Specifying a task through language
Prompt engineering changes a model's input without changing its trained weights. Instructions, examples, and context specify what to do and what a useful answer should contain. Their effectiveness depends on the model, task, and evaluation setup.
This lesson uses authored examples to explain the mechanisms. None of the figures calls a language model or reports an accuracy gain. Treat a proposed prompt as a hypothesis to test on representative inputs.
The recipe analogy
A recipe can specify ingredients, quantities, and a sequence of operations. Similarly, a prompt can specify the audience, task, reference material, and output format. More detail can reduce ambiguity, but irrelevant or conflicting requirements can make the task harder. Specific instructions do not guarantee factual accuracy.
Compare three authored instructions, then step through their recipe analogy.
Instructions specify a task
Compare authored recipes and prompts. Detail can reduce ambiguity; adding words does not guarantee a better result.
Make pasta
Write about AI.With minimal instructions, the chef (model) has to guess your preferences. The intended audience, scope, and format are unspecified.
3 recipe steps · 3 prompt words
Prompt anatomy
Different sections serve different purposes. A system instruction can request behavior or tone; reference context supplies material to use; demonstrations show an input-output pattern; constraints specify requirements; and the query identifies the current task. A role such as “expert” does not create expertise or certify the answer.
There is no fixed attention percentage assigned to these sections. All their serialized tokens pass through the model's layers. The prompt information flow lesson separates possible causal connections from measured influence.
Toggle and reorder the sections below to inspect the composed text. The counts measure included sections and words, not quality. Real chat templates encode roles separately, so the text builder is not a role-permission simulator.
Assemble a prompt
Edit, toggle, and reorder authored sections to inspect the resulting text. The counts describe structure, not prompt quality or model attention.
Inspect and edit system instruction
The system instruction sets the persona and behavioral guidelines for the model. It requests a tone and approach; compliance must be evaluated.
Inspect and edit context
Context provides the reference material the model should use. Without it, the model relies solely on its training data, which may be outdated or too general.
Inspect and edit few-shot examples
Examples demonstrate the exact format, style, and depth you expect. They demonstrate a desired input-output mapping.
Inspect and edit constraints
Constraints set boundaries on format, length, style, and content. They specify requirements that can be checked after generation.
Inspect and edit query
The query is the actual question or task. It should be clear and specific. Everything above exists to ensure this query gets the best possible answer.
5 of 5 sections included · 91 whitespace-separated words
System Instruction:
You are a senior data scientist with 10 years of experience. You explain complex topics clearly using analogies and examples.
Context:
The following is a research paper abstract about transformer architectures and their application to computer vision tasks.
Few-Shot Examples:
Example:
Input: "Explain backpropagation"
Output: "Think of backpropagation like a teacher grading papers backwards..."
Constraints:
Keep your response under 200 words. Use bullet points. Do not include code snippets. Write at a college freshman reading level.
Query:
Explain how Vision Transformers (ViT) differ from traditional CNNs for image classification.Intermediate reasoning text
In a causal language model, generated tokens become part of the prefix used for later predictions. Writing intermediate calculations can therefore make partial results available to subsequent steps. A direct answer can also contain several generated tokens; it is not generally a single forward pass for the whole response.
Wei et al. (2022) evaluated chain-of-thought demonstrations on arithmetic, commonsense, and symbolic tasks, including PaLM 540B on GSM8K. Those experiments support results for their evaluated setups, not a universal improvement from requesting steps. Model behavior varies, and a fluent explanation can contain errors or fail to faithfully describe internal computation.
The next figure compares an intentionally incorrect direct attempt with authored worked solutions. Step to the final answer and check the arithmetic; these examples are not sampled model responses.
Inspect a worked solution
All responses below are authored teaching examples. The direct attempts intentionally contain errors; this is not a model evaluation or an accuracy comparison.
