I’ve seen a lot of vague posts about using Elicit for literature reviews, but few that dive into the concrete, messy details of configuring it for a real, high-stakes workflow. My team is responsible for screening clinical trial protocols against inclusion/exclusion criteria, and the default column setup in Elicit was slowing us down. It presented too much fluff and not enough actionable data at a glance. So I built a custom column set that cuts directly to what matters for regulatory and clinical feasibility reviews.
The goal was to transform the list of papers from an academic bibliography into a direct screening dashboard. We need to judge relevance in under 30 seconds per paper. The default "Abstract" column is a wall of text. The default "Interventions" or "Outcomes" extractions are too generic. Here’s the column set I implemented, which we now use as a team template:
* **Trial Phase:** (Custom property) Extracted via a prompt from the abstract. Shows "Phase I", "Phase II/III", etc., immediately.
* **Participant Count:** (Custom property) The actual N number from the study. Crucial for feasibility assessment.
* **Core Design:** Replaces the generic "Methodology" column. Prompted to output: "RCT", "Cohort", "Case-Control", "Systematic Review".
* **Key Eligibility Snippet:** (Custom property) This is the most important. I prompt Elicit to extract the *single most relevant sentence* from the methods section detailing the primary inclusion/exclusion criteria. Not the whole criteria, just the decisive one.
* **Primary Endpoint:** (Custom property) Extracted verbatim where possible.
* **Conflict Flag:** (Custom property) A simple "YES/NO" output from a prompt asking if the study's primary funder is a pharmaceutical company directly related to the intervention. Critical for our compliance layer.
* **DOI Link:** The default, but made into a clickable link column for quick access.
Setting this up required careful, iterative prompt engineering for each custom column. You cannot be vague. For the "Key Eligibility Snippet," my initial prompt was terrible: "Extract eligibility criteria." It returned huge paragraphs. The working prompt is:
```
From the 'Methods' section, extract the one sentence that describes the most restrictive primary inclusion or exclusion criterion for patient participants. Return only that sentence. If none is clearly the most restrictive, return the sentence containing 'inclusion criteria' or 'exclusion criteria'.
```
The implementation revealed Elicit's strengths and a significant pitfall.
* Strength: Once configured, the workflow speed increased dramatically. The table view lets us sort by `Participant Count` to prioritize larger studies and filter `Conflict Flag: YES` for secondary review.
* Pitfall: Extraction consistency is not 100%. For approximately 15% of papers, the `Core Design` or `Trial Phase` extraction fails or is mislabeled, requiring a manual cross-check. This is not a set-and-forget system; it's a force multiplier for a trained reviewer.
This column set is now our first-pass screening rig. It moves the cognitive load from parsing text to making decisions. I'm sharing this not as a perfect solution, but as a concrete example of how to bend Elicit into a specialized tool. If you're using it for anything beyond casual academic curiosity, you need to invest time in tailoring the output. The defaults are for beginners.
The template string for this column set is below. You can import it into your own Elicit workspace if your workflow is similar. Be warned: you will likely need to tweak the prompts for your specific therapeutic area or review type.
```
[Your Elicit Column Set JSON configuration would be pasted here. It would include the column keys, types, and the exact prompts used for each custom property.]
```
-- as
This is exactly the kind of practical detail that makes these discussions valuable. The pivot from a generic "Methodology" column to a "Core Design" one is smart - it forces the extraction to focus on what's operationally critical, like whether it's a crossover or parallel-group study.
I'm curious about your extraction for "Participant Count." Do you find the model reliably pulls the actual enrolled number versus the target? That distinction can be a major time-sink in screening.
—HR