Whatever you're working on, start in the Research Agent. It helps you sharpen your question, check what evidence exists, and draft your prompt before you commit to a workflow.

When a Systematic Review is the better fit, use the agent to get the setup right before you run anything: refine the research question, sharpen the screening criteria, and work out the extraction columns. Then start the review yourself in the Systematic Review workflow, where you see and approve each step.

A Research Report also has to be started yourself: the agent will not suggest one and cannot launch one.

"I need to understand the patient population for an upcoming launch in moderate-to-severe atopic dermatitis. Suggest three research prompts I could run and explain what each would give me."

Using the Research Agent to plan your review before you run it is best practice. A Systematic Review applies your question, criteria, and extraction columns to thousands of papers, so small gaps in the setup show up across the whole review. A few minutes with the agent beforehand saves rework (and usage!) later.

The Systematic Review itself will suggest screening criteria and extraction columns based on your research question, so a vague question gets you vague suggestions. Refining your question first gets you better suggestions. You can also use the agent to help you refine and write instructions for your own custom screening criteria or data extraction columns to use in your review.

Use the Research Agent to:

"I'm planning a systematic review on the efficacy of JAK inhibitors vs biologics in moderate-to-severe ulcerative colitis. Help me refine the research question into PICO format, tell me roughly how much literature exists and whether I should narrow the scope, then draft inclusion and exclusion criteria and a list of extraction columns with instructions for each."

For a large or high-stakes question, ask the Research Agent to confirm its understanding before it searches. Correct the scope before a lot of time or usage is spent.

"Before you start, restate my question in your own words, list the sources you plan to search, and describe the output you'll produce. Wait for my confirmation."

The strongest prompts cover five things:

Before:

"What's the competitive landscape for KRAS inhibitors?"

After:

"Map the competitive landscape for KRAS G12C and G12D inhibitors in solid tumors, covering approved drugs and programs in Phase 1 or later since 2021. For each program, give the sponsor, phase, indication, and the most recent efficacy readout with its source. Separate approved, clinical, and discontinued programs. Present it as a table, then a one-paragraph summary of where the field is heading for a portfolio review audience."

Before:

"Is CLDN18.2 a good ADC target?"

After:

"Assess CLDN18.2 as an ADC target in gastric cancer. Cover expression in tumor vs normal tissue, internalization, and clinical safety signals from existing CLDN18.2 programs. Treat target-related, payload-related, and linker-related toxicity separately. Weight human clinical data above in-vivo data, and in-vivo above in-vitro. For internalization, answer yes, no, or unclear, with the supporting evidence."

Elicit doesn't know which distinctions your team cares about unless you tell it. If two things should be kept apart, say so. If some evidence counts more than other evidence, say how much.

Naming a reader works better than asking for a "simple" or "detailed" explanation, which tends to come back too basic or too technical.

"Explain the mechanism of durable remission with PD-1 inhibitors for a medical affairs lead who knows oncology but not immunology. One paragraph, no assay-level detail."

Be as specific about the output as about the question. Name the number of slides, the table columns, the sections, or the length.

"Summarize the scientific consensus on this mechanism in 3 slides: one on what's established, one on what's contested, and one on open questions. Include citations on each slide."

Words like "important," "best," "key," or "relevant" mean different things to different people. Tell Elicit what you mean by them.

Instead of:

"Find the 50 most important papers on GLP-1 and cardiovascular outcomes."

Try:

"Find the 50 most important papers on GLP-1 receptor agonists and cardiovascular outcomes. Rank by study design (RCTs and meta-analyses first), then by citation count, then by recency. Note which papers you assessed from full text and which from the abstract only."

Follow up in the same session when you're refining, expanding, or reformatting what you already have. The agent keeps the context of the conversation, so you don't need to repeat yourself.

"Add a column for route of administration and re-sort by phase."

Start a new session when the question changes substantially, or when you want to work from a different set of sources or uploaded files.

In Systematic Reviews, each extraction column and screening criterion works like its own prompt. The same rules apply: say exactly what to extract, in what form, and what to do when the data is missing. Draft them with the Research Agent first, using the planning steps under Planning a Systematic Review? Start here first.

Instead of:

"Sample size"

Try:

"Total number of randomized participants across all arms. Report as a number. If only per-arm numbers are given, add them up and note that you did. If not reported, write 'Not reported.'"

Instead of:

"Adult patients"

Try:

"Include only studies where all participants are 18 or older. Exclude studies of mixed adult and pediatric populations unless adult results are reported separately."

When a prompt gives you the result you want, you don't need to rewrite it each time.

PICO

"For [indication], identify the populations, interventions, comparators, and outcomes studied in Phase 2 and 3 trials since [year]. Present a PICO table with one row per trial and a link to each source."

Targeted literature review

"Run a targeted literature review on [topic] covering [timeframe]. Include [study types]. Summarize the findings by [theme 1], [theme 2], and [theme 3], and list the key evidence gaps."

Target assessment

"Assess [target] in [indication]. Cover genetic evidence, expression, mechanism, safety signals, and competitive activity. Weight human data above preclinical. End with an overall assessment and the three biggest risks."

Competitive landscape

"Map all programs targeting [target or mechanism] in [indication] at Phase 1 or later. Give sponsor, modality, phase, latest readout, and expected next milestone in a table."

Running the same prompt twice gives you similar results, but not always the exact same set of papers. Sources are updated over time, and the agent may take a slightly different path. If you need the same structure every time, save the prompt as a Skill.