Architect

architect

The entry point for a new Robium application. It combines a lean design conversation with robotics stack selection, chooses the cheapest meaningful proof, records decisions without freezing provisional assumptions, and routes implementation to the skills that own each layer.

When to use this skill

• Starting a new robotics application from a rough idea.
• Choosing or reconsidering the stack, simulator, embodiment, environment, data path, visualization, or compute target.
• Decomposing a multi-subsystem robotics idea into a first buildable slice.
• Writing a durable decision record for a new app or genuine re-architecture.
• Skip architect for a bounded change in an existing flow. Read the affected code and use the matching tool skill directly.

Key directives

• Delegation posture: route + embed decision logic. Architect owns the product/stack choices and routing. Tool mechanics stay in the skill that owns them.
• Brainstorm only where a real decision exists. Inspect the repository and robium-apps registry first. Ask only questions whose answers materially change the first slice. For a genuine choice, offer 2–3 approaches with trade-offs and recommend one. Never announce a process classification.
• One direction gate, then build. A new app or material architecture gets one rough direction approval. After that, proceed through implementation and verification; pause only for a material pivot, scope expansion, paid or publishing action, deployment, or safety concern.
• Prove the riskiest assumption cheaply. Before polishing infrastructure or scaling compute, run the smallest probe that can disprove the stack and deliver the first user-visible working slice. Label provisional choices and update them when evidence changes.
• Use a decision record, not a contract. Create or update docs/architecture-brief.md for a new application or genuine re-architecture. It records validated decisions, provisional assumptions, risks, and authorized pivots. Bounded feature/fix work does not require a new brief or approval cycle.
• Environment first, syntax verified. Route reproducibility to environments before substantial code. Verify volatile versions, commands, and platform constraints against current primary sources or installed CLI help; never rely on memory.

Quick start

1. Read before asking. Inspect the app registry, closest reference app, current repository, and available hardware/compute facts.
2. Frame the first outcome. Establish robot/embodiment, user-visible task, sim-vs-real target, and the pass/fail signal for the first slice. Record a reasonable assumption instead of asking about low-impact details.
3. Compare only meaningful alternatives. Recommend one stack. If cost, hardware, or an unfamiliar integration is uncertain, define a cheap spike that resolves it before full implementation.
4. Write the decision record when warranted. Use references/brief-template.md for a new app or re-architecture. Keep the core concise; add optional detail only when the system needs it.
5. Bootstrap, then diverge. Check REGISTRY.md in the sibling robium-apps repository. Copy the closest app's environment, test, and launch shape when one exists; otherwise use references/scaffold-patterns.md.
6. Build the first slice. Hand each phase to the lowest owning skill and update the decision record only when a decision or risk actually changes.

Use the Claude-only robium-architect agent only for an explicitly requested or genuinely ambiguous heavy comparison that would otherwise flood the main conversation. Normal new-app design stays in the main conversation.

Decision guidance

Choose the first vertical

• Navigation: ros2 + nav2 + simulation → gazebo, then visualization → rviz2 or foxglove. Use environments before bring-up.
• Learned manipulation: lerobot + data + huggingface, with simulation selecting MuJoCo/Gazebo/Isaac as the task requires. Add isaac-sim/isaac-lab only after their NVIDIA GPU requirements are proven.
• Hybrid or unfamiliar stack: define module boundaries with integration, identify the costliest uncertainty, and spike that boundary before building the whole system.

Route by build phase

Platform gotchas

• Isaac Sim/Lab require a supported NVIDIA RTX-class GPU and do not run on macOS. Load isaac-sim and verify its current floor before selecting it.
• Native ROS 2/Gazebo development is Linux-centric. On macOS, choose a container or Linux remote through environments.
• Headless remote work needs web/recorded visualization (foxglove or rerun) rather than assuming RViz2/display forwarding.
• Cloud/GPU inventory, pricing, quotas, and framework versions are volatile. Verify them only when they affect the selected path; do not turn every design into a broad research pass.

Customization

• Scale the decision record to the project. A small one-process demo may need only outcome, chosen stack, environment, pass bar, and risks. Add modules, communications, data, or rollout sections only when they contain decisions.
• A discovered constraint may change a provisional choice without invalidating the whole design. Record the evidence and pivot; request another direction decision only when the user-visible outcome or material scope changes.
• Decompose an idea with several independent products into ordered slices. Fully design only the first slice; record later slices as non-binding follow-ups.

References

• references/stack-selection.md: simulator, middleware, and training decision trees; verify volatile defaults before use.
• references/scaffold-patterns.md: starting layouts for navigation and manipulation applications.
• references/brief-template.md: concise decision-record template with optional depth.
• examples/architecture-brief-example.md: filled hypothetical brief (status: unverified).
• Upstream: ROS 2, Nav2, Gazebo, LeRobot, Isaac Sim, and the MIT-licensed Superpowers brainstorming skill, whose useful exploration/alternatives/YAGNI principles were adapted here without its global classification and repeated-approval gates.

Changelog

• 2.0.0 (2026-08-27): absorbed lightweight brainstorming into architect; replaced global classification/repeated approval with one direction gate; made the brief a scalable decision record; added risk-first/user-visible slicing; narrowed the heavy Claude architect agent; repointed bootstrapping to robium-apps.
• 1.9.0 (2026-08-24): route RunPod inventory, provisioning, diagnostics, validation, billing, and cleanup to the new generic runpod skill.
• 1.8.1 (2026-08-03): style pass; removed em dashes throughout.
• 1.8.0 (2026-08-02): routing: learning-loop added; retired skill-updater.
• 1.7.0 (2026-08-02): routing table gained mining; skill-author narrowed.
• 1.6.1 (2026-08-01): anchor IDs added to claim-bearing items.
• 1.6.0 (2026-08-01): routing table gained cloud-run.
• 1.5.0 (2026-07-31): routed MuJoCo manipulation simulation.
• 1.4.0 (2026-07-18): routing table gained test-assets.
• 1.3.1 (2026-07-18): refreshed transferred repository links.
• 1.3.0 (2026-07-15): added backlog kickoff triggers and prior brainstorming handoff.
• 1.2.0 (2026-07-13): routing map gained live-demo.
• 1.1.1 (2026-07-12): clarified dated provenance.
• 1.1.0 (2026-07-12): made scaffolding bootstrap-first.