K N OW L E D G E B A S E — 2 0 PA P E R S D I S T I L L E D

The Complete Guide to AI Prompt &

Context Engineering
Everything you need to know from 20 landmark research papers — explained in
plain language, no coding required.

What's Inside

00 Essential Vocabulary — 15 terms you must know before reading anything

01 The Foundation — How LLMs learn from prompts (2 papers)

02 Core Reasoning — Teaching AI to think step-by-step (3 papers)

03 Advanced Reasoning — Trees, Graphs & Structured Thinking (3 papers)

04 Grounding in Reality — RAG and real-world knowledge (3 papers)

05 AI Agents — Systems that reason AND take actions (3 papers)

06 Automated Prompting — AI that optimizes its own prompts (3 papers)

07 Context Engineering — The future of AI systems design (3 papers)

08 Quick-Reference Cheatsheet — When to use what technique

09 The Big Picture — How everything connects

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 PHASE 00

Essential Vocabulary
Before diving into papers, here are the 15 terms you'll see everywhere. Master these and you'll
understand 90% of any AI research paper's abstract.

The 15 Terms That Unlock Everything

LLM (Large Language Model)

An AI system trained on massive amounts of text that can generate, understand, and reason
with language. Examples: GPT-4, Claude, Gemini, LLaMA.

Prompt
The input text you give to an AI. It can be a question, instruction, or a detailed briefing. Think
of it as the "job description" you hand to the AI before it starts working.

Context Window
The total amount of text an AI can "see" at once — including your prompt, any examples,
documents, and its own response. Think of it as the AI's desk space — it can only work with
papers that fit on the desk.

Zero-Shot
Asking the AI to do something with zero examples. Just instructions, no demonstrations. Like
telling a chef "make me something Italian" without showing them any recipes.

Few-Shot
Giving the AI 2-5 examples of what you want before asking it to do the task. Like showing a
chef three finished dishes before saying "now make one like these."

Chain-of-Thought (CoT)
Asking the AI to show its reasoning step by step, rather than jumping to a final answer. Like
asking a student to "show your work" on a math exam.

Fine-Tuning
Permanently retraining a model on specific data to change its behavior. Expensive and slow.
Prompt engineering is the alternative — getting better results without retraining.

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 Hallucination
When an AI confidently generates false information. Like a student who doesn't know the
answer but writes something convincing anyway.

RAG (Retrieval-Augmented Generation)

A system that searches for relevant information from a database before answering. Instead
of relying only on memory, the AI can "look things up" — like a doctor consulting a medical
reference during your appointment.

Agent
An AI system that doesn't just generate text but also takes actions — searching the web,
running code, calling APIs, making decisions. It's like upgrading from an advisor to an
assistant who actually does things.

Token
The smallest unit of text an AI processes. Roughly 1 token = ¾ of a word. "Hello world" = 2
tokens. Matters because AI pricing and context windows are measured in tokens.

System Prompt
Hidden instructions given to the AI before the user starts chatting. It defines the AI's
personality, rules, and behavior. Like the training manual an employee reads before their first
day.

Temperature
A setting that controls how creative vs. predictable the AI is. Low temperature (0.0) = strict,
factual answers. High temperature (1.0) = creative, varied, sometimes wild answers.

Inference
The moment when the AI actually generates a response. Training teaches the AI, inference is
the AI using what it learned. Every time you send a message to ChatGPT or Claude, that's
inference.

Context Engineering
The next evolution of prompt engineering. Instead of just writing a good prompt, you design
the entire information environment — instructions, retrieved documents, memory, tools, and
examples — that feeds into the AI at runtime.

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 PHASE 01 — THE FOUNDATION

How LLMs Learn from Prompts

These two papers explain the fundamental discovery that makes everything else possible: you
don't need to retrain an AI for every new task — you just need to write a better prompt.

🧠 The Core Idea

Before 2020, if you wanted an AI to do a new task (say, classify emails), you had to
collect training data, retrain the model, and wait weeks. These papers showed that
large enough models can learn new tasks just from the instructions and examples in
your prompt — no retraining required. This single discovery created the entire field of
prompt engineering.

Paper #1 2020 • OpenAI

Language Models are Few-Shot Learners (GPT-3)

Tom Brown and 30+ co-authors at OpenAI

WHAT THEY DISCOVERED

OpenAI built GPT-3 — a model with 175 billion parameters (think of parameters as
"neurons") — and found something remarkable. When you show this model just 2-3
examples of a task inside the prompt, it can figure out the pattern and do the task for
new inputs, without any retraining. They called this "few-shot learning."

