Kimi Long-Context Prompt Builder

Order matters: document first, instructions after

Moonshot's Kimi is built for length. Its standout strength is a very long context window, which lets you feed in an entire contract, a full research report, a long paper, or several files merged together and then ask questions across all of it at once. But a long window only helps if the prompt around it is structured for long text. The single most useful habit is ordering: put the document first and the instructions after. When the model reads the full source before it sees "now do this with it", it locates the relevant passages more accurately and skips less than when the task is buried at the top above a wall of text. This builder enforces that order for you — your pasted document sits at the very top, and the task, output format and constraints follow, with a fixed grounding tail appended automatically.

After ordering, the second lever is the task itself. Long-document work falls into four common shapes: summarise the whole thing, extract specific points, answer questions, and rewrite or polish. Naming which one you want — and being concrete about it ("summarise the milestones in chronological order", "list every payment clause and its amount", "answer only using section 3") — tells the model what to pull out of a large body of text instead of guessing. It also helps to ask the model to first map the document — list its headings or structure — before answering. That gives it an internal map of a long source, so its citations and its sense of where things live become much sharper. Remember too that long context is not infinite context: anything past the window is truncated, so the most critical material belongs near the start or end, and named explicitly in the task.

Citation and "say when it is missing" make long-text answers trustworthy

The constraints are where long-context prompting earns its reliability. Two lines do most of the work. First, ask the model to quote the relevant passage or sentence, and to cite the page, heading or location. Quoting ties every claim back to the source, lets you verify in seconds against the original, and pushes the model to answer from the material rather than from a hazy impression of it — which is exactly what you want for contracts, filings, regulations and anything where traceability matters. Second, tell the model to say plainly "not stated in the document" when an answer is not there, instead of inventing one. That single instruction is the strongest defence against confident hallucination over a long source: when the material genuinely lacks the information, you get an honest gap rather than a fluent, unsupported paragraph. This builder appends both of these as a standing tail so you never forget them.

Because the output is structured plain text, the prompt is portable: it works whether you paste the document straight into Kimi's chat box or upload the file and paste only the instruction part, and the same structure carries over to other long-context models. Write the task in the language you want the answer in — the document can be in any language. And because the whole tool runs locally in your browser, you can iterate freely on long, sensitive material: tweak the task, copy again, and test, without anything you paste ever leaving your device, being sent to a model, or being stored. Treat the first prompt as a draft — run it, see where the answer drifts or over-reaches, then tighten the matching field. A round or two of that usually turns a vague long-document summary into precise, cited, verifiable answers, and you keep a clean, reusable prompt at the end.