15 sample prompts from The Senior Engineer's AI Prompt Vault — production-tested prompts for Claude Code, Cursor, and other AI coding assistants. Each one encodes how a staff engineer actually uses AI: forcing structured thinking over pattern-matched answers, naming the failure modes that matter, and making the model commit instead of hedge.
The full Vault — all 50 prompts, 5 .cursorrules files, and 3 CLAUDE.md starters (Go, Elixir, AI-agent projects) — is free on sublimecoding.com.
- Blast Radius Analysis — Before merging a PR that touches shared infrastructure.
- Security-First Review — Before merging anything that touches auth, input handling, or external IO.
- Database Migration Safety Review — Reviewing any DDL change against a live, large production database.
- Hypothesis-Driven Debugging — When you've been staring at a bug for 30 minutes and need a reset.
- Reproduce a Race Condition Reliably — When the bug appears once a week in production but never on your laptop.
- Production Incident: First 10 Minutes — Page just fired. You're alone. What now.
- Decompose a Long Function the Right Way — 200-line function that does sequential things.
- Make Illegal States Unrepresentable — When a struct has 4 optional fields and only some combos are valid.
- Untangle a God Class — 1500-line UserService that does everything.
- Design Idempotent APIs — Any API that mutates state and might be retried (i.e. all of them).
- Pick the Right Queue — When you need async processing and have to choose between Kafka, SQS, Pub/Sub, RabbitMQ, or a DB-backed queue.
- Design a Rate Limiter — Adding rate limits to a public or partner API.
- PR Description from Diff — Same as code-review-10, used for self-authoring not review.
- Architecture Decision Record — When a non-trivial decision deserves a written record.
- Postmortem Draft from Incident Notes — Writing the postmortem the day after, while context is fresh.
Drop the prompt into your chat or system prompt. Replace the <paste here> marker with your diff, function, endpoint, or context. Most are tuned to refuse hedging — if the model gives you a wishy-washy answer, paste the prompt again with "Pick one and commit." appended.
These prompts assume you've done the basics: written the failing test, run the linter, read the error twice. They're for the part where the work is actually hard.