Why did the emphasis shift away from wording?

Because the ceiling was reached. Systems became far better at interpreting ordinary requests, which removed most of the advantage careful phrasing used to carry, while the effect of supplying the right material stayed exactly as large. The lever that still moves is the one that gets the attention.

Is more material generally better?

No, and this is the useful counter-intuitive part. Attention concentrates at the beginning and end of what is supplied, so a very large pile buries its middle. Cost also rises with every request rather than once. A smaller well-chosen set is often better and cheaper at the same time.

What makes a well-assembled request go wrong anyway?

Stale material, most often. The selection can be exactly right and the documents behind it out of date, which produces a confident answer built on something that stopped being true. That failure is invisible from inside the process, because nothing in the assembly is checking whether its sources still hold.

Where do permissions come into it?

Directly, because assembly decides what somebody effectively has access to. Material gathered for one person and shown to another is a disclosure regardless of how it got there, which is why enterprise products scope what they assemble to the asker's own access rather than treating that as a later filter.

Glossary

Context engineering

Context engineering is deciding what material should be in front of the system for a given task, which now decides the result more often than the wording of the request does.

In plain terms

Choosing what it gets to see. The instruction matters, and what is placed alongside it matters more: the right documents, the relevant history, the record it should be working from. Getting the wording perfect while handing over the wrong material is the commonest way for a result to disappoint.

Why it matters

Because effort spent on phrasing has a low ceiling and effort spent on material does not. A well-worded request against the wrong documents cannot produce a right answer, and a plainly worded one against the right documents usually does. Most teams discover this in the opposite order, after a period of trying to write better.

How it works

The assembly is the work, and it is mostly selection. What is placed in front of the system typically comes from several sources at once: standing instructions, retrieved passages, the history of the conversation, and structured records pulled from other systems. Deciding which of those to include, and in what proportion, is the activity the term names.

More is not better past a point, and the evidence for that is unusually clear. Attention concentrates at the beginning and end of what is supplied, so material in the middle of a very large pile receives less of it. Adding everything available is therefore not a safe default, and can produce worse answers than a smaller, better-chosen set.

The cost moves with it, which makes this an economic decision as much as a quality one. Everything supplied is charged on every request, so a generous assembly is paid for repeatedly. That is the same reason retrieval exists at all, and it is why the choice is rarely settled by quality arguments alone.

Freshness is part of the assembly and is easy to leave out of the design. What is put in front of a request is only as current as its source, so a system faithfully reporting a document nobody has updated will be confidently wrong in a way that looks exactly like being right. Nothing about the assembly notices that, because staleness is a property of the material rather than of the process.

Permissions belong here and are frequently forgotten. What is assembled for one person should not necessarily be assembled for another, and enterprise products treat this as a first-class concern rather than an afterthought: one in this guide answers with citations while showing each person only what their own access allows. Assembly without that becomes a disclosure question.

Two levers, and how far each moves

The shift is worth understanding as a consequence rather than a trend, because that determines whether it will reverse. Careful phrasing mattered most when interpretation was the weak point: a request had to be constructed to avoid predictable misreadings, and people who learned those constructions got noticeably better results. As interpretation improved, that advantage shrank on its own, without anybody deciding phrasing was passé. What did not shrink is the fact that nothing can answer from material it was not given. That constraint is structural rather than a limitation of the current generation, which is why the emphasis is unlikely to move back. The practical reading for a team is that time spent building a reliable way to put the right documents in front of a request will keep paying, while time spent perfecting wording pays less each year.

Seen in the wild

Enterprise search that answers with citations while showing each person only what their permissions allow, so what is assembled differs by who is asking.
Glean
A tool whose answers track whatever your own workspace currently says, which makes the selection of material the thing that decides quality.
Notion AI
Grounded briefings built from uploaded sources, where what was uploaded is the whole determinant of what comes back.
NotebookLM

Common misconceptions

People assume

It is prompt engineering under a newer name.

In fact

They address different levers. Wording shapes how a request is interpreted; assembly decides what there is to interpret. The second has the larger effect on most real tasks, which is why the vocabulary shifted, and the shift describes a genuine change in where the effort pays rather than a rebrand.

People assume

Supplying more material is the safe choice.

In fact

It is neither free nor reliably better. Everything supplied is charged on every request, and attention concentrates at the edges of a large pile rather than spreading evenly. A smaller, better-chosen set is frequently both cheaper and more accurate, which is an unusual combination worth taking advantage of.

Telling them apart

Context engineering vs Prompt engineering

Context engineering

What material is in front of the system.

Prompt engineering

How the request itself is worded.

Perfect wording against the wrong material fails. Plain wording against the right material usually does not.

Questions

Why did the emphasis shift away from wording?: Because the ceiling was reached. Systems became far better at interpreting ordinary requests, which removed most of the advantage careful phrasing used to carry, while the effect of supplying the right material stayed exactly as large. The lever that still moves is the one that gets the attention.
Is more material generally better?: No, and this is the useful counter-intuitive part. Attention concentrates at the beginning and end of what is supplied, so a very large pile buries its middle. Cost also rises with every request rather than once. A smaller well-chosen set is often better and cheaper at the same time.
What makes a well-assembled request go wrong anyway?: Stale material, most often. The selection can be exactly right and the documents behind it out of date, which produces a confident answer built on something that stopped being true. That failure is invisible from inside the process, because nothing in the assembly is checking whether its sources still hold.
Where do permissions come into it?: Directly, because assembly decides what somebody effectively has access to. Material gathered for one person and shown to another is a disclosure regardless of how it got there, which is why enterprise products scope what they assemble to the asker's own access rather than treating that as a later filter.

Key takeaways

Selection of material, not wording, is now the lever that moves most results.
More is not safely better: attention concentrates at the edges of a large pile.
Everything supplied is charged on every request, so generosity is paid for repeatedly.
Assembly decides effective access, which makes permissions part of the design.