REAL-WORLD ANALOGY

Imagine hiring a new employee who has read every book, article, and document ever
written. On day one, instead of months of training, you just show them 3 examples of
how you want emails written — and they immediately get it. That's what GPT-3 proved
AI could do.

The Takeaway You Need: The more examples you give in your prompt (few-shot),
the better the output. But even zero examples with clear instructions (zero-shot)
works surprisingly well for large models. The key is that the quality of your

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 prompt directly determines the quality of the output — which is why prompt
engineering matters.

arXiv: 2005.14165 →

Paper #2 2021 • Carnegie Mellon / Multiple

Pre-train, Prompt, and Predict: A Systematic Survey of

Prompting Methods
Pengfei Liu, Weizhe Yuan et al.

WHAT THIS PAPER DOES

This is the first major "map" of the prompting world. It organizes every known
prompting technique into categories and explains how AI went through three ages: first
we had "feature engineering" (manually picking data features), then "fine-tuning"
(retraining for each task), and now "prompt engineering" (writing smart instructions). It
defines the vocabulary and concepts that every subsequent paper builds on.

The Takeaway You Need: AI has shifted from "train a model for each task" to "one
model, many prompts." Understanding this shift is crucial because it means the
most valuable skill in AI is no longer coding models — it's designing the right
instructions. This paper gives you the complete mental framework for thinking
about prompts.

arXiv: 2107.13586 →

PHASE 02 — CORE REASONING

Teaching AI to Think Step-by-Step

These three papers are the most practically useful in the entire list. They introduced the
techniques that every AI power-user applies daily.

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 🔑 Why This Matters for You
When you ask an AI a hard question and it gives a wrong answer, the problem usually
isn't that the AI is stupid — it's that you asked it to jump to the answer without
thinking. These papers discovered that making the AI "show its work" massively
improves accuracy. This is the single most important practical skill in prompt
engineering.

Paper #3 2022 • Google Brain

Chain-of-Thought Prompting Elicits Reasoning in Large

Language Models
Jason Wei, Xuezhi Wang et al.

WHAT THEY DISCOVERED

Google researchers found that if you show the AI an example where the reasoning
steps are written out (not just the final answer), the AI starts reasoning through its own
problems step-by-step too. A 540-billion parameter model with just 8 chain-of-
thought examples beat a fine-tuned GPT-3 on math problems — without any
retraining.

REAL-WORLD ANALOGY

If a maths teacher just writes "Answer: 42" on the board, students don't learn how to
solve problems. But if the teacher writes out every step — "First I add these... then I
multiply... then I subtract..." — students can follow the logic. Chain-of-Thought does the
same thing for AI. You show it one worked example, and it learns to work through
problems methodically.

HOW TO ACTUALLY USE THIS

Simple version: Add "Think through this step by step" or "Let's work through this
carefully" to your prompts. Advanced version: Include one fully worked example in
your prompt, showing the reasoning process from question to answer. The AI will mirror
that reasoning pattern for your new question.

Critical limitation discovered later: Chain-of-Thought only helps on models with

100+ billion parameters. Smaller models actually get worse when you ask them to
reason step-by-step — they produce fluent but logically wrong reasoning. Also,

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 newer "reasoning models" (like Claude's extended thinking) have this built in, so
manually adding CoT to them can sometimes be redundant.

arXiv: 2201.11903 →

Paper #4 2022 • Google Brain

Self-Consistency Improves Chain of Thought Reasoning

Xuezhi Wang, Jason Wei et al.

WHAT THEY DISCOVERED

One chain-of-thought path might lead to a wrong answer (just like one person might
make a calculation error). This paper's solution: ask the AI to solve the same problem
5-10 different ways, then pick the answer that appears most often. It's a majority
vote system. This simple trick improved accuracy dramatically across math, logic, and
commonsense reasoning tasks.

REAL-WORLD ANALOGY

If you ask one doctor for a diagnosis, they might be wrong. If you ask five doctors
independently and four say the same thing, you can be much more confident. Self-
Consistency applies this "second opinion" principle to AI reasoning.