A train travels at 60 mph for 2.5 hours, then at 80 mph for 1.5 hours. What is the total distance traveled?
Ready. Play or Next reveals the authored response.
Choosing an intervention
Few-shot examples add demonstrations. A format request specifies a checkable output structure. Intermediate steps expose a proposed calculation. Self-consistency samples multiple candidate answers and aggregates them, adding generation cost. A role request changes wording and context but does not guarantee domain knowledge.
Select a task and technique to see the exact instruction change and a task-specific validation method.
What changes in the prompt?
Choose a task and technique to inspect its concrete instruction. These authored transformations are not measured gains.
A train travels at 60 mph for 2.5 hours and 80 mph for 1.5 hours. Find the total distance.19 whitespace-separated words · 0 example
How to check the result: Check both products and their sum against the problem.
No technique is best for every task. Compare a simple baseline with one change at a time, using the same checkpoint, test inputs, scoring rubric, and decoding settings. Record token use and latency alongside task performance. For tool workflows, evaluate actual tool results and error handling.
Zero-Shot
- Mechanism
- Direct question with no examples or extra context
- Possible use
- Simple, well-defined tasks
- Limitations
- Struggles with multi-step reasoning, ambiguous tasks, or uncommon formats
Few-Shot
- Mechanism
- Include representative input/output examples before the query
- Possible use
- Format-sensitive tasks, classification, structured output
- Limitations
- Token-expensive, examples may bias output, useful example count depends on the task and model
Chain-of-Thought
- Mechanism
- Add "Let's think step by step" or provide step-by-step examples
- Possible use
- Math, logic, multi-step reasoning, analysis
- Limitations
- Increases latency and token usage, may produce plausible but wrong steps
Self-Consistency
- Mechanism
- Sample multiple CoT paths and pick the most frequent answer via majority vote
- Possible use
- Math, code debugging, tasks with checkable final answers
- Limitations
- Requires multiple API calls, significantly higher cost, latency tradeoff
Tree-of-Thoughts
- Mechanism
- Explore multiple reasoning branches, evaluate each, prune and continue the best
- Possible use
- Planning, puzzles, creative problem-solving
- Limitations
- Very token-expensive, complex orchestration, overkill for simple tasks
ReAct
- Mechanism
- Interleave Reasoning and Acting steps, using external tools between thoughts
- Possible use
- Tool use, fact-checking, knowledge-grounded QA
- Limitations
- Requires tool integration, non-trivial error handling, can loop on failures
These are mechanism descriptions, not a ranking or universal accuracy guarantee. Compare alternatives on the same held-out tasks with a fixed model and token budget.
Common pitfalls
Conflicting or unnecessary requirements
Keep the objective clear and include the reference material needed to answer it. Test whether each extra example or instruction helps. Prompt length alone is not a quality measure.
Ambiguous success criteria
“Make it better” leaves the target unclear. “Rewrite this paragraph for a first-year student in under 100 words” supplies an audience and a checkable length requirement. Factual correctness still needs separate verification.
Treating demonstrations as evidence
A few examples can illustrate a pattern without representing the deployment distribution. Use held-out inputs to test generalization, and verify requested citations against their sources. A model repeating a claim or several samples agreeing does not establish truth.
Related concepts
Explore how prompt token positions connect through causal self-attention, and why connectivity and attention weights do not establish causal influence.
Explore neural scaling laws in deep learning: power law relationships between model size, data, and compute that predict AI performance.
Explore an illustrative region-tiling policy: resize tiles, calculate visual tokens, and distinguish analytical savings from measured model quality.
Understand cross-entropy loss for classification: interactive demos of binary and multi-class CE, the -log(p) curve, softmax gradients, and focal loss.
Understand dilated (atrous) convolutions: how dilation rates expand receptive fields without extra parameters and how to avoid gridding artifacts.
Explore emergent abilities in large language models: sudden capabilities at scale thresholds, phase transitions, and the mirage debate.
