Over-prompting. The 800-word system prompt with 20 numbered rules and 14 edge cases inline. Models follow the first few instructions reliably and degrade afterward. Symptoms: you add rule #21 to fix a bug, rule #4 starts being violated. The fix is structural: decompose into multiple smaller calls (classify intent → route → handle), move edge cases into few-shot examples (show, don't tell), enforce structure via schema rather than prose.

No eval. Shipping prompt changes based on vibes and three cherry-picked examples. This is the single biggest failure mode in production GenAI and it's exactly what the FDE role exists to fix. You need: a frozen golden set (~30–100 representative inputs with known-good outputs), automated scoring (exact match where possible, Gemini-as-judge where not — with the judge prompt itself versioned), and a regression suite that runs on every prompt change. Without this, you have no idea if v23 of your prompt is better than v22 or just different. This is the muscle the recruiter doc calls "Operational Excellence" applied to prompts.

No versioning. Prompts edited in place in a config file or — worse — in the Vertex AI Studio UI, with no git history, no diff, no rollback. When quality drops on Tuesday you have no idea what changed. Treat prompts as code: in the repo, code-reviewed, with the eval suite attached as a CI check. Tag prompt versions; log which version served each request so you can correlate quality drops to prompt deploys.

Treating prompts as the whole solution. Some problems are not prompt problems. If the model doesn't know the customer's product catalog, no prompt fixes that — you need RAG. If the model needs to take actions in a live system, you need function calling/tools. If a task is high-volume and stable, fine-tuning a smaller model may be cheaper than prompting a frontier model. Senior judgment is knowing which layer the problem lives at: prompt, retrieval, tools, fine-tune, or "this isn't an LLM problem." An FDE who keeps iterating on the prompt when retrieval is broken is a junior FDE.