The Takeaway You Need: When accuracy matters more than speed (for example,
financial analysis or medical questions), asking the AI to generate multiple
answers and looking for agreement is one of the most reliable ways to improve
quality. The trade-off is that it uses more tokens (costs more money) since
you're generating 5-10 responses instead of 1.

arXiv: 2203.11171 →

Paper #5 2022 • University of Tokyo

Large Language Models are Zero-Shot Reasoners

Takeshi Kojima et al.

WHAT THEY DISCOVERED

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 This paper made a beautifully simple discovery: you don't even need worked examples.
Just adding the phrase "Let's think step by step" to the end of any prompt activates
reasoning behavior in the AI. No examples needed. This "zero-shot CoT" approach
works across math, logic, common sense, and many other tasks — making it the
simplest and most widely applicable prompt engineering trick ever discovered.

The Takeaway You Need: This is your default starting move for any hard question.
Before you do anything else fancy, just add "Let's think step by step" or "Think
carefully before answering" to your prompt. It's free, it's easy, and it works. If the
answer is still wrong, then escalate to more advanced techniques.

arXiv: 2205.11916 →

The Reasoning Ladder (simplest → most reliable):

Direct Question → Zero-Shot CoT ("think step by step") → Few-Shot CoT

(show worked examples) → Self-Consistency (multiple attempts, majority

vote)

PHASE 03 — ADVANCED REASONING

Trees, Graphs & Structured Thinking

Chain-of-thought thinks in a straight line. But some problems need the AI to explore multiple
paths, backtrack, and combine ideas. These papers teach structured exploration.

🌳 The Key Concept: Thinking Structures

Imagine you're solving a maze. Chain-of-thought walks down one path until the end.
Tree of Thoughts explores multiple paths, evaluates which ones look promising, and
backtracks when it hits dead ends. Graph of Thoughts goes even further — letting

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 different thinking paths merge and combine insights. These are different "shapes" of
thinking.

Paper #6 2023 • Princeton University

Tree of Thoughts: Deliberate Problem Solving with Large

Language Models
Shunyu Yao et al.

WHAT THEY DISCOVERED

Chain-of-thought follows one thinking path to the end — and if that path has an error
early on, the whole answer goes wrong. Tree of Thoughts (ToT) fixes this by having the
AI generate multiple possible "next steps" at each point, evaluate which ones are most
promising, and explore the best ones — just like a chess player thinking several moves
ahead. If a path seems bad, the AI backtracks and tries another.

WHEN TO USE THIS

Use ToT for creative tasks (brainstorming, writing), planning tasks (project plans,
strategies), puzzles, and any problem where the first approach might not be the best
one. It's overkill for simple factual questions — stick with basic CoT for those.

The Takeaway: When a problem has multiple possible approaches and you're not
sure which is best, ask the AI to propose 3 different approaches, evaluate each
one's pros and cons, then develop the most promising one in detail. You're
manually doing what ToT does automatically.

arXiv: 2305.10601 →

Paper #7 2023 • ETH Zurich

Graph of Thoughts: Solving Elaborate Problems with Large

Language Models
Maciej Besta et al.

WHAT THEY DISCOVERED

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 In a tree, branches never reconnect — each path is independent. But real thinking
often involves combining ideas from different lines of thought. Graph of Thoughts
(GoT) allows thinking paths to merge, so the AI can take the best insight from Path A,
combine it with an insight from Path B, and create a synthesis that neither path could
reach alone.

The Takeaway: When you need the AI to synthesize information from multiple
sources or perspectives, explicitly ask it to: (1) analyze from perspective A, (2)
analyze from perspective B, (3) combine the best insights from both. This
mirrors the graph structure and produces richer, more nuanced answers.

arXiv: 2308.09687 →

Paper #8 2022 • Google Brain

Least-to-Most Prompting Enables Complex Reasoning

Denny Zhou et al.

WHAT THEY DISCOVERED

When a problem is too complex for the AI to solve in one shot, break it into a staircase
of sub-problems — from simplest to hardest. Solve the simplest first, use that answer
to solve the next one, and keep climbing. This "least-to-most" approach let AI solve
problems that were far harder than anything it could handle directly — including
problems longer than anything in the training data.

HOW TO USE THIS RIGHT NOW

When the AI gives you a bad answer on a complex question, don't just rephrase the
question. Instead, ask: "What are the sub-questions I need to answer first before I
can answer this main question?" Then solve each sub-question one at a time. Feed
each answer into the next question. You'll often get a dramatically better final answer.

