Hey,
You're telling Claude you're a junior developer. That's why it gives you junior answers.
Most people never set an experience level. They open a chat, type their question, and get back something competent but generic. Safe explanations. Broad recommendations. The kind of advice you'd give someone you just met.
Set your experience level to 20 years and the same model produces structurally different output. More nuanced trade-offs. Deeper technical detail. Fewer caveats hedging every sentence. It stops explaining basics you already know and starts giving you the kind of answer a senior peer would give at a whiteboard.
Where to set it depends on your setup:
System prompt: Most direct. Works with any Claude integration.
Inline context: Add it to the top of your prompt when you don't control the system prompt.
CLAUDE.md: Claude Code users can drop it in their project config. It persists across every session.
Same prompt: "How should I structure error handling in a REST API?"
Default (no experience set):
You should use try-catch blocks and return appropriate HTTP status codes. Consider using a middleware layer for consistent error responses. Make sure to log errors for debugging purposes.
20 years experience set:
Layer it: domain errors at the service boundary (mapped to status codes via a shared enum), infrastructure exceptions caught at the middleware level (never leak stack traces), and a dead-letter queue for anything unhandled. Don't abstract error handling into a generic handler class too early. Start with explicit handling per route, then extract the pattern after the third repetition. The mistake most teams make is wrapping everything in a catch-all that swallows context.
Same model. Same cost. Different output because you told it who it's talking to.
Add this to your Claude Code CLAUDE.md or your system prompt:
# Experience calibration
You are assisting someone with 20+ years of professional experience
in software engineering. Skip foundational explanations. Focus on
trade-offs, edge cases, and production-ready patterns. Assume the
user understands basic concepts and wants expert-level guidance.Takes 30 seconds to add. Changes every response that follows.
June 4, 60 minutes, free. We're going beyond prompting into actual time architecture. How to use AI to multiply the hours in your week, not just speed up individual tasks.
Before you write a single line of code, hit /plan in Claude Code. It reads your codebase, maps dependencies, and proposes an implementation plan. Review it, adjust it, approve it. Then it executes. Building without planning is the coding equivalent of driving blindfolded.
Until next week,
Merryl