AI-driven development
Build reliable AI features and use coding assistants with clear requirements, safe tools, and verifiable results.
Showing 1–12 of 20 articles
Debug an AI feature that works locally but fails in production
Debug production-only AI failures by separating application, runtime, network, provider, and configuration boundaries with a safe request-level evidence bundle.
Decide when your AI feature needs human approval
Place human approval before consequential AI actions and bind it to an exact reviewable proposal, with permission checks, expiration, and safe handling of changed state.
Design agent tools with narrow permissions and clear contracts
Design agent tools as narrow server-enforced capabilities with validated arguments, explicit ownership, bounded results, and reviewable action contracts.
Reduce AI feature costs before adding another cache
Reduce AI feature spending by measuring request volume, context size, retries, and quality before choosing prompt changes, model routing, or a carefully scoped cache.
Handle slow and failed AI responses in a normal web application
Model slow AI requests as explicit application states with bounded timeouts, cancellation, request ownership, and recovery that preserves the user's work.
Build a document Q and A feature with verifiable citations
Build document Q&A around permission-aware retrieval and claim-level evidence so citations can be checked and unsupported questions produce a useful fallback.
Stop a retrieved document from becoming an instruction to your agent
Keep retrieved documents from controlling an AI agent by separating content from authority, enforcing server-side permissions, and testing malicious instruction-shaped text.
Return structured AI output without trusting its JSON
Validate AI generated JSON at both the shape and business-rule boundaries, then handle invalid, unsupported, and incomplete output without silently trusting it.
Build a small evaluation set before changing your AI prompt
Create a small representative evaluation set before editing an AI prompt so improvements, regressions, refusals, and ambiguous cases become visible and repeatable.
Choose between a normal function and an LLM for a product feature
Choose deterministic code or an LLM by separating exact product rules from language interpretation, then compare uncertainty, verification, latency, and operating cost.
Build an evidence checklist for AI assisted changes
Build a change evidence checklist that distinguishes observed behavior, executed checks, skipped boundaries, and remaining risk in AI assisted engineering work.
Detect when an AI assistant invented an API or dependency
Detect invented APIs in AI suggestions by checking installed package versions, exports, type declarations, and a minimal executable example before integrating new code.
Explore topics
Subscribe to Newsletter
Get the latest articles and updates delivered to your inbox.