# Trial Protocol Desk

> Paste a clinical study synopsis: the sample size and the power a planned enrolment buys are worked out free in the
> browser, the synopsis is checked for what a protocol needs, and the Schedule of Activities is drawn. A metered run
> critiques the design or drafts the core protocol sections, held to the page's numbers. A drafting aid for
> investigators, sponsors and CROs - not medical, statistical or regulatory advice. Built on the SkillSafe skill
> @anthropics/clinical-trial-protocol-skill (anthropics/healthcare).

https://trial-protocol-desk.skillsafe.ai/

## The input format
"Key: value" lines - Title, Intervention, Product type, Phase, Indication, Design, Control, Blinding, Primary endpoint,
Secondary endpoints, Endpoint type, Difference, SD, Control rate, Treatment rate, Margin, Alpha, Power, Dropout, Planned
enrolment, Inclusion, Exclusion, Visits, Assessments, Safety, DSMB, Analysis - with lists as bullets under their key.
Assessments read "Vital signs: Screening, Baseline, Week 12" (or "all").

## What is free
- Sample size: continuous endpoints by a two-sample t-test (normal approximation), binary endpoints by a two-proportion z-test; two-sided alpha for superiority, one-sided for non-inferiority (the margin replaces the effect); any allocation ratio (standard arm-size formula); enrolment inflated for dropout; sensitivity at 90% power and at a 10% smaller effect (or narrower margin). Matches the formulas of the skill's sample_size_calculator.py at 1:1 allocation.
- The power a planned enrolment buys after dropout.
- Checks: no primary endpoint, no timepoint, the primary timepoint not a scheduled visit, several primary endpoints without a multiplicity plan, no comparator, no randomisation, no blinding stated, no sample-size assumptions, a non-inferiority design without a margin, no dropout assumption, an underpowered planned enrolment, no analysis method or population, no inclusion or exclusion criteria, criteria without a measurable definition, no visit schedule, no safety monitoring, no data monitoring committee for a pivotal or high-risk study, promotional claims about the investigational product.
- The Schedule of Activities as a grid; CSV and Markdown exports.

## The metered lanes
Two tasks on gpt-terra, metered per run, routed by `task`. Input string fields: `task` (critique or draft), `synopsis`, `focus` (optional) and `facts` (the page's reading).
- `critique` replies with `verdict` (ready, revise, rework), `summary`, `findings`, `prescan`, `design_review` (Endpoint, Control and blinding, Population, Sample size, Analysis, Safety oversight, Operations, each with a risk), `sample_size`, `questions` (for biostatistics, clinical, regulatory or operations) and `caveats`.
- `draft` replies with `summary`, `sections` (synopsis, objectives, design, population, intervention, schedule, safety, statistics), `sample_size` (enrol and statement), `findings`, `prescan`, `open_questions` and `caveats`.
- The page requires every TP- error and warning answered by rule id, every sample-size number to be one of its own, no NCT numbers, guidance titles or publications that the synopsis does not contain, no promotional claims, and - in a draft - all eight sections, the page's enrolment in the statistics section and the primary timepoint wherever the endpoint is named. A critique can be handed to the draft as its focus.

## Links
- API tutorial: https://trial-protocol-desk.skillsafe.ai/api.html
- Source skill: https://skillsafe.ai/skill/@anthropics/clinical-trial-protocol-skill