AI for System Design

TL;DR

• AI can be a strong thinking partner for architecture: brainstorm options, identify risks, compare trade-offs. Claude Artifacts, Claude Code, Cursor Agent, and ChatGPT o3 each bring different strengths.
• It cannot make the final call. It doesn't know your org, team, or constraints.
• Use it to expand your thinking, not to replace it.

System design is judgment-heavy. AI can't attend your design review. But it can help you prepare, stress-test, and fill gaps before you get there.

What AI Is Good At

• Generating options — "What are 3 ways to add caching to this system?" You get a starting menu.
• Trade-off analysis — "Compare event-driven vs. request-response for this use case." Structured pros/cons.
• Identifying risks — "What are the failure modes of this design?" Often surfaces things you hadn't considered.
• Pattern recall — "What's the standard pattern for X?" CQRS, saga, circuit breaker, etc.
• Documentation — "Turn this whiteboard sketch into an architecture doc." Draft, not final.

What AI Can't Do

• Know your constraints — Team size, budget, legacy systems, compliance, politics. You supply that.
• Make the decision — "We should use Kafka" vs "We should use SQS" depends on your context. AI can't weigh it.
• Account for org fit — "We don't have Kafka expertise" — AI doesn't know that.
• Invent novel solutions — AI recombines known patterns. Truly novel architectures are still human.

Tools That Actually Help (Feb 2026)

• Claude Artifacts UI — Renders code, diagrams, and React components side-by-side. Strong for architectural review.
• Claude Code — 200k+ token context for full codebase review and legacy refactors. Great for "what if we add a caching layer to this whole thing?"
• Cursor Agent Mode — Traverses project folders, edits multiple files, refactors architecture autonomously. Use for implementation-level design validation.
• Devin — Specializes in legacy migration (COBOL, Fortran, Objective-C to modern). Nubank saw 12x engineering efficiency and 20x cost savings on a 6M+ line migration. Best for bulk, well-scoped migrations.
• ChatGPT o3 — "Deep research" to pull design patterns from many sources. Overkill for simple questions; use when you need breadth.

A Practical Workflow

1. Draft your design — Rough architecture, main components, data flow.
2. Paste into AI — "Here's my design. What am I missing? What are the risks?"
3. Iterate — "We have to use Postgres, not a new DB." "What changes?"
4. Stress-test — "How does this fail under load? Under partial outage?"
5. You decide — Take the feedback. Make the call. Own it.

Prompt Patterns for Design

• "Compare approach A and B for [goal]. Consider: scale, ops complexity, cost. We're a team of 5."
• "What are the failure modes of this design? Rank by likelihood and impact."
• "We're adding [feature] to [existing system]. What's the least disruptive approach?"
• "Explain the trade-offs of [pattern] (e.g., event sourcing) for our use case: [brief description]."

Common Pitfalls

• Over-engineering — AI loves to suggest the "right" pattern. Sometimes simple is better. Push back.
• Outdated advice — Managed services and cloud offerings change. Verify "best practices" against current docs.
• Ignoring humans — The best design is one your team can build and operate. AI doesn't know your people.

Do This Next

1. Paste a design into Claude Artifacts — Get code + diagrams side-by-side. See how it handles your architecture.
2. Run your next design through Claude Code — Use its 200k context if you have a large codebase. Ask for risks and alternatives.
3. Create a "design review" prompt — Template that includes your constraints (team size, cloud provider, latency). Reuse it.