Turn a Drug Development Evidence Gap Into a Structured Research Brief Using Only Prompts

The task

Clinical researchers, medical affairs staff, and evidence-generation leads regularly get asked, "What do we actually know about drug X in population Y — and where are the gaps?" Answering means pulling unstructured study summaries, endpoints, subgroups, and safety signals into one reviewable brief that a medical director can sign off on. This workflow does the extraction and cross-linking with prompts, so you spend your time judging the evidence instead of formatting it.

Before AI

Today, this is a full afternoon of copy-paste. You pull study abstracts and internal summaries into a Word doc, hand-tag endpoints and populations, build a comparison table, and try to spot where two trials disagree on subgroup effects. The relationships between studies — same endpoint, different dose; overlapping population, opposite safety signal — live in your head or a scratch spreadsheet, and get lost when you hand the brief off.

The failure mode is well-described in a recent MedCity News piece on evidence in drug development: drug development isn't limited by the amount of data we collect, but by our ability to preserve the context, meaning and relationships that transform data into evidence. Context includes everything surrounding a clinical observation, such as disease stage, prior treatments, patient characteristics, care setting, timing and concurrent events. Without this information, AI sees numbers. This workflow forces the model to keep that context attached to every extracted fact.

The workflow

Important: Do not paste identified patient data. Use published trial summaries, conference abstracts, or de-identified internal synopses. This workflow assumes HIPAA-covered records stay in your EHR — you are working with study-level evidence, not patient records.

Step 1 — Extract facts with their context intact

Paste your unstructured study summaries into the prompt below. The prompt forces the model to keep each finding tethered to its population, dose, endpoint definition, and timing — the context that usually gets stripped when someone builds a summary table by hand.

You are an evidence-extraction assistant for a clinical research team preparing an internal drug development brief. You will receive one or more unstructured study summaries below.

For EACH study, extract a JSON object with these fields. If a field is not stated, write "not reported" — do not infer.

- study_id: short label (e.g., first author + year, or trial acronym)
- design: trial phase and design (RCT, single-arm, retrospective, etc.)
- population: disease, stage, line of therapy, key inclusion criteria, N
- intervention: drug, dose, schedule, duration
- comparator: control arm or "none"
- primary_endpoint: exact definition as reported (e.g., "PFS by BICR per RECIST 1.1")
- primary_result: point estimate + CI + p-value if reported
- key_subgroups: list any subgroup effects reported, each with population + effect
- safety_signals: notable AEs with grade and rate
- context_flags: anything that limits generalizability (single-center, prior-treatment mix, small N, open-label, industry-funded, etc.)
- source_span: quote the exact phrase from the input that supports primary_result

Return a JSON array. Do not summarize across studies yet. Do not drop fields to save space. Preserve the exact endpoint wording — do not paraphrase "PFS" as "progression" or vice versa.

Studies to extract:

STUDY A — ARIA-2 (Chen et al., 2025)
Phase 2 randomized open-label trial of veltronib 80mg QD vs. investigator's choice chemo in 142 patients with previously-treated HER2-low metastatic breast cancer (≥2 prior lines, ECOG 0-1). Primary endpoint: PFS by investigator assessment per RECIST 1.1. Median PFS 7.4 mo (95% CI 5.9-9.1) vs 4.2 mo (95% CI 3.1-5.0), HR 0.58 (0.40-0.84), p=0.004. Prespecified subgroup: patients with prior CDK4/6 inhibitor (n=88) showed HR 0.51; those without (n=54) HR 0.79. Grade ≥3 AEs: neutropenia 22%, diarrhea 9%, ILD 3% (one grade 5 event). Single-region trial, 11 sites in East Asia. Industry sponsored.

STUDY B — VELOCITY (internal synopsis, data cut March 2026)
Single-arm phase 2, veltronib 60mg QD in 61 patients with HER2-low mBC, 1-3 prior lines, including CNS mets allowed if stable. Primary endpoint: confirmed ORR per BICR RECIST 1.1. ORR 34% (95% CI 22-47). Median PFS (secondary) 6.1 mo. In the CNS-met subgroup (n=18), intracranial response rate 28%. Grade ≥3 neutropenia 15%, ILD not reported at this cut. US + EU sites, open-label. Sponsor: internal.

STUDY C — Registry analysis (Okafor et al., 2024)
Retrospective real-world analysis of 312 HER2-low mBC patients receiving off-label veltronib across 6 US academic centers, 2022-2024. Median 3 prior lines. Median rwPFS 5.2 mo (IQR 2.8-8.9). No formal comparator. ILD reported in 5% (16/312), including two fatal cases; authors note under-ascertainment likely. Heavy pretreatment and payer-driven dosing variability noted as limitations.

Step 2 — Build the relationship map

This is where most evidence briefs fall apart: two studies use "PFS" but mean different things, or a subgroup finding in Study A directly informs the eligibility question in Study B — and nobody notices. This prompt makes the model surface those links explicitly.

