Skip to content

Prompt engineering as a discipline

  • Agents
  • Asked in most interview loops
  • About 1 h 30 min to learn
  • 6 vetted resources
Build a plan that includes this

What this topic covers

Prompt engineering is the cheapest lever you have over a model: the system prompt sets the rules and persona, the user turn carries the task, and examples, step-by-step reasoning and chained sub-prompts shape the output without touching weights. Interviewers rarely ask for clever wording; they ask when prompting is enough versus when you need retrieval (missing knowledge) or fine-tuning (missing behaviour), how you keep prompts in version control and regression-test them against a fixed eval set, and how you defend a prompt against injection and leaking. Expect to be handed a vague task and asked to draft a prompt out loud, then justify every line.

Builds on

Leads to

You’re ready when you can

What interviewers check, from the resources below.

  • Explain what belongs in a system prompt versus the user turn, and why
  • Rewrite a vague instruction so it states the goal, the audience and the format explicitly
  • Describe when adding examples helps and when they over-anchor the model
  • Explain why you would split a task into chained prompts instead of one long prompt
  • Compare asking the model to reason step by step with enabling extended thinking
  • Explain the difference between zero-shot, few-shot and chain-of-thought prompting with one example each
  • State a task where few-shot examples are not enough and what you would do instead
  • Describe prompt injection and prompt leaking and give two defence tactics for each
  • Explain why none of the listed defences is sufficient on its own
  • Explain the developer/system, user and assistant role hierarchy and which wins on conflict

What to study, in order

Each link opened and checked by hand; the date says when.

  1. Start here

    Prompting best practices (opens in a new tab)

    Claude Docs · Anthropic · 40 min · Free · checked 21 Oct 2026

    Why
    This is the vendor's own reference for every technique an interviewer will probe: clarity, examples, roles, thinking, format control and chaining.
    What to read
    Read “General principles” (“Be clear and direct” through “Give Claude a role”), then “Control the format of responses”, “Leverage thinking & interleaved thinking capabilities” and “Chain complex prompts” under “Agentic systems”.
    What to skip
    “Model-specific guidance”, “Migration considerations” and the “Capability-specific tips” (vision, frontend, dosing) unless you are targeting a Claude-only shop.

  2. Free alternative

    Prompt Engineering Guide (opens in a new tab)

    Prompt Engineering Guide · DAIR.AI · 35 min · Free · checked 21 Oct 2026

    Why
    Vendor-neutral and paper-backed, so you can name the technique and the result it came from when an interviewer asks why it works.
    What to read
    Under “Prompting Techniques” read “Few-Shot Prompting” (including “Limitations of Few-shot Prompting”), “Chain-of-Thought (CoT) Prompting” (through “Zero-shot COT Prompting”) and “Prompt Chaining”. Then under “Risks & Misuses” read “Adversarial Prompting in LLMs” down to the end of “Defense Tactics”.
    What to skip
    “Automatic Chain-of-Thought (Auto-CoT)”, the “Prompt Hub”, “Models” and “Papers” sections, and the jailbreak examples beyond knowing the names.

  3. Go deeper

    Prompt engineering (opens in a new tab)

    OpenAI Developers · OpenAI · 20 min · Free · checked 21 Oct 2026

    Why
    Covers the operational side interviewers add after the basics: prompts as versioned code, role hierarchy, caching and how reasoning models want to be prompted differently.
    What to read
    Read “Message roles and instruction following”, “Version prompts in code”, “Few-shot learning”, “Include relevant context information”, “Planning for the context window” and “Prompting reasoning models”.
    What to skip
    “Choosing a model”, the Markdown/XML formatting notes and the coding, front-end and agentic sections, which are model-specific.

  4. Go deeper

    Language Models are Few-Shot Learners (opens in a new tab)

    arXiv · Tom B. Brown et al. · 45 min · Free · checked 22 Oct 2026

    Why
    The GPT-3 paper defined zero-shot, one-shot and few-shot prompting, and “why does in-context learning work without gradient updates?” is still a standard opener.
    What to read
    Read §1 with Figures 1.1–1.3 (in-context learning improves with scale), §2 opening and Figure 2.1 (zero, one and few-shot versus fine-tuning), and §2.1 with Table 2.1 on model sizes. Then read §4 on benchmark contamination and §5 on limitations.
    What to skip
    Skip most of §3 (per-benchmark results); skim §3.9 for the synthetic tasks if you have time. Skip §6 broader impacts and the appendices.

  5. Go deeper

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (opens in a new tab)

    arXiv · Jason Wei et al. · 30 min · Free · checked 22 Oct 2026

    Why
    Chain-of-thought is the root of today’s reasoning models, and interviewers expect you to know what it is, when it helps and why it only shows up at scale.
    What to read
    Read §2 with Figure 1 (the prompt format), §3.2 on results (gains appear only in large models) and §3.3, the ablation that rules out equation-only and extra-tokens explanations. Then read Appendix A.1–A.3 on scale, prompt sensitivity and when to expect it to help.
    What to skip
    Skip §4 and §5 beyond the headline figures, §7 related work and Appendices C–E.

  6. Go deeper

    DSPy: Programming—not prompting—Foundation Models (opens in a new tab)

    GitHub · stanfordnlp · 60 min · Free · checked 22 Oct 2026

    Why
    DSPy reframes prompt engineering as programs plus optimizers that tune prompts against a metric. Interviewers bring it up when they ask how you’d stop hand-tuning prompts and make prompt changes measurable.
    What to read
    Work through the RAG tutorial on dspy.ai (docs linked from the README): “Exploring some basic DSPy Modules”, “Evaluation in DSPy” (dspy.Evaluate with SemanticF1), “Build your first RAG Module” and “Using a DSPy Optimizer to improve your RAG prompt” (MIPROv2). In the repo, read dspy/signatures/signature.py, dspy/predict/chain_of_thought.py and dspy/teleprompt/bootstrap.py (BootstrapFewShot, the simplest optimizer); glance at mipro_optimizer_v2.py and gepa/ to see what the stronger optimizers search over.
    What to skip
    Skip the tutorial’s MLflow integration and saving/loading sections, the repo’s adapters, retrievers and integrations folders, and fine-tuning optimizers like BootstrapFinetune on a first pass.

Companies that test Agents most

Put Prompt engineering as a discipline in your plan

Tell us your interview date and hours a day; it lands on the right day, in the right order.

Build my plan