THE MINIMALIST CONTEXT ENGINEERING
Context Engineering Explained in 100 Words
Ravindra Dastikop
Understand less. Accomplish more.
Why This Book?
Context engineering can sound complicated.
Instructions.
Tools.
Retrieval.
Memory.
History.
Compaction.
Selection.
Tokens.
Context windows.
Judgment.
But understanding context engineering does not have to begin with complexity.
This book strips context engineering down to what matters.
10 ideas.
10 words each.
100 words.
Enough to begin.
How to Read This Book
Read one idea.
Stop.
Think about it.
Try it with AI.
Then move to the next.
The purpose is not to know everything about context engineering.
The purpose is to understand enough to accomplish something with context engineering.
01 — CONTEXT ENGINEERING
Context engineering curates information models receive to improve their performance.
Prompt engineering asks:
What should I tell the model?
Context engineering asks a broader question:
What should the model know right now?
The model may receive much more than a prompt.
Its context can contain:
Instructions.
Conversation history.
Retrieved information.
Tool definitions.
Tool results.
Memory.
Files.
Examples.
System information.
The challenge is therefore not simply writing better instructions.
It is assembling the right information at the right time.
02 — INSTRUCTIONS
Instructions define goals, behavior, constraints, and priorities for model execution.
Instructions remain fundamental.
They tell the model:
What it should accomplish.
How it should behave.
What constraints apply.
What output is expected.
What it should avoid.
Prompt engineering therefore does not disappear when context engineering begins.
It becomes one component of a larger system.
A beautifully written prompt cannot compensate for missing information, irrelevant context, poor tools, or outdated knowledge.
Think:
The prompt matters.
But so does everything surrounding it.
03 — TOOLS
Tools provide capabilities and descriptions models need for effective action.
Agents increasingly do more than generate text.
They can:
Search.
Retrieve files.
Run code.
Call functions.
Use APIs.
Interact with systems.
Take actions.
For the model to use tools effectively, information about those tools must become part of its working context.
Tool descriptions therefore consume context too.
04 — RETRIEVAL
Retrieval supplies relevant external information when models actually need it.
Imagine an organization has one million documents.
Should all one million documents be placed into the model's context?
No.
Instead, the system can retrieve information relevant to the current task.
This is just-in-time context.
The model receives information when it becomes useful rather than carrying everything continuously.
This connects several ideas in the Minimalist Series:
Retrieval answers:
What information does the model need now?
05 — MEMORY
Memory preserves useful information across interactions, tasks, and longer workflows.
Some information remains useful beyond the current turn.
That is where memory enters.
Memory might preserve:
Preferences.
Decisions.
Task progress.
Important facts.
Previous outcomes.
Workflow state.
Long-running agents especially need ways to preserve useful information without continually carrying every previous interaction.
The challenge is not simply remembering.
It is remembering what remains useful.
06 — HISTORY
History carries prior messages, actions, and results into subsequent reasoning.
Conversation creates information.
Every turn can add:
Questions.
Answers.
Corrections.
Tool calls.
Tool results.
Decisions.
Intermediate work.
Some of this history may matter for the next step.
Some may not.
If everything is continually preserved, the context can become crowded.
If too much is removed, the model may lose continuity.
So context engineering asks:
What history still matters?
07 — COMPACTION
Compaction compresses accumulated context while preserving information needed for continuation.
Long-running agents create a problem.
Their context keeps growing.
Messages accumulate.
Tool results accumulate.
Intermediate work accumulates.
Eventually, keeping everything becomes inefficient or impossible.
Compaction addresses this by reducing accumulated information while preserving what the agent needs to continue.
Think of compaction as:
The objective is not merely fewer tokens.
It is preserving useful continuity with fewer tokens.
08 — SELECTION
Selection prioritizes useful information while excluding distracting or redundant context.
More context is not automatically better context.
Imagine giving an AI:
Twenty irrelevant documents.
Old instructions.
Repeated tool outputs.
Unnecessary conversation history.
Outdated information.
And one relevant document.
The useful information is present.
But it is surrounded by noise.
Context engineering therefore involves subtraction as well as addition.
Ask:
Does the model need this information now?
If not, perhaps it should not occupy the context.
09 — BUDGET
Budget allocates limited context capacity across competing sources of information.
Context is not unlimited.
Even models with very large context windows operate within finite capacity.
And different information competes for that capacity:
Instructions.
Conversation.
Retrieved documents.
Tool definitions.
Tool results.
Memory.
Examples.
Intermediate work.
So context becomes a resource.
The engineering question becomes:
What deserves space?
This creates a useful principle:
Every token occupying context should earn its place.
10 — JUDGMENT
Judgment determines what context models receive, retain, retrieve, or discard.
Context engineering ultimately involves choices.
Should this information enter the context?
Should this conversation history remain?
Should this document be retrieved?
Should this memory persist?
Should this tool result be retained?
Should old information be compressed?
Should something be removed?
The system can automate many of these decisions.
But designers still determine the architecture, boundaries, retrieval strategy, memory policy, and evaluation criteria.
Context engineering therefore brings us back to human judgment.
The goal is not:
Give AI everything.
The goal is:
Give AI what helps it accomplish the task.
The Entire Book in 100 Words
CONTEXT ENGINEERING
Context engineering curates information models receive to improve their performance.
INSTRUCTIONS
Instructions define goals, behavior, constraints, and priorities for model execution.
TOOLS
Tools provide capabilities and descriptions models need for effective action.
RETRIEVAL
Retrieval supplies relevant external information when models actually need it.
MEMORY
Memory preserves useful information across interactions, tasks, and longer workflows.
HISTORY
History carries prior messages, actions, and results into subsequent reasoning.
COMPACTION
Compaction compresses accumulated context while preserving information needed for continuation.
SELECTION
Selection prioritizes useful information while excluding distracting or redundant context.
BUDGET
Budget allocates limited context capacity across competing sources of information.
JUDGMENT
Judgment determines what context models receive, retain, retrieve, or discard.
The Minimalist Context Engineering
You do not need to understand everything about context engineering before you begin.
Remember ten things:
Context Engineering.
Instructions.
Tools.
Retrieval.
Memory.
History.
Compaction.
Selection.
Budget.
Judgment.
Everything else can come later.
Define the goal.
Provide useful instructions.
Expose appropriate tools.
Retrieve information when needed.
Preserve useful memory.
Manage history.
Compact accumulated context.
Remove noise.
Protect the context budget.
Apply human judgment.
Then accomplish something.
One Final Question
The next time you work with AI, don't begin with:
“What is the perfect prompt?”
Begin with:
“What does the AI need to know right now?”
Then engineer the context.