You now have the extracted JSON array from the previous step. Do NOT re-extract. Work only from that structured data.

Produce a "relationship map" as a markdown table with these columns:
| Relationship type | Study pair | Shared element | Where they agree | Where they diverge | Implication for the brief |

Relationship types to look for:
1. Same endpoint, different definition (flag any endpoint labeled the same but assessed differently — e.g., investigator vs. BICR, or PFS vs. rwPFS)
2. Overlapping population, different dose or schedule
3. Subgroup in one study = full population in another
4. Safety signal reported in one, absent or under-ascertained in another
5. Context-flag conflicts (e.g., single-region vs. multi-region, RCT vs. retrospective)

After the table, list any endpoints or safety signals that appear in only ONE study and therefore cannot be triangulated. Be explicit — this is the evidence gap section.

Do not invent relationships. If two studies share nothing meaningful, say so.

Step 3 — Draft the reviewable brief

Now assemble the brief a medical director will actually read. The prompt keeps the context flags attached to every claim so reviewers can see the caveat next to the number, not buried in a footnote.

Using the extracted study data AND the relationship map from the previous two steps, draft an internal evidence brief in this exact structure:

# Evidence Brief: [infer drug + indication from the data]

## 1. Bottom line (3 bullets max)
State what the evidence supports, at what confidence, for whom. Every bullet must name the population it applies to.

## 2. Evidence summary table
One row per study. Columns: Study | Design | N | Population | Intervention | Primary result | Key context flags.

## 3. What the studies agree on
Only claims supported by ≥2 studies. Cite study_ids inline.

## 4. Where studies diverge
Pull directly from the relationship map. For each divergence, state which study's design gives it more weight and why (RCT > single-arm > retrospective, prespecified > post-hoc, BICR > investigator-assessed).

## 5. Evidence gaps — what we do NOT know
Use the "appears in only one study" list from step 2. Frame each as a question the next study would need to answer.

## 6. Recommended next actions
2-4 concrete items: additional analyses, studies to request, subgroups to explore, safety follow-up.

Rules:
- Never state a result without its population and context flags in the same sentence.
- Do not upgrade retrospective findings to causal language.
- If the CI crosses 1 (or the null), say the result is not statistically conclusive.
- Do not add any study, endpoint, or number that was not in the extracted data.

Gotchas

• The model will paraphrase endpoints if you let it. "PFS by BICR per RECIST 1.1" is not the same as "progression-free survival." Step 1's source_span field is there so a reviewer can spot-check that the model didn't collapse a defined endpoint into a generic term.
• Real-world data invites causal drift. A registry HR looks like an RCT HR on the page. Step 3's rule against upgrading retrospective findings helps, but a human still needs to read the "Where studies diverge" section carefully.
• Safety under-ascertainment is invisible to the model unless the source text flags it. If your input summary doesn't mention that ILD was not systematically collected, the brief will treat "not reported" as "not present." Add a manual safety-review pass.
• This is study-level work, not patient-level. Do not paste EHR extracts, MRNs, or anything traceable to a person into a general-purpose model. Keep patient records inside your covered EHR environment; use this workflow for published or de-identified synopses only.
• Small-N subgroups will get overweighted. The n=18 CNS subgroup in the sample input is hypothesis-generating at best. If the brief presents it as a finding, edit it down to a gap.

Time saved

A typical three-to-five study evidence brief takes a medical writer or researcher 3-4 hours end-to-end: reading, tabling, cross-linking, drafting. This workflow gets you to a reviewable draft in about 20-30 minutes of prompt runs plus review — call it 2-3 hours saved per brief. The saving comes from the cross-linking step, which is the part humans routinely skip when they're rushed. As the MedCity piece argues, that linking work is the actual evidence problem, not the data volume.

What we got when we ran it

We executed every prompt above, in order, against the sample input (model: claude-sonnet-4-6). Outputs below are real and unedited apart from length truncation.