The Takeaway: Complex problems become manageable when decomposed into

steps. This is the single most useful technique for business-related prompts —
market analysis, competitive research, strategic planning. Always break big
questions into smaller, answerable sub-questions.

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 arXiv: 2205.10625 →

Thinking Structures (linear → most complex):

Chain (A→B→C→Answer) → Tree (explore A,B,C → pick best → continue) → Graph

(A+B merge into D, C+D produce Answer) → Least-to-Most (solve small →

medium → large)

PHASE 04 — GROUNDING IN REALITY

RAG: Connecting AI to Real-World Knowledge

LLMs trained on old data hallucinate about current events. They don't know your company's
internal docs. RAG solves this by letting AI "look things up" at runtime.

📚 The Fundamental Problem RAG Solves

Every AI model has a "training cutoff" — a date after which it knows nothing. It also has
no access to your private data, your company documents, or today's news. RAG solves
this by adding a "search step" before generation — the AI first searches a knowledge
base for relevant information, then uses that information to answer your question. This
is how every useful enterprise AI chatbot works today.

Paper #9 2020 • Facebook AI / UCL

Retrieval-Augmented Generation for Knowledge-Intensive

NLP Tasks
Patrick Lewis et al.

WHAT THEY DISCOVERED

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 This team combined two systems: a "retriever" (that searches for relevant documents
in a database) and a "generator" (the LLM that writes the answer). They trained both to
work together. The result was a system that could answer questions far more
accurately than the LLM alone — because it had access to real evidence, not just
memory. They called it "Retrieval-Augmented Generation" — RAG.

REAL-WORLD ANALOGY

Imagine two types of doctors. Doctor A answers from memory only — sometimes
outdated, sometimes wrong. Doctor B has a medical database at hand and looks up
the latest research before answering. Doctor B is obviously more reliable. RAG turns
Doctor A into Doctor B.

The Takeaway: If you're building any AI product that needs to answer questions
about specific, current, or private data — you need RAG, period. This is not
optional for production AI. The concept is simple: search first, then generate. But
the implementation details (what to search, how to rank results, how many
documents to include) determine whether your system is great or terrible.

arXiv: 2005.11401 →

Paper #10 2023 • University of Washington

Self-RAG: Learning to Retrieve, Generate, and Critique

through Self-Reflection
Akari Asai et al.

WHAT THEY DISCOVERED

Basic RAG has a problem: it always retrieves documents, even when the AI already
knows the answer. This wastes time and sometimes the retrieved documents are
irrelevant or even misleading. Self-RAG teaches the model three skills: (1) decide
whether retrieval is needed, (2) retrieve if needed, and (3) self-evaluate whether the
retrieved info actually helps the answer. It's RAG with a built-in quality control
system.

The Takeaway: Not every question needs a database search. The most effective
AI systems are smart about when to look things up and whether the retrieved
information actually improves the answer. When designing your own AI

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 workflows, always include a "do I need to search?" decision point — don't just
blindly retrieve for every query.

arXiv: 2310.11511 →

Paper #11 2024 • Multi-institutional

Retrieval-Augmented Generation for AI-Generated Content: A

Survey
Penghao Zhao et al.

WHAT THIS PAPER DOES

This is the comprehensive map of the entire RAG ecosystem as of 2024. It covers RAG
for text, images, code, audio, and video. It categorizes all RAG approaches into a clear
framework: what kind of retriever you use, how you augment the generation, and how
to evaluate quality. If you want to understand the full landscape of RAG options, this is
your reference guide.

The Takeaway: RAG isn't just for text chatbots anymore. It's being applied to
image generation, code completion, video understanding, and more. The core
principle — "look up relevant information before generating" — applies to any
AI modality. This survey helps you understand which RAG architecture fits which
use case.

arXiv: 2402.19473 →

The RAG Pipeline (what happens under the hood):

User Question → Should I search? (Self-RAG decides) → Search knowledge

base → Rank top documents → Insert relevant docs into prompt → Generate
answer grounded in evidence → Self-check for accuracy

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 PHASE 05 — AI AGENTS

Systems That Reason AND Take Actions

Until now, everything was about generating better text. These papers take the leap to AI that
can act in the world — searching, coding, browsing, and making decisions.

🤖 The Agent Revolution

An AI chatbot just talks. An AI agent thinks and does. It can search the web for
information, run calculations, call APIs, browse websites, manage files, and chain
multiple actions together to complete complex tasks. The prompt doesn't just say "tell
me about X" — it defines the agent's goals, available tools, and decision-making
process. This is the frontier of what prompt engineering has become.

