Curia

Multi-project development platform powered by Claude Code.

Concept

• Place applications under src/ and operate them with shared processes, agents, and plugins
• src/ is gitignored — anyone can clone and place their own applications
• Generated documents are also gitignored (docs/generated/). Only templates are committed

Setup

git clone <repository-url>
cd Curia

# Place your application
cp -r /path/to/my-project src/MyProject

# Create a project-specific CLAUDE.md
# Describe tech stack, architecture, etc. in src/MyProject/CLAUDE.md

# Link generated docs (optional)
ln -s ../../src/MyProject/docs docs/generated/MyProject

Directory Structure

Curia/
├── .claude/
│   ├── settings.json          # Shared settings (permissions + plugins + hooks)
│   ├── settings.local.json    # Local overrides (gitignored)
│   ├── agents/                # Custom agent definitions
│   │   ├── pdm.md             # PdM (Product Manager)
│   │   └── tech-lead.md       # Tech Lead
│   └── skills/                # Process definitions (slash commands)
│       ├── requirements/      # /requirements — Requirements definition
│       ├── develop/           # /develop — Development
│       ├── acceptance/        # /acceptance — Acceptance testing
│       └── qa/                # /qa — Cross-cutting QA (ad-hoc)
├── docs/
│   ├── templates/             # Document templates (committed)
│   └── generated/             # Generated documents (gitignored)
├── src/                       # Applications (gitignored)
├── CLAUDE.md                  # Platform design philosophy & process definitions
└── README.md

Plugin-First Architecture

Plugins are the primary actors. Custom agents only cover areas that plugins cannot.

Plugins (Primary)

Plugin	Type	Role	Phase
feature-dev:feature-dev	Skill	Guided development — analysis → design → implementation → quality review	Development
feature-dev:code-explorer	SubAgent	Codebase exploration — execution path tracing, architecture mapping (read-only)	Requirements, QA
feature-dev:code-architect	SubAgent	Architecture design — pattern analysis → implementation blueprints (read-only)	Requirements
feature-dev:code-reviewer	SubAgent	Code review — bugs, logic errors, quality checks, confidence ≥ 80 only (read-only)	Review, QA
pr-review-toolkit:review-pr	Skill	Comprehensive PR review — 6 specialized agents in parallel	Review, QA
code-review:code-review	Skill	GitHub PR auto-review — confidence ≥ 80 comments only	Review
frontend-design:frontend-design	Skill	Design-quality frontend implementation for new pages/components	Development (frontend)
security-guidance	Hook	Auto-detect XSS, command injection, etc. on Edit/Write	All phases (automatic)

Custom Agents (Complementary)

Agent	Role	Why plugins can't cover this
PdM	Business value judgment, acceptance testing (Playwright MCP), UI/UX review	Business domain knowledge and user value assessment
Tech Lead	Security review (OWASP), API integration verification (curl), business logic consistency	Security judgment, real API execution, domain-specific logic

Note: The team leader role (process control, task assignment, integrated verdict) is embedded in each skill (/develop, etc.). No dedicated agent is defined.

Development Process

3-phase structure for all domains:

/requirements <project> <domain>  →  Requirements definition
/develop <project> <domain>       →  Development + review + merge
/acceptance <project> <domain>    →  Acceptance testing by PdM

Ad-hoc QA (outside the main process):

/qa <project> <portal>            →  Full-screen verification + code quality review

Phase skipping is prohibited. Merging before acceptance testing is prohibited.

Workflow

/requirements:
  code-explorer (parallel) + code-architect (design) // PdM (business) → TechLead (supplement) → consensus

/develop:
  [New feature] feature-dev:feature-dev (analysis → design → implementation → quality review)
  [New UI]      frontend-design:frontend-design (design-quality frontend)
  [Bug fix]     code-explorer (root cause) → fix → code-reviewer (quality)
  → pr-review-toolkit:review-pr + code-review:code-review (PR review)
  → TechLead (security / API / business logic)
  → Skill integrates verdict and merges

/acceptance:
  PdM (Playwright MCP screen testing + UI/UX review)

/qa (ad-hoc):
  PdM (full-screen verification) + pr-review-toolkit (code quality) + TechLead (security/business)

Project CLAUDE.md

Each project declares its specifics in src/{project}/CLAUDE.md:

Section	Content
Overview	Project purpose and background
Tech Stack	Languages, frameworks, databases
Architecture	Layer structure, directory layout, naming conventions
Reference Projects	Pattern sources (paths and reference points)
CI Commands	Build, test, lint commands
Docker Config	Container names, execution methods
UI Hostname	Frontend access URL (direct localhost access is blocked by hooks)
Implementation Phases	Domain order and phase plan

Hooks

Hook	Matcher	Behavior
localhost block	Bash, Playwright navigate	Block direct localhost:<port> access. Use project CLAUDE.md UI hostname
Pre-commit UI check	Bash (git commit)	Remind to verify UI with Playwright MCP before committing
Secret file guard	Bash (git add)	Block staging of .env, credentials, secrets files
Force push block	Bash (git push)	Block git push --force / -f
Main branch guard	Bash (git push)	Block direct push to main / master
Screenshot path guard	Playwright screenshot	Block saving screenshots outside docs/generated/
Security patterns	Edit, Write	Auto-detect XSS, command injection, unsafe deserialization (via security-guidance plugin)