Step 1 output

```json [ { "study_id": "ARIA-2 (Chen et al., 2025)", "design": "Phase 2, randomized, open-label trial", "population": { "disease": "HER2-low metastatic breast cancer", "stage": "Metastatic", "line_of_therapy": "Previously treated, ≥2 prior lines", "key_inclusion_criteria": "ECOG 0-1", "N": 142 }, "intervention": { "drug": "Veltronib", "dose": "80 mg", "schedule": "QD (once daily)", "duration": "not reported" }, "comparator": "Investigator's choice chemotherapy", "primary_endpoint": "PFS by investigator assessment per RECIST 1.1", "primary_result": { "veltronib_arm": "Median PFS 7.4 mo (95% CI 5.9–9.1)", "comparator_arm": "Median PFS 4.2 mo (95% CI 3.1–5.0)", "HR": "0.58 (95% CI 0.40–0.84)", "p_value": "0.004" }, "key_subgroups": [ { "population": "Patients with prior CDK4/6 inhibitor (n=88)", "effect": "HR 0.51" }, { "population": "Patients without prior CDK4/6 inhibitor (n=54)", "effect": "HR 0.79" } ], "safety_signals": [ { "AE": "Neutropenia", "grade": "≥3", "rate": "22%" }, { "AE": "Diarrhea", "grade": "≥3", "rate": "9%" }, { "AE": "ILD (interstitial lung disease)", "grade": "≥3", "rate": "3%", "note": "One grade 5 event" } ], "context_flags": [ "Single-region trial (East Asia only), 11 sites", "Open-label design", "Industry sponsored", "Investigator assessment used for primary endpoint (not blinded central review)" ], "source_span": "Median PFS 7.4 mo (95% CI 5.9-9.1) vs 4.2 mo (95% CI 3.1-5.0), HR 0.58 (0.40-0.84), p=0.004" }, { "study_id": "VELOCITY (internal synopsis, data cut March 2026)", "design": "Single-arm, phase 2", "population": { "disease": "HER2-low metastatic breast cancer", "stage": "Metastatic", "line_of_therapy": "1–3 prior lines", "key_inclusion_criteria": "CNS metastases allowed if stable", "N": 61 }, "intervention": { "drug": "Veltronib", "dose": "60 mg", "schedule": "QD (once daily)", "duration": "not reported" }, "comparator": "None", "primary_endpoint": "Confirmed ORR per

…output truncated at 250 words.

Step 2 output

Relationship Map | Relationship Type | Study Pair | Shared Element | Where They Agree | Where They Diverge | Implication for the Brief | |---|---|---|---|---|---| | **1. Same endpoint, different definition** | ARIA-2 vs. VELOCITY | Both report PFS | Both report median PFS as a time-to-event efficacy measure in HER2-low mBC | ARIA-2 uses investigator assessment per RECIST 1.1 (primary); VELOCITY reports median PFS of 6.1 mo as a *secondary* endpoint, assessor not specified | PFS estimates are not directly comparable across studies; VELOCITY's 6.1 mo cannot be benchmarked against ARIA-2's 7.4 mo without knowing assessment method. Brief should not present these as concordant PFS data points. | | **1. Same endpoint, different definition** | ARIA-2 vs. Okafor 2024 | Both report a PFS-type metric | Both suggest median time-to-progression in the 5–7 mo range in pretreated HER2-low mBC | ARIA-2 uses investigator-assessed PFS per RECIST 1.1 in an RCT; Okafor uses real-world PFS (rwPFS) with no BICR, no formal primary endpoint, and dosing variability | rwPFS and trial PFS are methodologically distinct. The apparent numerical similarity (7.4 mo vs. 5.2 mo) may reflect different patient selection and ascertainment, not a true efficacy signal. Flag explicitly in brief. | | **2. Overlapping population, different dose** | ARIA-2 vs. VELOCITY | Veltronib in HER2-low mBC, overlapping line-of-therapy range | Both enroll pretreated HER2-low mBC patients; both use veltronib monotherapy QD | ARIA-2 uses 80 mg QD; VELOCITY uses 60 mg QD. No head-to-head dose comparison exists. | Dose-response

…output truncated at 250 words.

Step 3 output

# Evidence Brief: Veltronib in HER2-Low Metastatic Breast Cancer (Pretreated) --- ## 1. Bottom Line - In previously treated (≥2 prior lines) HER2-low mBC patients with ECOG 0–1, enrolled across 11 East Asian sites, veltronib 80 mg QD demonstrated a statistically significant PFS benefit over investigator's choice chemotherapy (median PFS 7.4 mo vs. 4.2 mo, HR 0.58 [95% CI 0.40–0.84], p=0.004) in a phase 2 RCT (ARIA-2); this is the only controlled efficacy estimate in the package and its generalizability to non-Asian populations is uncertain. - In a broader pretreated HER2-low mBC population including patients with stable CNS metastases (1–3 prior lines, US + EU sites), veltronib 60 mg QD produced a confirmed ORR of 34% (95% CI 22–47%) by BICR in a single-arm phase 2 trial (VELOCITY); this is a single-arm result with no comparator, and ILD data at this dose are not yet mature. - Across all three data sources, ILD — including fatal cases — is a consistent safety signal that cannot be fully characterized at either dose level; it must be treated as a priority risk in any regulatory or prescribing communication, and no dose-specific ILD fatality rate can be stated from the current evidence package. --- ## 2. Evidence Summary Table | Study | Design | N | Population | Intervention | Primary Result | Key Context Flags | |---|---|---|---|---|---|---| | ARIA-2 (Chen et al., 2025) | Phase 2 RCT, open-label | 142 | HER2-low mBC; ≥2 prior lines; ECOG 0–1 | Veltronib 80 mg

…output truncated at 250 words.

---

This content is for informational purposes only and is not medical advice. AI tools used with patient data must meet your organization's HIPAA and privacy requirements.