Three weeks ago I had a 412-page planning dossier, four competing vendor contracts and a client call in ninety minutes. The sensible move was to skim, panic, and bluff. Instead I dropped the entire pile into Kimi AI and typed one question. It came back with a comparison table and clause numbers attached.
That is the thing about Kimi. It isn’t a nicer chatbot, it’s a different shape of tool. Most assistants quietly truncate your input and hope you don’t notice. Kimi was built around enormous context windows, and that changes what you can reasonably ask it to do. What follows is the practical version: the actual steps, the prompts that work, and the habits that turn it into something you’d use on an ordinary Tuesday afternoon.
Step 1: Know what you’re actually working with
Kimi is made by Moonshot AI, and its headline feature is context measured in the hundreds of thousands of tokens. That’s enough to hold whole books, legal bundles or a mid-sized codebase in one conversation without the model forgetting page 12 by the time it reaches page 300. If you want the mechanics before you start feeding it files, this breakdown of how Kimi holds a million tokens of memory is a good twenty-minute primer.
Three practical facts follow from that:
- It reads far more than it writes. You can upload 400 pages, but you’ll get a better answer if you cap the output at a table or a 500-word memo.
- Web search and file upload behave differently. Searching pulls in live pages and can muddy a document analysis. Keep the two modes separate.
- Free tier first, then metering. Heavy multi-file uploads are what burn through quota, so batch them deliberately.
Kimi also isn’t the only option. Moonshot sits in the same long-context race as Alibaba’s Qwen, and that contest is part of a wider story about China’s push against American AI dominance. Which is a roundabout way of saying: if Kimi’s limits annoy you, the alternative is one tab away.
Step 2: Upload the source, not your summary
The most common mistake is pre-digesting. People paste a two-paragraph summary of a report and ask Kimi to analyse it, which throws away the entire reason to use Kimi in the first place.
A real example. I had a 60-page commercial lease and needed to know what happens when a tenant wants to sublet. Rather than read it, I uploaded the original PDF and asked:
“Find every clause that affects subletting or assignment. For each one, give me the clause number, the page, the requirement, and the notice period. If two clauses conflict, point that out rather than picking one.”
It returned five clauses, three of which cross-referenced each other in a way a keyword search would never have surfaced. Total elapsed time: about ninety seconds.
Step 3: Ask for structure, not prose
Kimi defaults to helpful paragraphs. Paragraphs are hard to verify and harder to reuse, so force a shape onto the answer.
Compare these two prompts on the same annual report:
- Weak: “What are the main risks in this report?”
- Strong: “List every risk factor mentioned. Columns: risk name, page, stated likelihood, stated impact, and whether management gives a mitigation plan. If a factor is mentioned twice, merge the rows and note both page numbers.”
The first gets you fluent generalities. The second gets you a table you can sort, filter and paste into a deck.
The page-number rule
Always ask for citations back to the source: clause numbers, page numbers, section headings, timestamps. You’ll spot-check maybe 10% of what comes back, and exact locations turn that into a two-minute job rather than a half-hour one. When Kimi can’t cite something, treat that as a signal the answer was generated rather than extracted.
Step 4: Put several documents side by side
Single-document analysis is table stakes. The real payoff is comparison: four vendor contracts, three research papers, two drafts of the same policy.
Name your files before uploading. “Contract A (Acme, 2023)” beats “scan_0042.pdf” when you’re asking which agreement has the weakest data-breach notification window.
A prompt that earns its keep: “Compare all four contracts on three points: termination notice, liability cap, and breach notification. One row per contract, one column per point, plus a final column flagging anything unusual.”
That produced a table I dropped straight into a client email. The flagging column caught a liability cap sitting roughly ten times below the other three, something I would probably have missed reading them in sequence over two evenings.
Step 5: Turn the analysis into the deliverable
Don’t stop at the table. Ask Kimi to draft the thing you actually owe someone: “Using only the table above, write a 400-word memo for a non-lawyer explaining which contract I should sign and why. Plain English. Flag the two biggest risks in bold.”
You’ll still edit it. You’ll still own the judgement call. But editing a draft is a different job from writing one, and it’s usually a much faster one. If you’d rather run the same pattern in a different assistant to compare output quality, this walkthrough on getting real work out of Qwen Chat follows a similar rhythm.
Step 6: Keep one thread alive per project
Starting a fresh chat for every new document means re-explaining context over and over. Instead, keep a single conversation per project and add files as they arrive. Kimi retains the earlier material, so a question in week three can still draw on the spec you uploaded in week one.
The caveat: very long threads drift. Once a conversation gets genuinely enormous, the model’s grip on an early detail loosens. When precision matters more than continuity, open a clean thread and re-upload the two files that matter.
Where Kimi still falls over
- Arithmetic across big tables. It will occasionally add a column wrong. Re-check any total that feeds a decision.
- Page numbers on scanned PDFs. Poor OCR makes citations unreliable. Test with one file before trusting a hundred.
- Confident invention in obscure corners. Footnotes and appendices are where fabrications hide.
- Autonomy. Anything that clicks, buys, sends or deletes on your behalf should keep a human in the loop, and the case for that caution is made well in this piece on why rogue AI is no longer science fiction.
The four-part prompt I reuse every time
Every long-document task now starts the same way. Paste this and swap the brackets:
“Attached: [files, named clearly]. Task: [one specific question]. Output format: [table / memo / bullet list], maximum [length]. For every claim, cite [clause number / page / timestamp]. If the documents don’t answer the question, say so and tell me what’s missing.”
That last sentence does more work than the rest combined. It converts Kimi from a machine that always produces an answer into one that occasionally tells you the honest thing: this isn’t in here. Twenty minutes of practice with that template is usually enough to make the difference obvious, and it costs nothing but a little patience while you learn where the tool’s edges are.