Claude Certified Engineers logoCCE

CCAR-P · Study Guide

← Domain 2: Claude Models, Prompting & Context Engineering

2.3 · Lesson 3 of 5

Apply prompt engineering techniques (zero-shot, few-shot, chain-of-thought)

What You Need to Know

Prompt engineering techniques are tools matched to failure modes. Zero-shot works when the task is clear and well-specified — a sharp rubric often beats ceremony. Few-shot helps when format consistency or edge-case behavior matters: a few strong exemplars teach structure better than more adjectives. Chain-of-thought (or explicit scratchpads) helps multi-step reasoning when the extra tokens improve measured accuracy enough to justify latency and cost.

The architect’s rule: pick the lightest technique that fixes the named failure. Do not default to CoT on latency-sensitive simple tasks. Do not skip few-shot when the bug is format drift. Technique shopping without evals is theater.

Technique map

  • Zero-shot — clear task, clear rubric, no format pain yet
  • Few-shot — schema/format drift or edges need demonstration
  • Chain-of-thought — multi-step reasoning failures with measured gain
  • Always: name failure mode → try lightest fix → measure
  • Hot path: respect latency SLAs when adding reasoning tokens

Decision rules

  • Zero-shot for clear, well-specified tasks.
  • Few-shot when format or edge cases are the failure mode.
  • CoT when multi-step reasoning earns its tokens on evals.
  • Lightest technique that fixes the failure — not the heaviest by default.
  • Do not default CoT on simple latency-sensitive paths.

Earning tokens

Every technique spends context and often latency. Few-shot examples and CoT traces are not free. Report Δ quality against Δ latency/cost before promoting a heavier technique to the production hot path. A hard-slice-only CoT route can be smarter than CoT everywhere.

Review checklist

  • Is the failure mode named?
  • Was zero-shot with a sharp rubric tried first when appropriate?
  • Do few-shot examples teach format/edges without contradiction?
  • Is CoT justified with measured accuracy gain and SLA fit?
  • Is the hot path still on the lightest clearing technique?

Stems that push CoT “for quality” on a simple classifier reward refusing unnecessary reasoning tokens.

Exam application

Match technique to failure mode and prefer the lightest fix. Distractors: always-CoT, few-shot for unrelated problems, deleting rubrics, or raising temperature instead of structure.

Exam traps

  • Default CoT on simple, latency-sensitive tasks

    Extra reasoning tokens cost latency and money. Exam stems reward lightest technique that fixes the failure.

  • Zero-shot when format or edge cases are the real bug

    Clear instructions without examples often still drift on schema. Few-shot examples of the desired format fix format failure modes.

  • Few-shot examples that teach the wrong pattern

    Misleading or contradictory exemplars poison behavior. Choose examples that cover edges you care about.

  • Technique shopping without naming the failure mode

    Pick zero-shot, few-shot, or CoT because of a measured problem — not because a blog listed them.

Practice scenario

A classifier must map tickets to five labels using a clear rubric. Latency is tight. The team proposes chain-of-thought on every request “for quality.” What is the best first move?

Choose one answer

Build exercise

Choose PE techniques for a latency-sensitive ticket classifier

35 minutes

What you'll learn

  • Name the failure mode before picking a technique
  • Prove zero-shot with a sharp rubric when possible
  • Add few-shot for format and edges only
  • Justify CoT with measured gain vs latency

  1. Step 1

    Name the failure mode

    For a ticket classifier, record whether failures are unclear task specification, format drift, edge-case confusion, or multi-step reasoning errors.

    Why: Technique choice follows the failure — not habit.

    You should see: A short failure-mode card with one primary label.

  2. Step 2

    Try zero-shot with a sharp rubric

    Write a precise task + acceptance criteria and run the gold set before adding examples or reasoning scaffolding.

    Why: Many well-specified tasks clear the bar without extra tokens.

    You should see: Zero-shot prompt + eval score on the gold set.

  3. Step 3

    Add few-shot only for format or edges

    If format or rare edges fail, add a few high-quality examples that show the schema and those edges — not a giant noisy set.

    Why: Few-shot earns tokens when examples teach structure or edges the model misses.

    You should see: 3–5 exemplars covering happy path + two hard edges.

  4. Step 4

    Justify CoT with measured gain

    Enable chain-of-thought (or scratchpad) only on the subset where multi-step reasoning fails and measure Δ accuracy vs Δ latency/cost.

    Why: CoT must earn its tokens against an SLA — especially on hot paths.

    You should see: A/B note: CoT vs no-CoT on the hard slice with p95 impact.

Sources