Quick answer: Ten patterns, ordered by how often they show up in production: role-priming, few-shot, structured output, delimiter discipline, self-critique, decomposition, tool-call schemas, retrieval-augmented reasoning, long context with pointers, and guardrails. Memorise these — they move the needle.

1. Role-priming

Start with a role: “You are a patent examiner with 20 years' experience in mechanical engineering.” Improves tone, depth, and reduces hedging. Overused in 2022, still useful in 2026 for specialised tasks.

2. Few-shot examples

Two to five input/output pairs. The biggest single win across the last five years of LLM research. Put examples in the order you want the model to weight them (recency bias is real).

3. Structured output

“Respond only with a JSON object matching this schema: { name: string, score: number }.” Use JSON schema / tool-call schemas for production. Parser errors drop by an order of magnitude.

4. Delimiter discipline

Wrap user-provided content in XML tags: <user_input>...</user_input>. Makes prompts clearer, defends against prompt injection, and helps the model separate instructions from data.

5. Self-critique

After generating, ask the model to critique its own output against a rubric, then revise. A two-pass self-critique loop adds 5–15% accuracy on complex tasks (Anthropic and Meta research, 2024).

6. Decomposition

Break a hard task into steps. Solve each step, feed the result into the next. Similar to human sub-goal reasoning. Works especially well for reasoning models — each step gets its own thinking budget.

7. Tool-call schemas

Define tools precisely: name, description, parameter schema (JSON schema), when to call. Most 2026 LLM failures in agent loops are bad tool schemas, not bad reasoning. Spend time here.

8. Retrieval-augmented reasoning

Retrieve relevant context, inject into the prompt with explicit provenance (<doc src="policy.pdf p.12">...</doc>), ask the model to cite sources. Reduces hallucination dramatically.

9. Long context with pointers

With 200k–1M-token context windows, you can paste entire codebases or document corpora. But models still weight early and late tokens more than middle. For critical facts, repeat them at the start and end, or add a short “focus on” section pointing the model at specific sub-regions.

10. Guardrails

System-level instructions that cannot be overridden, explicit refusal patterns for out-of-scope queries, and a classifier on top of user input for red-team surfaces. In 2026 prompt-injection attacks are a real security concern; guardrails matter.

Practice these on GeraQuest

Each of these ten patterns maps to a GeraQuest challenge level. You can drill them in real time and see your prompt's output scored.

Ten Prompt-Engineering Patterns That Actually Work in 2026