How to Do Research with AI (Without Losing the Plot)
AI research rarely fails because the model gave you a wrong answer. It fails because you lost the plot: the tangent that felt important is buried forty messages up, you can't remember which claims you verified, and the conversation that started as "evaluate this decision" has drifted into something else entirely. The answers were fine. The structure collapsed.
That's not a discipline problem — it's a tool-shape problem. A linear chat forces every thought into one lane, so real research (which branches, doubles back, and compares) gets flattened until it stops being research. This post is the workflow that survives real questions: five moves, one worked example, and the honest limits.
Three ways linear chat kills a research session
1. Tangents are destructive. Every "wait — what about…" moment forces a choice: chase it and derail the main line, or drop it and hope you remember later. Both options lose something. Researchers have far more good tangents than they ever get to explore, and the scroll eats the evidence that they existed.
2. You can't audit what the AI knew. Twenty minutes in, an answer arrives that feels off. Was it responding to your framing from message 4 or message 24? In a scroll, context is invisible and cumulative — every earlier message silently shapes every later answer, and you can't isolate anything.
3. One model, unchecked. Serious research cross-examines sources. A single model answering in a single thread is one witness with no cross-examination — and models disagree with each other far more than most people expect.
The tree workflow: five moves
The fix is structural: run research as a tree instead of a scroll. Each move below maps to one thing trees do that scrolls can't.
- Branch every tangent instead of chasing it. When the "what about…" hits, open a branch and park the question there. The main line stays clean; the tangent stays alive. You triage tangents at the end of the session instead of mid-thought.
- Criticize in a side branch. Got a confident answer? Branch it and ask the model to argue the opposite, list failure cases, or name what evidence would change the conclusion. The criticism lives next to the claim — not twenty messages after it, silently poisoning everything in between.
- Cross-examine models at checkpoints. At each decision point, send the same node to a second or third model. Because branches inherit identical context, the disagreement you see is real disagreement — not an artifact of divergent conversations. More in comparing models side by side.
- Prune, then name what survived. Research produces mostly dead ends — that's the point of it. Archive the branches that didn't survive criticism and look at the map: the surviving path is your conclusion, with its reasoning attached.
- Export the surviving path. Turn the path from question to conclusion into a document — the deliverable keeps the chain of reasoning, not the noise. Why documents and branches are different animals: branches are where you think.
A worked example: "RAG or fine-tuning?"
Take a decision with real stakes: should your product build on RAG or fine-tuning? In a tree it looks like this:
- Root: the question, with your actual constraints — data freshness, team size, budget.
- Branch 1: the case for RAG. Sub-branch: "argue against this — where does RAG bite back?"
- Branch 2: the case for fine-tuning, criticized the same way.
- Branch 3: the same root question to a second model. It proposes a hybrid neither of you considered — and its clarifying questions expose which constraint actually drives the decision.
- Prune and export: the hybrid survives criticism; the map shows why. The export is a one-page decision memo with the losing arguments preserved as context.
Total time: about the same as a messy linear session. The difference is what you're left with — an auditable decision instead of a vague feeling that you discussed it once.
Can't I just do this with folders and multiple chats?
Partially — and people try. Separate chats per subtopic recreate the branches but lose the shared context: each new chat starts blind, and you become the courier pasting background between windows. Folders don't fix this because the problem isn't storage, it's structure. Even ChatGPT's own edit-a-message forking creates branches — but invisible ones you can't see or navigate. The workflow needs the tree to be a first-class object: visible, navigable, comparable, exportable.
Run one research question as a tree — KnowTree has a free tier. No API keys, no setup, nothing to install.