What It Is
Terminology · LLM Engineering · Language Models · Benchmarking · AI · ExplainerTratopedia · 15 Aug 2026
Second of five · the layer around the prompt
“All the Context” Lasted Six Days
On 19 June 2025 Shopify's CEO said context engineering was the art of providing all the context for a task to be plausibly solvable. Six days later Andrej Karpathy endorsed the term and quietly changed it: filling the window with just the right information for the next step. Three weeks after that, Chroma measured eighteen models and found that performance gets steadily less reliable as input grows. By the end of September, Anthropic's definition had inverted the original — find the smallest possible set of high-signal tokens. In about a hundred days the term went from a question about whether you had given the model enough, to a question about whether you had given it too much.
Four definitions, in fifteen weeks Put side by side, they do not say the same thing. The difference is the article.
- 6days between the term being endorsed and its meaning being narrowed
- 102days from “all the context” to “the smallest possible set”
- 18models Chroma tested, finding none of them uses its context uniformly
- 0Wikipedia articles for the term at capture — it is a paragraph inside another entry
Tobi Lütke · 19 Jun 2025
“Providing all the context”
- In full: the art of providing all the context for the task to be plausibly solvable by the LLM.
- Note “the task” — singular, and done once.
- His own words are “I really like the term”, which reads as endorsement rather than coinage. Nothing consulted names an earlier user either.
Andrej Karpathy · 25 Jun 2025
“Just the right information for the next step”
- “+1 for context engineering over prompt engineering. People associate prompts with short task descriptions… when in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window…”
- From “the task” to “the next step”: from something done once per job to something done once per turn.
- What was retrieved here stops mid-sentence. The ending — “with just the right information for the next step” — is quoted by Anthropic, which links to this post. Corroborated, not read whole.
Chroma · 14 Jul 2025
Not a definition — a constraint
- The assumption they test: a model “should handle the 10,000th token just as reliably as the 100th”. Their finding: it does not hold.
- Models “do not use their context uniformly; instead, their performance grows increasingly unreliable as input length grows” — across GPT-4.1, Claude 4, Gemini 2.5 and Qwen3.
- It lands between the coinage and the formalisation. That is suggestive, and this page does not treat it as proven cause.
Anthropic · 29 Sep 2025
“The smallest possible set of high-signal tokens”
- Formally: the set of strategies for curating and maintaining the optimal set of tokens during LLM inference, including all the other information that may land there outside of the prompts.
- It also makes the cadence explicit: “the curation phase happens each time we decide what to pass to the model.”
- A vendor's definition, and the only formal one here. anthropic.com is blocked to this session — the text was supplied by djTratoh.
How firmly each part stands Almost all of it was read at source. Two things were not, and one was calculated.
| How firm | What | On what |
|---|---|---|
| Read at source | Lütke's wording: “the art of providing all the context for the task to be plausibly solvable by the LLM” | His post on X, fetched 15 Aug 2026 |
| Read at source | Karpathy's wording, as far as “filling the context window” | His post on X, fetched 15 Aug 2026. The retrieved text is truncated there |
| Read at source | Chroma tested 18 LLMs; models “do not use their context uniformly”; performance “grows increasingly unreliable as input length grows” | Context Rot, Chroma technical report, 14 Jul 2025, fetched at source. Only the opening was read — no individual experimental result is quoted here |
| Read at source | There is no Wikipedia article for context engineering; it is a paragraph inside “Prompt engineering” | A 404 at capture, 15 Aug 2026. An absence is a finding: the encyclopaedia still treats it as a sub-topic |
| Derived | The two posts are six days apart — 19 June 03:01 UTC and 25 June 15:54 UTC 2025 | Calculated from the posts' own IDs, which encode a millisecond timestamp. A derivation, not a reading — repeatable, but not the same as seeing a date on the page |
| Supplied, host blocked | Anthropic's formal definition, the attention budget, the n² attention argument, and the three long-horizon techniques | Effective context engineering for AI agents, 29 Sep 2025. anthropic.com returns 403 to this session for every tool; djTratoh supplied the text |
| Our reading | The term inverted: a sufficiency test became a scarcity test | Assembled from the four definitions above. No source consulted remarks on it |
| Our reading | Chroma's measurement is a plausible cause of the narrowing | It falls between coinage and formalisation, and Anthropic cites it — but nothing says the definition changed because of it |
Timeline
A hundred and two days The tightest chronology of the five terms — four documents, and a measurement in the gap.
- Dec 2024The work, without the word. Anthropic describes the “augmented LLM” — a model plus retrieval, tools and memory — as the building block of everything agentic. All of it is context engineering; none of it is called that yet.
- 19 Jun 2025All the context. Tobi Lütke posts that he prefers “context engineering” to “prompt engineering” because it describes the core skill better: providing all the context for the task to be plausibly solvable.
- 25 Jun 2025Six days later, just the right context. Karpathy adds his +1 — and reframes it as filling the window with just the right information for the next step. The word “all” does not survive the week.
- 14 Jul 2025Someone measures it. Chroma publishes Context Rot: eighteen models, and none of them handles the ten-thousandth token as reliably as the hundredth. Performance grows steadily less reliable as input grows.
