Most prompt engineering guides cover the same basics: be specific, give examples, use system prompts. Here are the techniques beyond that level — the ones that produce meaningfully better results on difficult tasks.
Chain of Thought and Its Limits
Chain-of-thought prompting (“think step by step” or “let’s reason through this”) measurably improves performance on reasoning-heavy tasks: mathematics, logic problems, multi-step analysis. The mechanism: it pushes the model to produce intermediate reasoning tokens rather than jumping to an answer, which improves accuracy on problems where the answer is difficult to compute directly. The limits: chain of thought does not reliably help on tasks that require memory, factual recall, or creative generation — and it increases output length and therefore cost. For simple factual questions or straightforward generation tasks, chain of thought adds overhead without benefit. Use it specifically for analysis, reasoning, and step-by-step problem-solving tasks.
Constitutional and Role-Based Framing
Framing the model’s role and operating constraints in the system prompt substantially changes output quality for specific domains: “You are a senior software engineer reviewing code for production deployment. Your primary concerns are security vulnerabilities, performance issues, and maintainability. You identify problems directly without qualification and suggest specific fixes.” This type of framing works because it activates a consistent persona with implicit priorities — the model interprets ambiguous situations through the lens of the stated role. The difference between a generic coding assistant and a “senior production engineer” persona is measurable in the specificity and technical depth of the output. The technique extends to tone, format, and domain expertise.
Few-Shot Examples for Format Control
When you need precise output format — structured data, specific style, particular tone — 2–4 examples in the prompt are more reliable than detailed instructions. Format instructions are interpreted; examples are followed. The pattern: “Here are examples of the analysis I want: [example 1] [example 2]. Now apply this to: [actual task].” For JSON output, one complete valid example of the expected schema is more reliable than a schema description. For writing style, two paragraphs in the target style are more reliable than an adjective list. The caveat: long few-shot examples consume context and increase cost — use them when format precision matters and instructions alone prove insufficient.
Self-Consistency Sampling
For high-stakes one-off decisions (code architecture choices, analysis of complex situations), run the same prompt 3–5 times with temperature > 0 and compare the results. If all outputs converge on the same answer, confidence is higher. If they diverge substantially, the question may not have a clear single answer — the divergence itself is informative. This is computationally expensive (3–5x cost), so it is only warranted for genuinely high-stakes decisions where the cost of being wrong is significant. Automated self-consistency (majority voting across samples) is used in research and some production AI systems for this reason.
XML/Structured Tags for Long Contexts
Anthropic’s Claude models respond particularly well to XML-style structured prompts for complex tasks with multiple inputs. Wrapping different input types in descriptive tags (<document>, <instructions>, <example>, <user_query>) helps the model correctly attribute and weight different parts of a long prompt. For prompts with >2,000 tokens of input, structured tagging measurably reduces confusion between instructions and content. The technique is less critical for shorter prompts but becomes valuable as prompt length increases. Include the tags even for single-item inputs when the prompt is long.
