Agent skill · personamanagmentlayer
llm-engineering-expert
Build reliable applications on large language models: prompt design, structured output, evaluation, guardrails, and cost and latency control. Use when the user mentions LLMs, prompts, prompt engineering, few-shot examples, structured or JSON output, function calling, hallucination, model evaluation, token costs, streaming, or when the task involves choosing a model, writing a system prompt, or making model output dependable enough for production.
What it needs
About 6k tokens when loaded.
What this skill does
LLM Engineering Expert Building on a language model is engineering against a component that is non-deterministic, occasionally confident and wrong, and priced per token. The discipline is in the surrounding structure, not in the wording of prompts. Core Concepts The Model Is a Component, Not the System A production LLM feature is a pipeline: input validation, context assembly, the model call, output parsing and validation, then a decision about what to do when any of that fails. Treating the model call as the whole feature is the root of most reliability problems. Determinism Is Not Available Even at temperature zero, output can vary across model versions, infrastructure and batching. Design for a distribution of outputs, not a fixed one: validate what comes back, and make the failure path as considered as the success path. Context Is a Budget Everything competes for the same window: system instructions, examples, retrieved documents, conversation history, the user's input, and room for the answer. Attention is not uniform across a long context — material at the beginning and end is used more reliably than material buried in the middle. Spend the budget deliberately. Evaluation Precedes Iteration Without a scored test set, prompt changes are superstition. Ten to fifty representative cases with expected properties are enough to start and will catch most regressions. Prompt Design Structure that holds up Order matters: instructions first, reference material next, the specific request last. Ending with the request keeps it close to generation. What earns its place: an explicit role, hard constraints stated as rules, a defined behaviour for the out-of-scope case, and delimiters that separate untrusted content from instructions. What does not: politeness, threats, "you are the world's best", and repeated emphasis. They consume budget and do not measurably improve output. Few-shot examples Two to five examples usually beat both zero and twenty. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills personamanagmentlayer/llm-engineering-expert