Give AI enough context for a long-form draft without flooding the prompt

Separate permanent context from task context

Permanent context changes slowly: the intended reader, the project's central claim, voice preferences, source policy, terminology, and non-negotiable constraints. Task context changes every time: the selected passage, its place in the outline, the current revision goal, and what must remain untouched.

Keeping the two layers separate stops prompts from becoming giant diaries. Save the permanent layer as a one-page project card. Attach only the relevant task layer to each request. When the argument changes, update the card once rather than correcting the same assumption in ten chats.

Set a context budget before attaching material. Give first priority to the current passage, the section's job, and the facts it may use. Give second priority to the project promise and voice rules. Supply older drafts only when comparison is the task. This ranking is more useful than filling an arbitrary token limit because it protects the information that changes the answer and removes history that merely happens to exist.

Write a six-part project card

Use six headings: reader, outcome, central claim, evidence boundary, voice, and constraints. Under each, write one to three lines. The evidence boundary should say what the assistant may treat as factual - For example, only attached interview transcripts and the three named reports - And what it must mark as unknown.

The voice section should describe choices rather than adjectives. ‘Use concrete nouns, short openings, and qualified conclusions’ is actionable. ‘Sound compelling and human’ is not. Constraints include length, required terminology, reading level, publication rules, and claims that need specialist review.

A practical project card can fit in twelve lines. It might name an audience of first-time nonprofit directors, an outcome of choosing a reporting method, a claim that simple measures beat unused dashboards, and a source boundary of two interviews plus one internal dataset. Add three voice rules and three prohibitions. If a collaborator cannot understand the card in a minute, it probably contains notes rather than settled direction.

Give every section a job

Before requesting prose, state where the passage sits and what changes for the reader by its end. A section job might be ‘show why the existing explanation fails, using the customer interview, then hand off to the new model’. This is more useful than a topic label because it defines movement.

Include the sentences immediately before and after the selection when transitions matter. For a structural task, provide the whole sentence outline instead of the whole manuscript. The aim is enough surrounding shape to avoid repetition and contradiction, without burying the operation in unrelated material.

Give the section a small contract: starting point, required movement, evidence, and handoff. For example, the reader begins believing more metrics mean better oversight; the section shows the maintenance cost of weak measures; it must use Interview C and end by introducing a three-measure test. That contract exposes repetition because two sections cannot quietly perform the same movement without someone noticing.

Name the operation and the acceptance test

Use a verb that describes one kind of change: diagnose, reorder, compress, expand from supplied notes, identify objections, or produce alternatives. Avoid ‘improve’, which silently combines structure, style, fact, and voice into a result you cannot evaluate cleanly.

Then say how you will judge the result. An acceptance test can be simple: preserve all three sourced claims, cut 20 percent, make the objection explicit before the response, and do not add facts. A result either passes those tests or becomes material for another pass.

Separate generation from judgment. One request can propose three orders for the supplied material. The next can test the chosen order against the brief. A final request can compress the accepted version without changing claims. Each stage has a visible input and decision. If you ask for an authoritative, concise, well-structured rewrite in one move, you cannot tell which requirement caused a useful fact to vanish.

Context is not every file you own. It is the smallest set of decisions and evidence that prevents the wrong answer.

Treat sources as a closed set

If accuracy matters, tell the assistant to use only the material you supplied and to label gaps rather than fill them. Give sources short names and, where possible, locators. Ask it to place a source key beside every factual addition so verification does not become a memory exercise later.

A closed set does not guarantee accuracy; it prevents invisible research from entering the draft. You still open the cited passage and check it. But the provenance remains visible, which makes that check possible.

Label source excerpts with durable keys such as S1 and S2, followed by page or timestamp. Ask the assistant to place a key after every source-dependent sentence and to write [GAP] where the closed set cannot answer. Remove those working marks only after checking the prose. This simple convention separates supplied evidence from connective language and makes accidental invention much easier to spot.

Refresh context at real turning points

Do not paste the same packet forever. Refresh it when the thesis changes, a new audience is chosen, an editor settles terminology, or evidence overturns an earlier assumption. Add a short decision log so you know why the card changed.

At the end of a session, save what the assistant was told along with the accepted result. Weeks later, you can distinguish a bad transformation from a prompt that accurately followed an obsolete brief. That distinction is essential in a project that outlives one conversation.

Keep a short change log under the card: date, decision, reason, and sections affected. When the intended reader changes, the log tells you which old outputs are now suspect. When only a headline changes, the permanent context can stay put. Treating every edit as equally important creates needless re-prompting; recording consequential decisions keeps the context current without rewriting it after every session.