AI Prompts

Plan a library or runtime upgrade with changelog risk, lockfile strategy, CI gates, and a rollback that does not leave half the fleet on a broken version.

Threat-model a service for authz gaps, tenancy leaks, injection, and abuse paths, then rank mitigations that engineering can ship before launch.

Investigate a latency or throughput regression with profiles, deploys, and query plans, then recommend a safe fix instead of a speculative rewrite.

Draft an architecture decision record with real options, consequences, and reversal cost so the team can commit without burying disagreement in chat.

Audit a retrieval-augmented generation stack for recall, ranking, chunking, and silent misses so generator quality is not blamed for a retrieval failure.

Design an offline and CI eval harness for an LLM feature with task metrics, slice bugs, and a promotion gate that prompt diffs cannot skip.

Set cost, latency, and error budgets for an LLM gateway, then design routing, caching, and fallback that stay inside those numbers at peak.

Review a tool-calling agent for loops, bad arguments, auth mistakes, and irreversible side effects, then add stops the runtime can actually enforce.

Plan groundedness and hallucination tests so answers that lack sources fail the eval, including abstain cases and citation checks you can run in CI.

Design input, tool, and output guardrails that the runtime can enforce, with allowlists and audits, instead of hoping the system prompt will refuse.

Write a decision memo among prompting, RAG, and fine-tuning with data needs, ops cost, and a reversal path instead of a default to fine-tune.

Build a labeled eval dataset from production traces and slice quotas so prompt changes are judged on real failures, not toy questions.