Paper #12 2022 • Princeton / Google

ReAct: Synergizing Reasoning and Acting in Language Models

Shunyu Yao et al.

WHAT THEY DISCOVERED

Before ReAct, AI could either reason (chain-of-thought) OR take actions (call APIs) —
but not both at the same time. ReAct proved that interleaving reasoning with actions
produces far better results. The AI thinks: "I need to find X" (reason), then searches
Wikipedia (act), then thinks: "This page says Y, but I still need Z" (reason), then searches
again (act). This think-act-think-act loop is now the foundation of every agent
framework.

REAL-WORLD ANALOGY

Imagine a detective solving a case. They don't just sit in their office thinking (pure
reasoning). And they don't just randomly search for clues without thinking (pure
acting). They think about what clue to look for, go find it, think about what it means,
then decide what to search for next. ReAct gives AI this same think-act cycle.

The Takeaway: This is the most important paper for understanding how modern
AI agents work. Every time you see Claude searching the web, writing code, or
using tools — it's using the ReAct pattern. When you design AI workflows,

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 always structure them as loops of "think about what to do next → do it →
think about what the result means → decide the next action."

arXiv: 2210.03629 →

Paper #13 2023 • Zhejiang / Microsoft

HuggingGPT: Solving AI Tasks with ChatGPT and its Friends

Yongliang Shen et al.

WHAT THEY DISCOVERED

What if one AI could act as a "manager" that delegates tasks to other specialized AI
models? HuggingGPT does exactly this: you give ChatGPT a complex request like
"describe this image, translate the description to Hindi, and create an audio version."
ChatGPT breaks this into three sub-tasks, picks the best AI model for each (an image
model, a translation model, a speech model), sends each sub-task to the right model,
and combines the results.

The Takeaway: The future of AI isn't one model doing everything — it's one
"brain" model orchestrating many specialized models. This is relevant for
anyone building AI products: you don't need one perfect AI. You need a smart
prompt that knows which tool to use for each part of a complex job. Think of it
like a project manager coordinating a team of specialists.

arXiv: 2303.17580 →

Paper #14 2023 • Stanford / Google

Generative Agents: Interactive Simulacra of Human Behavior

Joon Sung Park et al.

WHAT THEY DISCOVERED

Stanford researchers created a virtual town where 25 AI characters lived, worked,

formed friendships, planned a Valentine's Day party, and made decisions — all powered
by carefully engineered prompts and a memory system. Each agent had three layers: a
memory stream (recording everything that happened), a reflection system (making

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 sense of memories), and a planning system (deciding what to do next). The result was
eerily human-like behavior.

The Takeaway: This paper is the ultimate proof-of-concept for context

engineering. It shows that with the right architecture of memory + reflection +
planning, AI agents can exhibit complex, believable behavior over extended
periods. The secret wasn't a better model — it was better context
management. For anyone building AI products, this paper demonstrates that
how you manage the AI's memory and context over time matters as much as the
model itself.

arXiv: 2304.03442 →

PHASE 06 — AUTOMATED PROMPTING

AI That Optimizes Its Own Prompts

What if instead of you writing prompts by trial and error, another AI could systematically find
the best prompt? That's automatic prompt engineering.

⚡ The Shift from Craft to Engineering

Manual prompt engineering is like hand-tuning a guitar by ear. Automatic prompt
engineering is like using a digital tuner. Both get you there, but one is systematic,
repeatable, and scalable. These papers show how AI itself can generate, test, and
refine prompts far faster than any human.

Paper #15 2022 • University of Toronto

Large Language Models Are Human-Level Prompt Engineers

(APE)
Yongchao Zhou et al.

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 WHAT THEY DISCOVERED

APE (Automatic Prompt Engineer) works in three steps: (1) give an LLM a task
description and some examples, and ask it to generate many candidate prompts; (2)
test each prompt on a set of examples to see which ones work best; (3) pick the
winner. The result? Machine-generated prompts that often outperform the best
prompts written by human experts.

The Takeaway: You can use this idea right now, manually. When a prompt isn't
working well, ask the AI itself: "Rewrite this prompt in 5 different ways that
might produce better results." Then test each version and keep the best one.
You're doing manual APE. The paper proves this approach works systematically.

arXiv: 2211.01910 →

Paper #16 2023 • Stanford

DSPy: Compiling Declarative Language Model Calls into Self-

Improving Pipelines
Omar Khattab et al.

