Context Goes First, Not Third — And The Docs Put 30% On It

This card is an anatomy, and that word is doing real work. It is not a checklist of eight things to include; it is eight parts drawn down a vertical spine, numbered one to eight, with the colour deepening as you descend. Part 1 is Purpose, and its own description makes the ordering claim explicit: “One clear sentence gives Claude the whole picture before it reads anything else.”

So the card is not only saying what to include. It is saying what to put first. That is a checkable claim, because Anthropic publishes exactly one rule about where things go in a prompt — and it points the other way.

The one ordering rule in the documentation

From the prompting best-practices page, under Long context prompting:

“Put longform data at the top: Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples. This improves performance across all models.”

“Queries at the end can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs.”

Read that against the card’s spine. The card’s Parts 1 and 2 — Purpose and Task — are the query. Part 3 is the Context. The documented rule puts the context above the query; the card puts the query above the context.

When the card’s order is perfectly fine

This needs saying plainly, because it is the difference between a useful correction and a cheap one. The documented rule has a condition attached: it opens “When working with large documents or data-rich inputs (20k+ tokens)…”

Below that threshold — which is most prompts most days — nothing in the documentation says position matters, and the card’s order is a sensible way to write for a human reader: say why, say what, then fill in the detail. If you are asking for a paragraph of copy, write it in the card’s order and do not think about it again.

The rule earns its keep the moment you paste something in. A contract, a transcript, three competitor pages, last quarter’s export — at that point the card’s Part 3 stops being “context” in the background sense and becomes the bulk of the prompt, and it belongs at the top with the ask underneath it.

There is a second half of that same guidance the anatomy has no part for at all:

“Structure document content and metadata with XML tags: When using multiple documents, wrap each document in <document> tags with <document_content> and <source> (and other metadata) subtags for clarity.”

Eight parts, and none of them is “label which bit is which”. For a single-document prompt that is a nicety. For four documents it is the difference between an answer that cites the right file and one that blends them.

And the eight parts are really about five things

Lay the eight descriptions side by side and a cluster appears. Four of them are the same instruction with four names.

That is not padding, and the four are not identical — a stop condition and an output format are genuinely different objects. But if you are trying to remember an eight-part structure under time pressure, it is easier to remember five, with “describe an acceptable answer” as the big one that happens to have four faces.

Part 4 is the one you cannot type

Part 4 says: “Tell Claude how hard to think. ‘High effort’ means deeper reasoning and careful checks. Skip this and Claude picks a default level.”

Both halves are true and both are easy to misread. Effort is a real control with five levels, but it is a setting, not a sentence: /effort in Claude Code, output_config.effort on the API. Writing “high effort” in your prompt is English going to the model, not a parameter going to the request.

And “Claude picks a default level” is worth finishing: on most current models that default is already high. So for the majority of work, skipping Part 4 costs you nothing — the interesting move is usually stepping down for routine jobs, not demanding more. We went through the five levels and the per-model defaults in the piece on the effort ladder.

The anatomy, rebuilt in the documented order

Keep all eight ideas. Move three of them. This is the skeleton for the case the rule is about — a prompt with real material pasted into it.

What the card gets right

Three things, and the first one is the reason this piece is a correction rather than a takedown.

The fine print is unusually self-aware: “A simple framework, not an official format. Model names and behaviour change often, so test your prompts and check Claude’s official docs.” A card titled “The Anatomy of a Perfect Prompt” that then tells you it is not official and asks you to test is doing something most prompt cards never do.

Part 5 is the underrated one. “Say what not to do as clearly as what to do… Without them, Claude adds things you never asked for.” That matches guidance in the documentation about models that “overengineer by creating extra files, adding unnecessary abstractions, or building in flexibility that wasn’t requested”, with the recommended fix being to add specific guidance to keep solutions minimal. The card got there on its own.

And the whole premise is endorsed by the docs’ own golden rule:

“Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they’d be confused, Claude will be too.”

Eight labelled parts is a decent way to stop being confusing. The ordering is the bit to adjust, not the idea.

One line in the caption does need flagging: “Most prompts miss 5 of these 8 parts.” That is a statistic about other people’s prompts, and nobody is in a position to measure it. Read it as the rhetorical flourish it is, and judge the framework on the eight parts instead.

Four honest caveats

1. The 30% is the documentation’s number, not ours. It is reported as “up to”, from Anthropic’s own tests, “especially with complex, multidocument inputs”. We did not reproduce it and neither should you assume it on a short prompt.
2. Below 20k tokens, we are not claiming the card is wrong. The published rule simply does not speak to short prompts, and the card’s order reads well to humans.
3. “Four parts say the same thing” is a reading, not a measurement. Someone who writes specifications for a living would reasonably keep stop conditions and output format apart.
4. We did not A/B the two orders. This compares a card against published guidance. The card’s own footer asks you to test your prompts, which is the right answer for both of us.

The one-line version

It is an anatomy, so the order is part of the claim — and the one ordering rule Anthropic publishes says to put the long material at the top and the ask at the end, worth up to 30% on data-rich prompts, which is close to the reverse of Purpose, Task, Context.

Keep the eight parts. On a short prompt, keep the order too. The moment you paste a document, move it to the top and put the ask last.