📝 Prompt Engineering

Description

< What is it? >

Prompt engineering is the practice of writing and evaluating instructions that guide an LLM toward a useful, accurate, and well-formatted result. A prompt supplies the task, relevant context, constraints, and desired output—not new model knowledge or permanently changed weights.

Application prompting needs more rigor than a personal chat prompt: it should work reliably across varied inputs, be evaluated on representative cases, and define what to do when information is missing or uncertain.

Key points

< Writing a reliable prompt >

• Be clear: use precise, unambiguous instructions.
• Be complete: provide the context, constraints, and data the model needs.
• Specify the output: state the format, length, audience, and any required fields.
• Organize the prompt: place critical instructions prominently and use consistent terminology.
• Test systematically: evaluate representative and difficult cases instead of trusting one good-looking answer.
• State positive actions: say what the model should do, not only what it should avoid.

< Chain-of-thought (CoT) prompting >

Chain of thought (CoT) asks a model to produce a logical sequence of intermediate steps before its final answer. It can help a capable model break a complex task into smaller, checkable subproblems instead of jumping directly from the question to an answer.

• Word-problem example:

  Prompt: Solve step by step. You have 7 bananas, give 4 to a friend,
  receive 5 more, then throw out 3 rotten bananas. How many remain?
  
  Concise reasoning:
  1. Start with 7 bananas
  2. Give away 4: 7 - 4 = 3
  3. Receive 5: 3 + 5 = 8
  4. Throw out 3: 8 - 3 = 5
  
  Final answer: 5 bananas

• Business-calculation example: suppose a $2,400 project receives a $15\%$ discount, then $8\%$ tax is applied, and three people split the final bill equally:

  1. Discounted price: 2400 × (1 - 0.15) = 2040
  2. Tax: 2040 × 0.08 = 163.20
  3. Final bill: 2040 + 163.20 = 2203.20
  4. Per-person share: 2203.20 ÷ 3 = 734.40
  
  Final answer: $734.40 per person

CoT prompting changes the prompt, not the model weights. Post-training can additionally teach a reasoning model to spend more tokens on difficult, verifiable problems through supervised fine-tuning or reinforcement learning.

< When CoT helps >

Task type	How CoT can help
Mathematics and quantitative analysis	Makes intermediate calculations and assumptions inspectable
Coding and debugging	Encourages a plan, hypothesis checks, and a path from symptom to likely cause
Rule-based or symbolic problems	Applies explicit conditions in a deliberate order, such as logic puzzles or multi-step transformations
Complex questions and planning	Separates subquestions, trade-offs, and dependencies before synthesizing an answer
Contextual or commonsense questions	Makes the relevant constraints explicit, rather than relying on an unexplained conclusion

For a simple fact lookup, classification, or creative-writing request, CoT often adds tokens and latency without improving the outcome.

< Designing a CoT prompt >

Ask for the task, the needed reasoning format, and the final output explicitly. A practical pattern is:

1. State assumptions or missing information.
2. Solve the relevant subproblems in order.
3. Check constraints or calculations.
4. Return a concise final answer in the requested format.

Useful instructions include Explain why, State your reasoning briefly, First determine ___, then identify ___, and If ___, then do ___; otherwise ___. For a recurring task, add verified examples and evaluate the result on cases that differ from those examples.

The visible rationale is generated text, not proof that every step is correct or a faithful record of the model’s internal computation. Use it to make assumptions and calculations easier to review, then verify important answers with tools, tests, or source material.

< Common CoT variants >

Variant	What it adds
Zero-shot CoT	A generic cue such as “reason step by step,” with no worked examples
Few-shot CoT	A few demonstrations that show both the intermediate steps and final-answer format
Least-to-most prompting	Decomposes a hard problem into easier subproblems, solves them, then combines the results
Auto-CoT	Automatically constructs or selects reasoning demonstrations for a task; it is not an automatic feature of every LLM
Multimodal CoT	Uses images, audio, video, or other modalities as part of the reasoning context when the model supports them

Comparison

< Zero-shot prompting vs. zero-shot image classification >

• Different scopes: zero-shot prompting describes the examples supplied in a prompt, while zero-shot image classification describes how a model is applied to a target classification task.

  Concept	Meaning	Example
  Zero-shot prompting	Request a task without providing worked input–output demonstrations in the prompt	Give a VLM a product image and ask, “Classify this product as dress, shirt, or shoes”
  Zero-shot image classification	Classify images into target categories without additional training on labeled examples from the target classification task	CLIP compares a dress image with text descriptions such as “a dress,” “a shirt,” and “shoes,” then selects the highest-scoring match

  The prompting distinction follows the in-context evaluation settings in Language Models are Few-Shot Learners; CLIP demonstrates classification without dataset-specific training.

• What counts as a demonstration: instructions, candidate labels, and the input to be classified do not constitute worked examples. An additional image paired with its correct label would be a demonstration. A zero-shot prompt can therefore include an image, an article, and a list of allowed labels.

• Overlap: a pretrained generative VLM can perform zero-shot image classification through zero-shot prompting. However, the absence of demonstrations in the prompt does not establish whether the model was previously fine-tuned on that target task. Zero-shot transfer also does not imply that relevant concepts were absent from pretraining.

• Examples in these notes: the CLIP dress example demonstrates zero-shot image classification through image–text similarity. The Qwen news example uses zero-shot prompting to request an impact label and explanation from an article and image.

< CoT vs other prompting patterns >

Technique	What the prompt provides	Best use	Main trade-off	Example
Standard prompting	Direct task or question	Straightforward requests	Little guidance for a complex process	“There are 3 bags with 4 apples each. How many apples are there?”
Zero-shot prompting	Instructions, context, and no examples	A clear task the model already understands	May be brittle for unfamiliar labels or formats	“Return only the total number of apples: 3 bags with 4 apples each.”
Zero-shot CoT	Zero-shot prompt plus a reasoning cue	Multi-step tasks when no examples are available	Longer output; not guaranteed to improve accuracy	“There are 3 bags with 4 apples each. How many apples are there? Work through the calculation step by step.”
Few-shot prompting	A few input–output examples	New label schemes, styles, and output formats	Uses context and may overfit to the examples	“Q: 2 bags, 3 apples each. A: 6. Q: 4 bags, 2 apples each. A: 8. Q: 3 bags, 4 apples each. A:”
Few-shot CoT	Examples that include intermediate reasoning	Difficult tasks needing both a process and a format	Most context-heavy prompting option in this table	“Q: 2 bags, 3 apples each. A: Multiply bags by apples per bag: 2 × 3 = 6 apples. Q: 4 bags, 2 apples each. A: Multiply bags by apples per bag: 4 × 2 = 8 apples. Q: 3 bags, 4 apples each. A:”

CoT is not automatically more accurate or more debuggable than every alternative. It is most valuable when the intermediate work can be checked; otherwise, a polished explanation can still hide a wrong conclusion.

Implementation

• https://github.com/yujie-hao/yujie-hao-prompt-engineering-with-langchain

Video Tutorial

Reference

• Language Models are Few-Shot Learners
• CLIP: Learning Transferable Visual Models From Natural Language Supervision
• Prompt Engineering with LangChain | DataCamp
• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., 2022)
• Least-to-Most Prompting Enables Complex Reasoning in Large Language Models (Zhou et al., 2022)
• Automatic Chain of Thought Prompting in Large Language Models (Zhang et al., 2022)
• Chain of thought prompting (aws)
• Language Models Perform Reasoning via Chain of Thought (Google Research)