- 18 Sep 2025The loop gets a definition. Simon Willison settles on “an LLM agent runs tools in a loop to achieve a goal”. That loop is what keeps producing the tokens somebody now has to curate.
- 29 Sep 2025The smallest possible set. Anthropic gives the term its first formal definition — curating and maintaining the optimal set of tokens during inference — and states the goal as the smallest set of high-signal tokens. A hundred and two days after “all”.
The Argument
Why a definition would invert in a hundred days Because the first one was a hope and the second one had been measured.
Lütke's definition is a sufficiency test. Have you given the model everything it needs? If not, the task is not plausibly solvable, and that is your fault rather than the model's. It is a generous, practical idea and it points at a real failure — an agent that cannot see the file cannot use it. Anthropic's is a scarcity test, and it points at the opposite failure. Context is finite, models have an attention budget, and every token spends some of it. The architecture explains why: a transformer lets every token attend to every other, so n tokens produce n² pairwise relationships, and attention gets stretched thinner as the window fills. The result is a gradient rather than a cliff — models stay capable at length but lose precision. What happened between the two is that somebody measured it. Chroma tested eighteen models against the assumption that the ten-thousandth token is as reliable as the hundredth, and reported that it is not — not in one model, in all of them. Once that is on the record, “provide all the context” stops being generous and starts being a way to degrade your own system. This page does not claim Chroma caused the change. The report sits between the coinage and the formalisation, and Anthropic cites it; that is suggestive and it is all that can be shown. What can be shown is the wording: “all the context” on 19 June, “just the right” on 25 June, “the smallest possible set” on 29 September.
- all → leastthe same term, a hundred and two days apart
- 18models in which context is not used uniformly, per Chroma
- n²pairwise relationships from n tokens — why attention thins as the window fills
What Others Add
What you actually do about it Anthropic's own practice, which is the most specific account available.
- 1Write at the right altitudeBetween hardcoded brittle if-else logic and vague guidance that assumes shared context. Aim for the minimal set that fully specifies the behaviour — and note that minimal does not mean short.
- 2Keep the tool set smallTools should be self-contained, robust to error and minimally overlapping. The test is blunt: if a human engineer cannot say which tool to use in a given situation, the model cannot be expected to do better.
- 3Fetch late, not earlyKeep lightweight identifiers — file paths, queries, links — and load data at runtime. Metadata is itself signal: a file called
test_utils.pyintests/means something different from the same name insrc/core_logic/. - 4For long jobs, throw things awayThree techniques: compaction (summarise a full window and restart from the summary), structured note-taking (write notes outside the window and read them back), and sub-agents (explore with tens of thousands of tokens, return one to two thousand).
The three long-horizon techniques, compared Anthropic's own guidance on which fits what.
| Technique | What it does | Best for |
|---|---|---|
| Compaction | Summarise a window nearing its limit and reinitialise from the summary. Claude Code keeps architectural decisions, unresolved bugs and implementation details, discards redundant tool output, and carries on with the five most recently accessed files | Tasks needing extensive back-and-forth |
| Structured note-taking | The agent writes notes to memory outside the context window and pulls them back later. A to-do list, or a NOTES.md | Iterative development with clear milestones |
| Sub-agents | Specialised agents work in clean windows and return condensed summaries — often 1,000–2,000 tokens from tens of thousands explored | Research and analysis where parallel exploration pays |
One point deserves separating from the vendor's framing, because it is the part most likely to age. Anthropic's advice ends with “do the simplest thing that works” and an expectation that smarter models will need less curation — that agentic design will “trend towards letting intelligent models act intelligently, with progressively less human curation”. That is a prediction, made by a party with an interest in it being true, and it should be read as one. What is not a prediction is Chroma's measurement, made by a different organisation, across eighteen models including three vendors' flagships. If the models were already handling long context uniformly, that report would say so. It says the opposite, and it says it about all of them. The two can both be right: curation may get easier while remaining necessary. But only one of them has been measured.
- 1independent measurement in this article. Everything else is a definition or a practice
- 1–2ktokens a sub-agent returns from tens of thousands explored
- 403what anthropic.com returns to this session. Its post was supplied, not fetched
Conclusion
What to take away Three things, and the first is a warning about quoting the term.
Ask which definition someone means
“Provide all the context” and “find the smallest possible set of tokens” are opposite instructions, and both have been published under this one term within a hundred and two days of each other. Somebody saying “context engineering” in 2026 could mean either. The word does not disambiguate itself.
The one independent measurement is the part to trust most
Three of the four definitions here come from people with something to sell or a position to advance. Chroma's report does not define the term at all — it tests whether models use long context uniformly, across eighteen of them, and finds they do not. That is the load-bearing fact, and it comes from the one party not defining anything.
Nobody consulted claims to have coined it
Lütke wrote “I really like the term” — endorsing something, not naming it. Karpathy wrote “+1”. Neither reads as an origin, and no earlier user turned up in anything consulted here. The term is routinely attributed to them; on the evidence available, they popularised it.