (SessionStart hook)
   • All agents, skills, hooks, MCP config are repo-committed
   • Agent Teams enabled

Option 2: CLI / Desktop / IDE

git clone https://github.com/YOUR_USER/YOUR_NEW_REPO.git
cd YOUR_NEW_REPO
bash setup.sh    # installs plugins (superpowers, ECC, codex, ...) + MemPalace
claude           # start working — workflow is active

Option 2: Setup Script (any existing repo)

curl -sL https://raw.githubusercontent.com/moeffel/claude-workflow/main/docs/claude-setup.sh | bash

Auto-detects language (TypeScript/Python/Go/Rust/PHP) and framework (Next.js/Vite/FastAPI/Django/Express).

Option 3: Manual Copy

curl -sO https://raw.githubusercontent.com/moeffel/claude-workflow/main/docs/CLAUDE.template.md
mv CLAUDE.template.md CLAUDE.md
mkdir -p .claude
curl -sO https://raw.githubusercontent.com/moeffel/claude-workflow/main/docs/settings.local.template.json
mv settings.local.template.json .claude/settings.local.json

Permissions

Equivalent to --dangerously-skip-permissions:

{
  "allow": ["Read", "Edit", "MultiEdit", "Write", "Bash(*)", "WebSearch", "WebFetch", "mcp__playwright__*"],
  "deny": ["Bash(rm -rf /)", "Bash(rm -rf ~)", "Bash(sudo rm -rf *)"]
}

Only catastrophic rm -rf and sudo are blocked.

After Setup

1. Fill in project sections at the top of CLAUDE.md (Tech Stack, Commands, Architecture)
2. The workflow works immediately — Claude reads the mandatory steps and follows them
3. Add .claude/skills/ for domain-specific knowledge (optional)
4. Add .claude/hooks/ for project-specific validation (optional)

Prerequisites

The workflow references skills from these plugins. Install them for the full experience:

Plugin	Install	Provides
superpowers	claude plugin marketplace add obra/superpowers-marketplace → claude plugins install superpowers@superpowers-marketplace	brainstorming, writing-plans, executing-plans, TDD, verification, debugging
Everything Claude Code	claude plugins install everything-claude-code	quality-gate, learn, instinct-status, context-budget, strategic-compact, build-fix, reviewers, agents
Codex	claude plugins install codex	/codex:rescue, /codex:review, /codex:adversarial-review
ui-ux-pro-max	claude plugins install ui-ux-pro-max	Design intelligence for UI tasks
MemPalace	pip install mempalace + claude mcp add mempalace -- python -m mempalace.mcp_server	Persistent cross-session memory (96.6% LongMemEval, local-first, free)

The workflow degrades gracefully — skills that aren't installed are simply skipped.

Works On

Environment	CLAUDE.md	Settings	Plugins/Skills
Claude Code CLI	Yes	Yes	Yes
Claude Desktop App	Yes	Yes	Yes
VS Code / JetBrains	Yes	Yes	Yes
Claude Cloud (claude.ai/code)	Yes	No*	Partial**

* Claude Cloud can't read ~/.claude/ (local). That's why the workflow lives in the repo's CLAUDE.md. ** Plugins may not be available on Cloud. The workflow steps still guide Claude's behavior.

License

MIT