Clinical Trial Design (ToolUniverse Claude Skill)

A ToolUniverse agent skill that stress-tests a proposed trial before a statistician is engaged — sizing the eligible population, checking whether the FDA has previously accepted the endpoint, sourcing a comparator, and scoring overall feasibility from precedent.


Type	Claude Skill (one of ToolUniverse’s pre-built agent skills)
Supplier	Zitnik Lab, Harvard Medical School
Availability	GA — part of the ToolUniverse skills collection (`skills/tooluniverse-clinical-trial-design/`)
Pricing	Free / OSS (Apache-2.0); wraps public sources (Open Targets, ClinicalTrials.gov, ClinVar, gnomAD, COSMIC, DrugBank, FDA Orange Book, openFDA, FAERS, PubMed)
Capabilities	Read-only — evidence gathering and scoring; runs no statistical computation
Verified	works · 2026-08-13
Security	cleared · 2026-08-13 — mims-harvard/ToolUniverse Apache-2.0, read-only public sources

How to install

This skill calls ToolUniverse tools, so the ToolUniverse MCP server must be installed first (see the ToolUniverse page). Simplest registration:

claude mcp add --transport stdio tooluniverse -- uvx tooluniverse

Then add the skills:

Claude Code — install the whole skill collection (the skill resolves as tooluniverse-clinical-trial-design):
```
npx skills add mims-harvard/ToolUniverse
```
Manual / other agents — copy just this skill directory into your skills folder:
```
git clone https://github.com/mims-harvard/ToolUniverse
cp -r ToolUniverse/skills/tooluniverse-clinical-trial-design ~/.claude/skills/
```
(replace ~/.claude/skills/ with your agent’s skills directory if you are not using Claude Code/Desktop.)

The skill sets disable-model-invocation: true upstream, so invoke it explicitly (e.g. ask Claude to “use the clinical-trial-design skill”) rather than relying on automatic dispatch.

What it does

Assesses six design dimensions — endpoint (OS, PFS, ORR, or biomarker-based), population (unselected vs biomarker-enriched), comparator (placebo, active control, or historical), effect size, duration, and regulatory pathway — via six parallel research paths:

Patient-population sizing from disease prevalence and biomarker frequency, through an eligibility funnel to an enrollment projection.
Biomarker testing logistics — assay availability, cost, reimbursement.
Standard-of-care identification and comparator sourcing.
Endpoint precedent — has the FDA accepted this endpoint in this indication before?
Safety monitoring — dose-limiting toxicity definitions and organ-specific surveillance, informed by FAERS signals for the class.
Regulatory precedent — breakthrough designations, prior approval pathways, Orange Book approval history.

Scoring rules it applies:

Feasibility score ≥ 75 = HIGH (proceed); 50–74 = MODERATE (additional validation needed); < 50 = LOW (de-risk before committing).
Component weights: patient availability 30%, endpoint precedent 25%, regulatory clarity 20%, comparator feasibility 15%, safety monitoring 10%.
Effect-size sizing rules of thumb: a 20% improvement in objective response rate needs roughly 100 patients per arm; a 50% improvement roughly 30 per arm.
Endpoints are graded A (regulatory acceptance, multiple precedents) through D (novel, unvalidated).

Output is a 14-section markdown report covering executive summary, disease background, population, biomarker strategy, endpoint, comparator, safety monitoring, study design, enrollment strategy, regulatory pathway, budget, risk, and success criteria.

Primary use cases: go/no-go feasibility on a proposed indication, endpoint selection defensible by precedent, enrollment projection and site-count sanity check, comparator and pathway strategy.

Notes

It is a reasoning layer over ToolUniverse; without the MCP server registered, the tool calls fail.

Upstream states plainly that this is precedent-based reasoning, not first-principles math — the per-arm numbers above are order-of-magnitude anchors drawn from what comparable trials did, not a power calculation. Do not put them in a protocol. Run the actual sample-size calculation with Power and Sample Size, and take group-sequential or adaptive features to Adaptive Designs. The 30/25/20/15/10 weighting is likewise a convention the skill imposes, not a validated instrument; a feasibility score is a way to compare two candidate designs, not an absolute probability of success.

Prevalence figures inherited from Open Targets and biomarker frequencies from gnomAD/COSMIC/ClinVar carry their own ascertainment biases, which propagate straight into the enrollment funnel — the eligible-population number is usually the most optimistic figure in the report.

Complements Clinical Trial Protocol (the Anthropic plugin that drafts the protocol document once the design is settled), Clinical Trial Matching (the patient-side counterpart), Drug Regulatory Research, and Trial Reporting. ToolUniverse ships ~68 such skills; other workflows are catalogued separately.

Sources

Installed this tool?

Share feedback — install path, OS, errors, workarounds. The form opens with this tool pre-selected and a link back to this page.