Genesis follows a four-phase workflow. Each phase must complete before the next begins, and you can review and adjust at every decision point.
Phase 1: Interview
Genesis extracts four key elements from your opening message:
- Project name (kebab-case identifier)
- Purpose (one-sentence description)
- Tech stack (language, framework, runtime)
- Key integrations (databases, APIs, auth, message queues)
The number of follow-up questions scales with how much detail you provide. A message like "Create a Go REST API called inventory-service with PostgreSQL and JWT" needs only one or two clarifying questions. Something vague like "I need a new backend service" triggers all four foundational questions.
Genesis reads your existing environment.md and personalisation.md to avoid re-asking setup information.
Context-aware scaffolding
Every generated file (CLAUDE.md, agents, skills, memory, settings) consumes context tokens. Project size must match your plan tier's capacity:
| Profile | Plan | Context | Agents | Skills | Cost |
|---|---|---|---|---|---|
| Lean | Pro | 200k | 1–2 | 1–2 | ~5–8k |
| Standard | Max | 200k | 2–3 | 2–3 | ~10–15k |
| Full | ProMax/API | 1M | 3–4 | 3–4 | ~15–25k |
Lean profile
- Condensed CLAUDE.md with merged sections
- Single consolidated MEMORY.md
- Only the most critical domain agent plus workflow agents
- Context management guidance built in
Standard profile
- Full CLAUDE.md template
- Complete memory files
- 2–3 domain agents
- 2–3 dynamic skills
Full profile
- Detailed CLAUDE.md with extended sections
- Comprehensive agent selection (3–4 agents)
- 3–4 dynamic skills
- Extended session capacity
All projects include a statusline displaying current model and context usage percentage, making it easy to know when to run /compact.
Phase 2: Plan
Genesis presents a structured plan showing:
- Project name and path
- Technology stack
- Domain agents (specialised roles)
- Workflow agents (test-runner, code-reviewer, doc-writer)
- Base and dynamic skills
- MCP servers
- Folder structure
You can request modifications before confirming: add integrations, remove agents, change folder structures. Nothing is generated until you approve.
Phase 3: Generate
Generation creates the entire project in a single pass.
Infrastructure files
- CLAUDE.md (project standards and rules)
- .claude/settings.json (permissions and hooks)
- .claude/agents/*.md (individual agent instructions)
- .claude/skills/*/SKILL.md (skill directories)
- .mcp.json (external integration configs)
Memory files
Written to ~/.claude/projects/-home-xeeva-claude-<name>/memory/, outside the project directory. Claude reads these automatically at the start of each session.
- MEMORY.md (index)
- user-profile.md (user preferences)
- project-context.md (project details)
Application boilerplate
- Entry points (main.go, src/main.py, src/index.ts, etc.)
- Package configuration files
- Test setup and sample tests
- Documentation (README.md, architecture.md)
- Stack-specific configurations
Template system
Genesis uses {{PLACEHOLDER}} templates from .claude/skills/genesis/templates/. Templates reference stack profiles for language-specific conventions, folder structures, linters, test frameworks, and error handling patterns.
Agent selection
Domain agents are selected from a catalogue based on project type, capped at 3–4. Three workflow agents are always included (test-runner, code-reviewer, doc-writer). Projects with external integrations also get a risk-evaluator agent.
Phase 4: Finalise
Git initialisation
cd ~/claude/<name> && git init && git add -A && git commit -m "Initial scaffold from Genesis"Every project starts with a clean git history containing one commit with the full scaffold.
Registry update
Genesis maintains a project registry in memory, logging each project's name, path, stack, date, and status.
Summary
A creation confirmation with all generated components and next steps for accessing the project.
Complete example
Here's a full walkthrough creating a Python FastAPI project:
Opening message
"Create a Python FastAPI service called user-service for managing user accounts.
It needs PostgreSQL for storage and JWT for authentication."All four information pieces are present, so no follow-up questions are needed.
Plan output
The plan shows 4 domain agents (api-designer, data-modeller, auth-specialist, security-reviewer), 3 workflow agents, 6 skills, and a PostgreSQL MCP server.
Generation
All infrastructure, application boilerplate, configuration files, and memory files are created at ~/claude/user-service/.
Finalisation
Project confirmed with component summary and next steps. cd ~/claude/user-service && claude to start working.