System DesignAI assistants are genuinely useful for exploring engineering tradeoffs, but they have a specific failure mode you need to understand: they are diplomats, not decision-makers. They will lay out options beautifully and then refuse to commitWhat is commit?A permanent snapshot of your staged changes saved in Git's history, identified by a unique hash and accompanied by a message describing what changed. to a recommendation. Knowing this limitation lets you use them effectively.
What AI does well and poorly
| Strength | Weakness |
|---|---|
| Listing pros and cons of multiple options | Making a strong recommendation for YOUR situation |
| Explaining technical concepts (CAP, PACELC, consistency models) | Knowing your team's skills, budget, or existing stack |
| Generating comparison matrices quickly | Providing accurate, up-to-date pricing information |
| Summarizing documentation and prior art | Understanding organizational politics and constraints |
| Drafting ADR templates with the right structure | Estimating real-world performance under your specific load |
| Identifying tradeoff dimensions you hadn't considered | Predicting which tradeoff will matter most for your product |
| Explaining what other companies have done | Knowing if that approach applies to your scale |
Effective prompts for tradeoff analysis
Comparing technologies
Bad prompt:
Should I use PostgreSQL or DynamoDB?This gets you a generic pros/cons list. Not useful.
Good prompt:
I'm building an e-commerce platform. Expected scale: 100K users in year 1,
potentially 2M by year 3. My team of 4 developers all have SQL experience,
none have NoSQL experience. Our access patterns are:
- Product catalog: read-heavy, complex queries (search, filter, sort)
- Order processing: write-heavy, needs ACID transactions
- User sessions: simple key-value lookups, high throughput
Compare PostgreSQL and DynamoDB for this use case. For each access pattern,
tell me which database is better suited and why. Be opinionated - I want a
recommendation, not a balanced summary.The key additions: scale, team expertise, specific access patterns, and the explicit instruction to be opinionated.
Stress-testing a design decision
I've decided to use Redis as our primary session store for a web
application expecting 50K concurrent users. Our Redis instance is
a single node (not clustered).
Challenge this decision. What are the failure modes? At what scale
does this break? What would you recommend instead and why?
Give me the worst-case scenarios, not just the happy path.This prompt explicitly asks AI to argue against your decision. This is valuable because AI's natural tendency to be agreeable works against you, it will validate whatever you suggest unless you tell it to push back.
Generating an ADRWhat is adr?Architecture Decision Record - a short document capturing one technical choice, why it was made, and what tradeoffs were accepted. draft
Write an Architecture Decision Record for the following:
Context: We're choosing between building a custom notification system
vs using a managed service (OneSignal, Firebase Cloud Messaging, or
Amazon SNS). We need push notifications, email, and SMS. Volume:
~500K notifications per day. Team size: 6 developers. Timeline:
product launch in 3 months.
Include at least 3 alternatives with honest pros/cons.
In the Consequences section, list what we're giving up with each option.AI is excellent at producing well-structured ADR drafts. The structure will be solid; your job is to fill in the context-specific details and make the actual decision.
What to verify
AI-generated tradeoff analysis frequently gets these things wrong:
Cost estimates: AI has outdated pricing and often confuses pricing tiers. DynamoDB pricing, in particular, is complex (on-demand vs provisioned, read vs write capacity units, data transfer costs) and AI regularly gets it wrong. Always check the vendor's current pricing page.
Performance claims: "PostgreSQL can handle millions of rows easily" is true but useless. Can it handle YOUR query pattern on YOUR hardware with YOUR indexes? Run your own benchmarks. AI cannot predict your p99 latencyWhat is latency?The time delay between sending a request and receiving the first byte of the response, usually measured in milliseconds..
Scaling limits: AI might say "Redis can handle 100K operations per second" which is technically true for a well-configured cluster but meaningless if you are running a single t3.micro instance. Always check the specific configuration.
Recency: AI training data has a cutoff. A new database version might have fixed the limitation AI warns you about. A managed service might have added the feature AI says it lacks. Check changelogs and release notes.
Your team's reality: AI does not know that your DBA quit last month, or that your company has an enterprise agreement with AWS that makes DynamoDB half price, or that your compliance team requires all data to stay in the EU. You must inject this context.
The hybrid workflow
Here is how to effectively combine AI with human judgment for tradeoff analysis:
- Research phase (AI-heavy): Ask AI to list options, explain tradeoffs, and generate comparison matrices. Ask it to challenge your assumptions. Use it to draft an ADRWhat is adr?Architecture Decision Record - a short document capturing one technical choice, why it was made, and what tradeoffs were accepted.. This should take 1-2 hours instead of 1-2 days.
- Validation phase (human-heavy): Verify cost estimates with actual pricing pages. Run a spike to test the critical unknown. Check with your team about their experience levels. Review compliance and organizational constraints.
- Decision phase (human only): Apply the DACI framework. The Approver makes the call based on the research and spike results. AI does not get a vote here, it does not understand your priorities.
- Documentation phase (AI-assisted): Use AI to polish the ADR draft. Have it check that you've addressed alternatives and consequences. But make sure the final document reflects your actual reasoning, not AI's generic analysis.
The pattern is: let AI do the breadth (exploring many options quickly) while you do the depth (validating the critical details for your specific situation). AI is your research assistant, not your architect.