Some problems the basic prompt — role, context, task, format — does not solve well. Complex multi-step reasoning. Tasks where a specific style or format is critical. Situations where you need the model to adopt a very precise perspective. For these cases, there are techniques that go beyond the simple prompt and produce substantial, documented improvements in output quality.

When the basic prompt is not enough

A well-constructed basic prompt is enough for most tasks: writing, summarising, classification, brainstorming. But when the task requires chained reasoning, the model tends to skip steps and reach incorrect conclusions. When the format or style is very specific and hard to describe in words, verbal description rarely produces the desired result. When you need a very particular expert perspective, a generic role does not activate the right patterns.

In those cases, these three techniques make the difference.

Chain-of-thought: think before answering

Chain-of-thought prompting consists of asking the model to show its reasoning step by step before reaching a conclusion.

Models make more reasoning errors when they produce the answer directly than when asked to reason explicitly. Showing the intermediate process reduces errors because the model “checks” its own reasoning as it generates it.

Without chain-of-thought:

Q: If I have 3 boxes with 12 bottles each and sell 
   2/3 of the total, how many bottles remain?
A: 12

With chain-of-thought:

Q: Solve this problem step by step, showing each 
   operation before giving the final answer.
   If I have 3 boxes with 12 bottles each...
A: Step 1: Total bottles = 3 × 12 = 36
   Step 2: Bottles sold = 36 × (2/3) = 24
   Step 3: Bottles remaining = 36 - 24 = 12
   Answer: 12 bottles remain.

The difference is greater in multi-step problems, logical reasoning, and problems where an initial error propagates. For simple tasks, chain-of-thought adds verbosity without benefit.

Useful variants: “Think out loud before answering,” “Explain your reasoning,” “What would the pros and cons be before recommending something?”, “Identify the assumptions first before reaching a conclusion.”

Few-shot: learn from examples

Few-shot prompting consists of including one or several examples of the type of response you want in the prompt, before the actual question. Instead of verbally describing the desired format or style — which is often imprecise — you show it directly.

Without few-shot (zero-shot):

Classify these reviews as positive, negative or neutral.
Review: "The product arrived late but works well."

With few-shot:

Classify these reviews as POSITIVE, NEGATIVE or NEUTRAL.

Example 1:
Review: "Fast delivery and product exactly as described."
Classification: POSITIVE

Example 2:
Review: "Terrible quality, broke on first use."
Classification: NEGATIVE

Example 3:
Review: "Works, nothing special."
Classification: NEUTRAL

Now classify this:
Review: "The product arrived late but works well."

Few-shot is especially useful when:

  • The output format is very specific and hard to describe
  • The style or tone must be consistent with examples you have
  • The task has ambiguities that examples resolve better than words
  • You need consistent results at scale (processing many inputs of the same type)

The optimal number of examples varies by model and task. Generally, 2 to 5 examples is sufficient. More examples do not always improve the result and may consume valuable context.

Role prompting: experts on demand

Role prompting is more than adding “you are an expert in X” at the start of the prompt. When done well, it activates specific representations from training that produce qualitatively different responses.

The difference between basic and advanced role prompting is in the level of detail of the role description:

Basic:

You are a marketing expert. Review this text.

Advanced:

You are a marketing director with 15 years of experience 
in B2B software startups. You have worked primarily in 
English-speaking markets but know the European market well. 
Your approach is data-driven: you never recommend anything 
without being able to measure it. When reviewing texts, 
you always address the value proposition first, then 
the call-to-action.

The detailed role description works because it orients the model towards specific response patterns from training. The model does not “become” that person, but the precise description of perspective, experience and evaluation criteria produces responses more consistent with what someone with that profile would produce.

An advanced use of role prompting is asking the model to adopt contradictory perspectives on the same problem: “now respond as a sceptical critic of this idea” after having asked “respond as an enthusiastic advocate.” The tension between the two perspectives usually produces a more complete view than either one alone.

Combining all three techniques

The three techniques complement each other and can be used together:

ROLE: You are a senior financial analyst specialising 
     in technology companies in emerging markets.

TASK: Evaluate whether this startup should raise a 
      Series A round now or wait 12 months.

EXAMPLES (few-shot):
The analysis must follow this format:
[Market context]: 2-3 sentences on timing
[Key metrics]: list of the most relevant metrics
[Argument for]: the strongest case for acting now
[Argument against]: the main risk of acting now
[Recommendation]: a single sentence with your position

REASONING (chain-of-thought):
Before your recommendation, list the three factors 
that carry the most weight in your decision and why.

STARTUP INFORMATION:
[data here]

This type of compound prompt — detailed role + format examples + explicit reasoning instruction — produces results significantly more useful than any simple prompt, especially for complex or high-stakes tasks.

The technique has a cost: these prompts are longer and require more tokens. For frequent, repeatable tasks, the prompt design time is quickly amortised. For occasional queries, a simple prompt is often enough.