1. Include what changes the answer

Context is not decoration; it is steering. The test for every piece of background: would the answer change if the model knew this? Your industry, the reader's seniority, the thing you already tried that failed, the constraint that makes the obvious answer wrong. Those change answers.

What does not change the answer dilutes it. Ten paragraphs of history the model must wade through compete with the two facts that matter.

2. Summarize aggressively, quote selectively

For long source material, you have three options: paste it all, summarize it, or quote the load-bearing part. Pasting everything works when the material IS the task (edit this document). Summarizing works when the material is background ('we sell B2B software to hospitals; sales cycles run long').

Quote exactly when the wording matters: the clause being disputed, the sentence being rewritten, the number being checked. A tight quote plus a one-line summary of the rest often beats both extremes.

3. Conversations drift; anchor them

In a long conversation the model weighs everything said so far, including the tangent you abandoned and the draft you rejected. Drift is normal: the fix is a periodic anchor. 'To reset: we are writing the board memo, the audience is non-technical, the decision is the March launch. Continue from the last draft.'

For a big change of direction, a fresh conversation with a clean summary usually beats fighting the accumulated history.

4. Stable context belongs in one place

Some context never changes: who you are, how you write, what you care about, the formats you prefer. Retyping it every conversation is the tax most people silently pay. Persistent instructions, a profile, pay that tax once.

The craft is separating STABLE context (who you are, put it in the profile) from TASK context (this document, this deadline, put it in the prompt). Mixing them bloats both.