Stop prompting from memory. Upload first, ask second.

When the answer comes back generic, it almost always means the model had nothing specific to read. Someone opens a chat, types 'what are the risks in vendor contracts like ours,' reads a tidy list that could have come from any business blog, and closes the tab convinced AI is overhyped. They asked a specific question and got a generic answer, so the model must be mid. Wrong diagnosis: they asked a specific question while giving the model nothing specific to read.

An AI model answering a cold prompt is working from the averages of everything it was trained on. It has never seen your vendor contract, your customer transcripts, or your competitor's pricing page. Describe those things in a sentence and it will guess politely. Hand it the actual documents and it stops guessing, because now the answer comes out of your material instead of the world's median take on the topic.

This is a different failure than the copy-paste commute. The commute is about where AI sits in your workflow: shuttling text between tabs instead of bringing work into the tool. This one is about briefing technique. You can have AI sitting right inside your workflow and still starve it, asking questions from memory when the source material is one drag-and-drop away. Knowing how to give AI context is the skill that separates the two outcomes.

Once the sources are in, stop asking open questions and start assigning synthesis, then audit what comes back. 'Compare what these vendor proposals actually promise.' 'Where do the customer transcripts contradict what sales is telling us?' 'What's missing from this report that a board member would ask about?' 'Read all of this as a skeptic and argue against our plan.' These prompts only work when there is material to work on. Cold, they produce filler. Grounded, they produce the kind of reading you'd pay an analyst for.

One amendment before you drag in the whole drive: upload first does not mean upload everything. Chroma's context rot research ran 18 frontier models through progressively longer inputs and found that accuracy degrades as context grows, that even a single distracting passage measurably hurts performance, and that models answered focused prompts better than everything-included ones. Knowing how to give AI context includes knowing what to leave out. Six documents that bear on the question beat sixty that mostly don't.

Then make the output auditable. Require a citation for every claim: which document, which section. This does two things. It forces the model to stay anchored to the sources instead of drifting back to its training data, and it gives you a checklist for the last step. Open the originals and verify the citations that matter most. Grounding AI in real documents cuts fabrication sharply, but it doesn't erase it. When Stanford researchers audited the legal research tools built on exactly this kind of document grounding, the tools still hallucinated on 17% to 33% of queries, well below the rates general chatbots show on legal questions and nowhere near zero. A confident summary with one invented claim is worse than no summary at all. The spot-check is the price of trusting the rest.

None of this requires new tools, and the mainstream ones now assume you'll work this way. ChatGPT and Claude both offer projects that keep a set of sources attached across conversations, Gemini reads a stack of files in a single prompt, and Google built NotebookLM around nothing else: it answers only from your uploads and links each claim to the exact passage. What's scarce is the habit, and habits come from reps on real work, which is how AI fluency actually forms. If research is a regular part of your job, the AI for research path goes deeper on this workflow, and how to learn AI covers where briefing technique sits in the broader skill stack. Upload first, ask second. Everything else in AI-assisted research builds on that order.