How do I decide what goes in the context window?

Keep the current task, the rules that constrain it, the evidence that would change the answer, and the output contract. Compress bulk that still holds those facts into labeled spans with provenance. Drop unlabeled dumps, duplicate retrievals, and history that belongs in a store. The Packing List at curatedcontext.com/packing-list/ is the desk for that keep / compress / drop decision.

What is context engineering?

Context engineering is the craft of building the information environment around an LLM: task, instructions, retrieved evidence, memory, examples, constraints, and output contracts. Curation — choosing the working set — is the hard part, not filling the window.

What is The Packing List?

The Packing List is Curated Context’s signature desk. You pack a job — invoice dispute, failing test, policy exception, or paper comparison — by putting each candidate in keep, compress, or drop. It is model-neutral and does not copy Claude context-window sizes.

Is RAG or long context better?

RAG selects from a large corpus. Long context holds a known bundle. Hybrid systems retrieve candidates, then pack a structured working set. Neither advertised window size nor dumping more tokens is a strategy.

Model-neutral context engineering

Choose what the model gets to see.

How do you decide what goes in the context window? Keep the current task, the rules that constrain it, the evidence that would change the answer, and the output contract. Compress bulk that still holds those facts. Drop unlabeled dumps and history that belongs in a store. The Packing List is the desk for that: keep, compress, or drop.

Open The Packing List Start With What Belongs

Goal

The user task, success criteria, and decision boundary.

Evidence

Sources ranked by relevance, authority, freshness, and conflict.

Rules

System constraints, policies, formatting contracts, and tool limits.

Memory

Stable preferences and prior decisions, not every historical token.

Output

The shape the answer must take so it can be judged and reused.

Key facts

Context Engineering Guide

Context engineering is the design of the full inference payload: task, instructions, evidence, memory, tools, examples, constraints, and output contract.
The central rule is curation over dumping: put answer-critical material in the prompt, citations in the audit trail, and everything else behind retrieval.
The Packing List at /packing-list/ is the signature desk: keep, compress, or drop each candidate before you fill the window.
The site is model-neutral and independent. Claude-specific context-window numbers belong at claudecontext.com, cited and linked.

The signature desk

The Packing List

Four jobs. A pile of candidates. Three bins. Pack an invoice dispute, a failing test, a policy exception, or a paper comparison and watch the working set become a brief instead of a warehouse. Claude-specific window sizes live onclaudecontext.com; this desk stays model-neutral and cares what you selected.

Pack the invoice job See every job

Working set

Current question
Billing rules that govern plan changes
Plan-change ticket from August 3

Compressed spans

Twelve-page invoice PDF

Left behind

Last forty support tickets
Similar billing FAQ pages from retrieval
Full CRM account timeline

Why this exists

Curation beats dumping because attention is not uniform.

Long context windows changed what is possible, but they did not remove the need for selection. Research on long-context use repeatedly shows that placement, length, and task complexity change how well models use evidence.

"performance is often highest when relevant information occurs at the beginning or end"

Lost in the Middle: How Language Models Use Long Contexts

Guide map

The context pipeline, page by page.

Each page answers one operational question. Read them in order when designing a new system, or jump to the failure mode you are debugging.

What Belongs in the Context Window

A practical rubric for deciding which goals, instructions, evidence, examples, metadata, and constraints belong in an LLM context window.

RAG vs Long Context

An evidence-backed comparison of retrieval augmented generation, long context prompting, context rot, and hybrid context pipelines.

Context Compression

How summarization, chunking, prompt compression, and hierarchical memory reduce tokens without throwing away the facts that matter.

Structuring Context

Ordering, delimiters, source cards, priority rules, and output contracts that make curated context easier for LLMs to use.

Evaluating Context Quality

How to measure whether curated context improves retrieval, faithfulness, answer relevance, latency, and downstream task success.

Context Window Management

A practical workflow for managing context-window budgets, drift, overflow, and refresh decisions in LLM systems.

Context Engineering Glossary

Plain-language definitions for context engineering, context curation, RAG, context rot, prompt compression, memory, and related terms.

Sources and Machine-Readable Files

Research sources, primary papers, and the llms.txt files behind Curated Context.

Context Engineering FAQ

Answers to common questions about context engineering, RAG, long context, compression, context rot, and evaluation.

Free tools

Estimate the window after you pack it.

After the packing is chosen, size it. The tools run entirely in the browser. Use them to estimate tokens, reserve output room, and see where a budget is going.

ESTIMATEToken CounterPaste text to estimate tokens with a transparent four-characters-per-token heuristic and compare the result against common context windows.PLANContext PackerPlan a context window by allocating tokens to system instructions, history, retrieved evidence, examples, memory, and reserved output.

Model-neutral, source-backed

This is about context, not one vendor's prompt style.

Curated Context complements claudecontext.comby staying model-agnostic. Claude-specific context-window numbers belong there, cited and linked; they are not packing rules here. The pages cite research on RAG, long-context behavior, prompt compression, memory systems, and evaluation so the guidance can travel across model families and tooling stacks.

Cite This Page

Curated Context. "Context Engineering Guide." Accessed July 12, 2026. https://curatedcontext.com/

https://curatedcontext.com/