Planning your question
Before it reads a single paper, the agent turns whatever you typed — a one-word topic or a fully formed hypothesis — into a precise, answerable research question and a plan for working through it. This planning step happens automatically; you never have to operate it yourself.
#From a topic to a precise question
Most research starts vague. You might open with “cancer treatment”, “diabetes”, or “gene editing”. A literal search on a phrase that broad would return millions of loosely related papers and answer nothing. The agent’s first job is to sharpen the target: it reads your input, infers the biology you almost certainly mean, and reframes it as a specific scientific question with a clear subject, context, and outcome.
| You start with | The agent plans around |
|---|---|
| cancer treatment | Role of immune checkpoint inhibitors in triple-negative breast cancer resistance mechanisms |
| diabetes | Gut microbiome influence on insulin sensitivity in type 2 diabetes patients |
| gene editing | CRISPR-Cas9 off-target effects and mitigation strategies in therapeutic applications |
This refinement is not a rewrite you have to approve step by step. When your question is already specific, the agent keeps it and moves on. When it is broad or ambiguous, the agent narrows it on your behalf and states, in the brief it produces, exactly which question it answered — so you can always see the framing it worked from.
#Breaking a question into sub-questions
A good research question is rarely answerable in one pass. “How does the gut microbiome influence insulin sensitivity in type 2 diabetes?” quietly contains several distinct questions: which microbial species are involved, through which metabolites, measured by which markers, in which patient populations. The agent decomposes your question into these focused sub-questions and then works through each one in turn.
Decomposition is what makes the rest of the workflow reliable. Each sub-question:
- Gets its own evidence. The agent searches and reads the literature for that specific slice, rather than skimming everything at once.
- Is answered independently. Findings for one sub-question don’t bleed into another, which keeps contradictions and gaps visible instead of averaged away.
- Reassembles into the brief. The synthesized answer is built up from the sub-answers, so the structure of the final brief mirrors the structure of your question.
#Clarifying scope and terminology
Planning also fixes the boundaries of the investigation. The agent resolves the ambiguities that would otherwise scatter the search across unrelated fields, settling questions such as:
- Which entity do you mean? A gene symbol, protein, drug, or pathway can share a name with something unrelated; the agent anchors on the intended biological entity.
- What is the biological context? Species, tissue or cell type, disease model, and developmental stage all change which papers are relevant.
- What is in and out of scope? The agent decides whether to include preclinical models, clinical populations, or both, and where the question’s edges lie.
The more of this context you supply up front — the disease model, cell type, pathway, or organism — the less the agent has to infer, and the more targeted its reading becomes. But supplying it is optional: if you leave a detail out, the agent chooses a sensible default and records the assumption in the brief.
#Research emphasis: fundamental, applied, convergent
The same question can be investigated with very different priorities. “What is the role of interleukin-6 in rheumatoid arthritis?” looks one way to a basic scientist and another to a clinician. As it plans, the agent chooses an emphasis that shapes which sub-questions it generates and how it weighs the evidence.
| Emphasis | The question is framed around |
|---|---|
| Fundamental | Definitions and scope, molecular mechanism, experimental strategy, comparative and evolutionary context, and reproducibility of the underlying findings. |
| Applied | Clinical relevance, the therapeutic landscape, biomarkers and diagnostics, safety and toxicity, patient stratification, and standard of care. |
| Convergent | Cross-system relevance, bench-to-bedside translation, multi-omics integration, precision medicine, systems biology, and open future directions. |
You do not pick an emphasis from a menu. The agent reads it from how you phrase the question — a question about a signalling cascade leans fundamental, one about treatment outcomes leans applied, one that spans several fields leans convergent — and you can nudge it simply by making your interest explicit in the wording.
#Writing an effective question
The agent plans well from a bare topic, but it plans best when you give it something to anchor on. A few habits noticeably sharpen the result:
- Name the biology. Include the specific gene, protein, pathway, drug, or disease rather than a broad field.
- Set the context. State the organism, tissue or cell type, and disease model when they matter to you.
- Signal your intent. Wording that points at mechanism, at clinical outcomes, or at cross-disciplinary links tells the agent which emphasis to plan around.
- Ask one thing at a time. A single, well-defined question decomposes cleanly; the agent will still split it into sub-questions for you.
Compare “tell me about Alzheimer’s” with “What is the evidence that microglial TREM2 signalling drives amyloid clearance in Alzheimer’s disease, and how does that inform therapeutic targeting?” Both are valid inputs. The second simply hands the agent the subject, the context, and the emphasis at once, so the plan it builds is closer to what you actually want on the first pass.