WHAT THEY DISCOVERED

DSPy is a framework where you don't write prompts at all. Instead, you define what you
want (e.g., "take a question and produce a factual answer with sources") using small
programmatic modules, and DSPy's "compiler" automatically figures out the best
prompts, examples, and chain structures to achieve your goal. When the LLM changes
or the task shifts, DSPy re-optimizes automatically.

REAL-WORLD ANALOGY

Think of the evolution from hand-coding websites in raw HTML to using tools like
WordPress or Shopify. DSPy is the same shift for prompts. Instead of crafting every
word of every prompt by hand, you declare your intent and let the system optimize the
implementation. It's "prompt engineering" evolving into "prompt compilation."

The Takeaway: The future of prompt engineering is not about writing better
prompts manually — it's about building systems that optimize prompts
programmatically. Even if you never use DSPy directly, understanding this
concept matters because it shows where the industry is heading. The skill of the

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 future is defining clear objectives and evaluation criteria — not handcrafting
prompt text.

arXiv: 2310.03714 →

Paper #17 2025 • Multi-institutional

A Survey of Automatic Prompt Engineering: An Optimization

Perspective
Wenhao Li et al.

WHAT THIS PAPER DOES

This 2025 survey frames the entire field of automatic prompt engineering as a
mathematical optimization problem. It covers every known approach — evolutionary
algorithms (prompts "breed" and evolve), gradient-based methods (using calculus to
find better prompts), reinforcement learning (prompts learn from reward signals), and
LLM-based methods (using one AI to improve another AI's prompts). It's the definitive
reference for understanding all the ways prompts can be automatically improved.

The Takeaway: This is a reference paper, not one you read cover-to-cover.
Bookmark it for when you need to understand a specific automatic prompting
technique. The most practical insight for non-coders: prompt optimization is
moving from "human art" to "engineering discipline" with measurable,
repeatable results.

arXiv: 2502.11560 →

PHASE 07 — CONTEXT ENGINEERING

The Future: Designing Complete Information

Systems

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 Prompt engineering focused on what you type. Context engineering focuses on everything the
AI sees — prompts, retrieved documents, memory, tool outputs, conversation history, and
more.

🏗️ From Prompts to Environments

Think of it this way: prompt engineering is writing a good email to your assistant.
Context engineering is designing their entire office — the briefing documents on their
desk, the tools at their fingertips, the filing system with past conversations, and the live
data feeds on their screen. The email (prompt) is just one small part of the total
information environment that determines the quality of the AI's work.

Paper #18 2024 • OpenAI / Google / Stanford

The Prompt Report: A Systematic Survey of Prompting

Techniques
Sander Schulhoff et al. (32 researchers)

WHAT THIS PAPER DOES

A team of 32 researchers from OpenAI, Google, Stanford, Microsoft, and Princeton

analyzed over 1,500 academic papers and catalogued every known prompting
technique into a unified system. They identified 58 distinct text-based prompting
techniques and 40 techniques for other modalities (images, audio, video). They also
created a standardized vocabulary of 33 terms. This is the most comprehensive survey
of prompt engineering ever published.

The Takeaway: You don't need to memorize 58 techniques. But this paper is your
encyclopedia. When you face a specific challenge — summarization quality,
reasoning accuracy, creative output, factual grounding — look up which
technique addresses it. The paper's taxonomy lets you quickly find the right tool
for the job. Think of it as a "Yellow Pages" for prompt engineering.

arXiv: 2406.06608 →

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 Paper #19 2025 • Multi-institutional

A Survey of Context Engineering for Large Language Models

Lingrui Mei et al.

WHAT THEY DISCOVERED

This paper formally defines "Context Engineering" as its own discipline — separate from
and broader than prompt engineering. It covers three pillars: (1) Context Retrieval &
Generation — finding and creating the right information; (2) Context Processing —
handling long documents, summarization, and structured data; (3) Context
Management — memory systems, compression, and optimization. It then shows how
these combine into real systems: RAG, agents, multi-agent systems, and tool-using AI.

The Takeaway: The term "prompt engineering" is becoming too small for what
builders actually do. The real skill is context engineering — designing the
complete information environment that an AI operates within. This paper
provides the definitive framework for thinking about it. If you read one paper from
this list in full, make it this one — it connects everything else together.

arXiv: 2507.13334 →

Paper #20 2025 • Stanford

Agentic Context Engineering: Evolving Contexts for Self-

Improving Language Models
Qizheng Zhang et al.

