Multi-agent orchestrator for Claude Code — from idea to shipped product
Named after Tom Jobim — like bossa nova harmonizes simplicity with complexity, Jobim orchestrates specialized AI agents to build complete applications.
Install · Agents · Pipeline · PRD Loop · Contributing
One prompt. Eight agents. Full-stack delivery.
You describe what you want. Jobim (Claude Opus) breaks it into phases and delegates to 7 specialized sub-agents. Each agent has deep domain expertise, structured JSON contracts, and quality gates.
No manual coordination. No context-switching. From idea to deployed product.
/jobim new "Create a task management SaaS with real-time collaboration"
Jobim plans the work, delegates to Scout for market research, Builder for code, Designer for the design system, Tester for quality gates, and ships it.
┌──────────────────────────────────────────────────────────────┐
│ USER — provides high-level objective │
└────────────────────────────┬─────────────────────────────────┘
▼
┌──────────────────────────────────────────────────────────────┐
│ JOBIM (Opus) — Orchestrator │
│ Plans, delegates, synthesizes, decides │
│ Maintains state in .jobim/state.json │
│ NEVER executes code — always delegates │
└───────┬──────┬──────┬──────┬──────┬──────┬──────┬────────────┘
▼ ▼ ▼ ▼ ▼ ▼ ▼
┌──────────────────────────────────────────────────────────────┐
│ SUBAGENTS (Haiku / Sonnet) │
│ │
│ Scout Builder Tester Designer UXer Shipper Launcher │
│ (Haiku) (Sonnet) (Sonnet) (Sonnet) (Sonnet) (Sonnet) (Sonnet) │
│ │
│ Execute specific tasks → Return structured JSON │
└──────────────────────────────────────────────────────────────┘
| Agent | Model | What It Does |
|---|---|---|
| Jobim | Opus | Orchestrates the full pipeline. Plans, delegates, synthesizes results. Never writes code directly. |
| Scout | Haiku | Market research, competitor analysis, tech stack recommendation, viability scoring |
| Builder | Sonnet | Full-stack implementation — React, Next.js, APIs, databases, responsive layouts |
| Designer | Sonnet | Design systems, HSL color tokens, typography scales, spacing, visual effects |
| UXer | Sonnet | Flow analysis, friction audit, Nielsen heuristics, WCAG accessibility, conversion optimization |
| Tester | Sonnet | Code review, OWASP security scan, test writing, quality gates |
| Shipper | Sonnet | Docker configs, GitHub Actions CI/CD, deploy pipelines, monitoring |
| Launcher | Sonnet | README generation, social posts (X/LinkedIn), Product Hunt copy, launch checklist |
Five phases, each with dedicated agents and quality gates:
Discovery → Prototype → Production → Ship → Launch
Scout Builder Builder + Shipper Launcher
(Haiku) (Sonnet) Tester (Sonnet) (Sonnet)
(parallel)
Jobim persists state across the session in .jobim/state.json:
{
"version": "2.0",
"phase": { "current": "discovery", "completed": [], "history": [] },
"context": {
"discovery": null,
"prototype": null,
"production": null,
"ship": null,
"launch": null
},
"decisions": [],
"artifacts": []
}git clone https://github.com/xiapeli/jobim.git
cd jobim
# Copy agents and commands (won't overwrite existing files)
mkdir -p ~/.claude/agents ~/.claude/commands
cp -n .claude/agents/*.md ~/.claude/agents/
cp -n .claude/commands/*.md ~/.claude/commands/git clone https://github.com/xiapeli/jobim.git
cd jobim
# Back up your existing config first
[ -d ~/.claude ] && cp -r ~/.claude ~/.claude.backup.$(date +%s)
# Install everything
mkdir -p ~/.claude/agents ~/.claude/commands
cp .claude/agents/*.md ~/.claude/agents/
cp .claude/commands/*.md ~/.claude/commands/
cp .claude/settings.json ~/.claude/settings.json# Full autonomous pipeline
/jobim new "Create a task management SaaS"
# Interactive mode — pause for approval at each phase
/jobim new "Create a meditation app" interactive
# Run a specific phase only
/jobim run design
# Check progress
/jobim status
# Use individual agents directly
/scout "Research design system trends 2026"
/builder "Create a responsive landing page"
/designer "Design system for a fitness app"
/tester "Review the auth module for security issues"Jobim includes a PRD-driven autonomous loop based on the Ralph Pattern. Define user stories in a prd.json, and Jobim implements them one by one:
cd prd-loop
cp prd.json.example prd.json # edit with your user stories
./jobim-loop.sh 10 # run up to 10 iterationsEach iteration:
- Reads the PRD and progress log
- Implements the next incomplete user story
- Commits the work
- Updates the progress file
- Stops when all stories pass or max iterations reached
Note: The PRD loop uses
--dangerously-skip-permissionsto run without interactive prompts. This allows fully autonomous operation but gives Claude unrestricted tool access. Review the prompt.md before running.
Jobim includes a curated design library with:
- 159 reference screenshots from award-winning sites
- Curated references organized by category (SaaS, portfolio, e-commerce, etc.)
- Design system patterns extracted from top implementations
See design-library/ for the full collection.
Every agent follows structured contracts:
- Input contract: Required and optional parameters with types
- Output contract: JSON schema defining the response format
- Quality gates: Conditions that must pass before Jobim accepts the output
When a quality gate fails, Jobim retries (up to 3 times), then escalates with full context.
- Claude Code with API access
- Claude Opus for the orchestrator (Jobim agent)
- Claude Sonnet/Haiku for sub-agents
Contributions welcome. See CONTRIBUTING.md for guidelines.
Ideas for contributions:
- New specialized agents (e.g., Analytics, Copywriter, SEO)
- Additional design library references
- Language translations for agent prompts
- Example PRDs for common project types
MIT — Phelipe Xavier, 2026