AI Systems/AI & reliability

Context Engineering

7 micro-topics in AI Systems, each with the evidence that proves you have it and the written reason behind every link. Before you start, it rests on 3 other domains. Downstream, it holds up 3 domains.

7 topics · depth 4–7 of 16 · 7 internal links · 4 in · 3 out

Prerequisites · 4 links in

Before you start

Load-bearing links first. Each one says what in this domain rests on what outside it, and why.

Model Foundations

2 links

load-bearing

Treat the context window as a budget rests on Explain the context window

A budget requires a limit.

load-bearing

Treat the context window as a budget rests on Compute token economics

The budget has money and latency on its axes.

Model Integration

1 link

load-bearing

Carry structured state instead of transcripts rests on Maintain chat context

You replace history replay with state.

Prompt Engineering

1 link

supporting

Treat the context window as a budget rests on Structure prompts for reliable parsing

Prompt structure is the first thing that fills the window.

7 topics · depth 4–7

What you will be able to do

In prerequisite order. Each idea names the artifact that closes it, the market demand that put it on the map, and what it stands on.

Depth 4 · conceptual · unassisted · M18 §18.1

Treat the context window as a budget

Allocates tokens against cost, latency and accuracy

Evidence: Cuts token spend on a working endpoint with no eval drop
Market anchor: Context engineering
Rests on: load-bearing
Explain the context window · Model Foundations
A budget requires a limit.
load-bearing
Compute token economics · Model Foundations
The budget has money and latency on its axes.
supporting
Structure prompts for reliable parsing · Prompt Engineering
Prompt structure is the first thing that fills the window.

Depth 5 · conceptual · scaffolded · M18 §18.1

Account for position effects

Knows the middle of a long context is weakest

Evidence: Moves critical instructions and measures the difference
Market anchor: Context engineering
Rests on: load-bearing
Treat the context window as a budget
Placement matters once you are deciding what goes in.

Depth 5 · conceptual · scaffolded · M18 §18.1

Recognize context rot

Knows long histories degrade rather than accumulate

Evidence: Identifies degradation across a long session
Market anchor: Context engineering
Rests on: load-bearing
Treat the context window as a budget
Rot is what happens when the budget is spent badly over time.

Depth 5 · procedural · scaffolded · M18 §18.1

Cache reusable context

Reuses a stable prefix across calls

Evidence: Cost per request drops measurably
Market anchor: AI cost management
Rests on: load-bearing
Treat the context window as a budget
Caching is a cost lever on the budget.

Depth 6 · conceptual · scaffolded · M18 §18.1

Carry structured state instead of transcripts

Replays a compact state object, not raw history

Evidence: Replaces transcript replay and holds quality
Market anchor: Context engineering
Rests on: load-bearing
Recognize context rot
Structured state is the answer to transcript growth.
load-bearing
Maintain chat context · Model Integration
You replace history replay with state.

Depth 6 · meta · scaffolded · M18 §18.1

Decide what to evict

Chooses what leaves the window when it is full

Evidence: Defends an eviction policy against a failure case
Market anchor: Context engineering
Rests on: load-bearing
Account for position effects
Eviction decisions depend on what position costs.
supporting
Compress and summarize history
Compression and eviction are the same tradeoff at different intensities.

Depth 7 · procedural · scaffolded · M18 §18.1

Compress and summarize history

Keeps meaning while dropping tokens

Evidence: Halves history tokens without losing task state
Market anchor: Context engineering
Rests on: load-bearing
Carry structured state instead of transcripts
Compression assumes a representation to compress.

Unlocks · 3 links out