PMx Skills Index

Public pharmacometrics skill files, ranked each month by Claude

A monthly, Claude-maintained ranking of public skill files for pharmacometrics work, in tool and task categories, scored on quality, author expertise, maintenance and evidence that people use them, by criteria Andy sets.

This list is compiled and ranked each month by Claude, an AI model made by Anthropic, running unattended. No person reviews the entries or the scores before they are published. Andy Stein sets the criteria in Specification 2 and does not adjust individual scores.

Every score runs from 0 to 100. The full definitions are in Specification 2, Section 4. The scores capture:

• S The total of the other scores.
• Q Quality
• E Expertise
• M Maintenance and usability
• U Usage

Grouping:

• Only skills with evidence of use are ranked.
• Flags. ❌ a major error found. ⚠️ no licence, archived repository, authors not found in OpenAlex, or an author matched by name alone.
• At most three skills per library are ranked in each category.

Tool Skills

Skills for particular software tools (e.g. NONMEM, Monolix, nlmixr2)

Task Skills

Skills whose instructions hold across tools (e.g. Model Evaluation)

Unproven

Skills scored on everything except use, in order of their total, with no evidence yet that anyone outside their authors has used them. A skill moves up to the ranked tables in the month that evidence appears.

Evaluation Resources

Test cases and benchmarks for measuring how well an assistant does pharmacometrics work. Not ranked: Specification 2 does not yet define a quality score for them.

This Month’s Run

October 2026, run on 2026-10-07, and reissued the same day under criteria version 2, which replaced adoption with use and ranks only skills with evidence of use. Five of the 40 skills have that evidence. Discovery searched GitHub code and repositories with the queries in discovery.yml, cloned 25 repositories and scored 40 skills. Copies were merged into their originals: 37 copies of K-Dense’s pkpd-modeling alone. Exclusions and copies are listed in list/2026-10.yml.

Four skills were added on 2026-10-09, outside the run, at Andy’s request. Discovery missed nlmixr2/AI-assisted-MIDD, an NCA, popPK and simulation project template created six days before the run. Its three analysis skills and its data-validation skill were scored by the same criteria, in the same single pass. The other 14 skills in the repository are the database, logging, report and QC steps around those four, and were not scored. The four have no evidence of use yet, so they are listed under Unproven.

Both known-answer checks passed. The error audit reported both errors in fixtures.yml as major before anything was scored.

This first run departs from the specification in four ways:

1. One scoring pass. The rubric and the error audit were each run once, by the same model. The second independent pass, the judge-agreement figure and the test-retest correlation were not produced, so a difference of a few points between two skills is within the noise of a single judgement.
2. Large libraries were pre-screened. In libraries of 30 or more skills, the most pharmacometric skills were chosen by keyword before scoring, and the per-library cap applied to those.
3. No benchmark score. No evaluation resource yet covers a listed skill’s task with expected findings.
4. Use is measured from GitHub issues, pull requests, comments and installations only. GitHub Discussions were not searched, and no listed skill ships in a package with download statistics.

Proposed Changes to the Criteria

The run proposes these for Andy to accept or reject in criteria.yml and rubric.md. Until he does, the criteria stay as they are.

1. Add a rubric item for domain content. A templated review skill with no named diagnostic meets most of the current items, so it scores close to a skill that explains how to read a VPC. An item such as “names at least five specific diagnostics or methods and says how to read each” would separate them.
2. Add a rubric item for verified output. Some tool skills require the assistant to check its result numerically before reporting it, such as a converted model reproducing the original engine’s predictions. The rubric gives that no credit.
3. Define a score for evaluation resources, from the number of test cases with expected answers and the tasks they cover, so the third category can be ranked.
4. Decide how organisations count for expertise. Skills published under an organisation or a bot account score 0 on E, however expert the team. A named-author list, where the organisation publishes one, could stand in.