Reference
The AI engineering skill glossary
Hiring managers describe the same work with different words. “Harness engineering,” “context engineering,” “agentic orchestration,” “forward-deployed AI” — they point at a small set of disciplines. Every alias on this page rolls up to one of the 8 core categories used across the index.
8 core categories
Category 01
Prompting
Treat prompts as a specification: inputs, refusals, and a regression eval.
Prompt design, iteration, and instruction-following for a specific job. At production scale this is versioned text, eval harnesses, and model-agnostic benchmarks — not a clever one-off instruction.
System prompt design
Writing and iterating the durable instructions that define a task: role, tools, refusal cases, and output shape.
Also called
Prompt evaluation
A/B testing variants, golden-set regression, and catching drift when a model or tool changes.
Also called
Structured output
Forcing reliable JSON, tool calls, or schema-bound answers instead of free-form prose.
Also called
Category 02
Retrieval / RAG
Index, chunk, rerank, and ground answers in a corpus — with a metric, not a vibe.
Retrieval-augmented generation: how documents become embeddings, how queries find the right passages, and how the model is required to cite what it used. Most demos fail here in production.
Hybrid retrieval
Combining lexical search (BM25) with dense embeddings and a reranker instead of trusting a single vector store.
Also called
Chunking and indexing
Deciding what a “document” is, how it is split, what metadata is stored, and when the index is rebuilt.
Also called
Retrieval evaluation
Precision@k, recall, and golden-question sets that prove the retriever — not the chat UI — is doing the work.
Also called
Category 03
Context management
What goes in the window, what stays out, and how state is carried across turns.
Context engineering is the discipline of packing, trimming, and remembering: conversation state, retrieved passages, tool results, and the budget of a finite window.
Window budgeting
Choosing what survives when tokens are scarce: summaries, citations, tool traces, and the current user turn.
Also called
State and memory
Carrying task state across steps without dumping the entire history back into the next prompt.
Also called
Harness engineering
The surrounding loop that feeds the model: files, tools, retrieval, and the policy for when to stop.
Also called
Category 04
Agents
Tool use, multi-step orchestration, and recovery when a step fails.
Agent engineering is planning, tool calling, multi-agent coordination, and the boring parts: budgets, retries, and an audit log. Autonomous does not mean untraceable.
Tool use and planning
Selecting tools, sequencing steps, and stopping when the budget or the policy says so.
Also called
Multi-agent orchestration
Splitting work across specialists, passing state, and failing closed when a worker hallucinates a result.
Also called
Tracing and recovery
Every step leaves a trace. Failed tools retry or abort; they do not silently invent an answer.
Also called
Category 05
Evaluation
Evals, traces, and a quality bar that is not a vibe check.
Designing the tests that decide whether a change ships: golden sets, graders, online metrics, and regression suites that run when a prompt, model, or retriever moves.
Offline eval suites
Fixed datasets and graders that catch regressions before a prompt or model reaches users.
Also called
Online quality
Traces, thumbs, and sampled reviews that show whether production quality is holding.
Also called
Task-specific graders
Rubrics and model-as-judge setups that measure the job you actually have, not a generic leaderboard.
Also called
Category 06
Guardrails
Policy, refusal, red-teaming, and the cases the model must not invent.
Safety and hardening: input classification, output validation, jailbreak resistance, and the refusals a domain requires — healthcare dosage, legal advice, customer PII.
Input and output filters
Classifying prompts and validating completions before they reach a user or a downstream system.
Also called
Red-teaming
Adversarial tests that try to bypass the policy, then become part of the eval suite.
Also called
Policy and compliance
Encoding what the system must never say or do, and proving it in an audit.
Also called
Category 07
Production hardening
Latency, cost, observability, and the path from a notebook to a service.
LLMOps for systems that stay up: inference routing, caching, timeouts, cost controls, and the dashboards that show when quality or spend slipped overnight.
Inference and serving
Choosing hosts, batching, and fallbacks so a model call is a reliable dependency.
Also called
Cost and latency
Caching, model routing, and budgets so the feature does not become an unbounded bill.
Also called
Observability
Traces, metrics, and alerts for prompts, tools, and retrieval — not just HTTP 200s.
Also called
Category 08
Fine-tuning
Data pipelines, adapters, and knowing when training is the wrong move.
Domain adaptation of foundation models: datasets, LoRA/QLoRA, evals against the base model, and the judgment to stay with prompting or RAG when training will not pay off.
Adapter training
PEFT methods that specialize a base model without a full pretrain.
Also called
Data pipelines
Collecting, cleaning, and versioning the examples a tune actually learns from.
Also called
Train-or-not judgment
Comparing a tuned model to a well-prompted or RAG-backed baseline before spending the GPU budget.
Also called
Use this taxonomy to find talent
Every engineer profile is classified with these categories — no buzzword bingo.