WHAT THEY DISCOVERED

ACE treats the AI's entire context (system prompts, instructions, memory) as a "living
playbook" that automatically evolves and improves over time. As the AI encounters
new tasks and gets feedback, its context grows smarter — accumulating strategies,
refining approaches, and organizing knowledge. The key result: using a smaller,
cheaper open-source model with ACE's evolving contexts matched or beat a top-
tier expensive model on agent benchmarks.

WHY THIS IS THE FUTURE

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 This paper proves a powerful idea: you can compensate for a weaker model with
better context engineering. A cheaper model with brilliant context can beat an
expensive model with mediocre context. For someone building AI products on a
budget, this is incredibly important. You don't always need the most expensive model
— you need the best-designed context.

The Takeaway: The most important insight from this entire roadmap: the context
you build around an AI matters as much as (or more than) the AI model itself.
Invest in designing better instructions, better retrieval, better memory, and better
tool integration — not just buying bigger models. This is the future competitive
advantage in AI.

arXiv: 2510.04618 →

Quick-Reference: When to Use What

A practical decision guide. When you face a specific challenge, this tells you which technique
to reach for.

Situation → Technique

"The AI gives a wrong answer on a hard question"

Start with Zero-Shot CoT: add "think step by step." If still wrong, use Few-Shot CoT with 2-3
worked examples. If it's a math/logic problem, try Self-Consistency (generate 5 answers,
pick the majority).

"The AI hallucinates facts"

This is a grounding problem. You need RAG — provide the AI with real documents/data in the
prompt. If building a product, implement Self-RAG so the system decides when to search.

"The problem is too complex for a single prompt"

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 Use Least-to-Most prompting: break the problem into sub-questions, solve from simplest
to hardest. Or use the manual Tree of Thoughts approach: ask for 3 approaches, evaluate,
then develop the best one.

"I need the AI to use tools / search / take actions"

Use the ReAct pattern: design your prompt so the AI alternates between thinking about
what to do and actually doing it. Define available tools clearly in the system prompt.

"I can't get the prompt wording right"

Use the APE principle: ask the AI to generate 5 different versions of your prompt, then test
each one. Let the AI optimize its own instructions.

"I need the AI to handle long conversations / maintain context"

This is a context engineering challenge. Design a memory system (key facts to remember),
use context compression (summarize old messages), and prioritize which information stays
in the context window.

"I want to build a production AI product"

You need full context engineering: system prompt + RAG pipeline + memory management +
tool integration + evaluation framework. Read the Context Engineering Survey (Paper #19)
for the complete architecture.

The Big Picture: How Everything Connects

Step back and see the complete evolution — from a simple prompt to a full context-
engineered AI system.

The Evolution of Interacting with AI:

2020 — "Just ask" (zero-shot prompting) — GPT-3 paper

2021 — "Show a few examples" (few-shot prompting) — Survey papers

2022 — "Show your reasoning" (chain-of-thought) — CoT / Self-Consistency

2022 — "Think AND act" (agents) — ReAct paper

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF
 2023 — "Explore multiple paths" (structured reasoning) — Tree/Graph of

Thoughts

2023 — "Look up real information" (grounding) — RAG / Self-RAG

2023 — "Optimize prompts automatically" (auto-prompting) — APE / DSPy

2024 — "Catalogue all 58 techniques" (comprehensive mapping) — The Prompt

Report
2025 — "Design the complete information environment" (context engineering)

— CE Survey / ACE

🎯 The One Sentence That Summarizes Everything

The quality of an AI's output is determined not by the model's intelligence alone, but by
the quality of the context you provide — the instructions, examples, retrieved
knowledge, memory, tools, and structure you design around it. Master context
engineering, and you can make even a modest model produce extraordinary results.

📖 If You Read Only 3 Papers, Read These

Paper #3 (Chain-of-Thought) — The most practically useful technique for daily use. It
will immediately make every AI interaction better.

Paper #12 (ReAct) — The foundation of how modern AI agents work. Understanding
this pattern lets you understand every AI product being built today.

Paper #19 (Context Engineering Survey) — The complete map of where the field is
heading. This one connects everything together and shows you the full picture.

Explore our developer-friendly HTML to PDF API Printed using PDFCrowd HTML to PDF