ClinPharm PMx Skills

Portable Agent Skills for clinical pharmacologists and pharmacometricians who review documents and analyses.

Turn any AI agent into a clinical pharmacologist and pharmacometrician.

It reviews, reconciles, and prepares evidence. It never selects a dose, signs off, or submits.

[!IMPORTANT] Skills review, reconcile, verify, structure, and flag. Qualified humans decide, approve, sign, and submit. No skill selects a dose, resolves a scientific disagreement, edits a controlled document, or replaces clinical, pharmacometric, medical, regulatory, or quality judgment.

Getting started — pick your door

New here, and not a software person? Take the green door. Everything else on this page assumes a comfort with GitHub that you do not need in order to use this.

Wondering whether an AI wrote all this and nobody checked? That is the right question — VALIDATION.md answers it without hedging.

The 60-second route, in full

1. Open skills/verify-nca-outputs/PASTE.md.
2. Paste it into Claude, ChatGPT, or any ordinary chat.
3. Attach the four files under examples/verify-nca-outputs/inputs/.
4. Ask:

Verify the NCA report against the parameter dataset and the analysis plan. Report every finding with its locator. Do not re-derive anything.

Every skill has the same web route: skills/<id>/PASTE.md. Review every finding before using the output.

More worked examples: examples/.

Install the library (desktop agents)

git clone https://github.com/malekokour/clinpharm-pmx-skills.git
cd clinpharm-pmx-skills

Ask in a new chat:

Which ClinPharm PMx Skills skill applies to reviewing PK sections of a CSR?

A loaded library names review-csr-pk-consistency. If the host answers without naming a package, the skills did not load.

Prefer library-router when the right skill is unclear. Keep each skills/<id>/ directory intact. Clone the whole repository — a single-folder copy is a degraded route (no router, no shared/).

What ✅ and ⬜ mean, and the machine-readable attestation: catalog/adapter-evidence.json. Loaded is not exercised — Claude Code listed all 151 packages after install; one skill was then run end to end. See docs/COMPATIBILITY.md.

What's in the library

Name the gate. released does not mean clinical validation, and it does not mean the evaluation suite has qualified a package's behaviour — three blocker-severity findings against that suite are open and frozen. CLAIM-LEDGER.md and AGENTS.md say this in the same words, and a gate fails the build if any of the three drifts.

The catalog spans CSR and protocol review, CTD 2.7.2 and USPI Section 12, DDI and FIH evidence, NCA verification, and cross-document reconciliation. Each package follows the Agent Skills specification: SKILL.md is the activation contract; scripts/, references/, and assets/ load only when needed.

The job map

Writing the profession out as tasks was uncomfortable. You see how much of the job nobody has built anything for.

That third row is the point. Live tree (open a domain, then a sub-domain): malekokour.github.io/clinpharm-pmx-skills/map/. On GitHub: map/ · A · B · C. Source: catalog/job-model-167.tsv.

One subdomain, zoomed — a job, then a task, then a skill you can run:

How it was built

Map the job. Decide what actually needs a skill. Build those. Write what each one refuses. Publish the gaps with the wins.

The same recording plays on the site with controls.

Evidence

Three claims that are often blurred stay separate here:

1. Package validity — structure, frontmatter, links, vendored dependencies, and clean installation are checked.
2. Diagnostic evidence — synthetic fixtures, planted defects, false-positive traps, and deterministic canaries show what the current test harness examines.
3. Qualification — risk-tiered behavioral comparison, accepted expert keys, held-out inputs, activation testing, practitioner review, and independent closeout are required before a package becomes released.

As of 2026-08-13, the public quality surface contains 151 evaluation suites, 1,197 cases, 4,510 assertions, 147 declared inputs, 67 portable scripts, and 230 repository tests. Those are engineering denominators—not a clinical-performance claim. Every one of them is re-derived from its source of record in CLAIM-LEDGER.md, and the build fails if this sentence and the sources disagree. A historical CSR deterministic-path result is retained under evals/benchmark/ as bounded diagnostic evidence after its broader qualification claim was retracted.

Regenerate counts from CLAIM-LEDGER.md before quoting them.

Safety and data boundary

Use only public, synthetic, or explicitly permitted information. Never upload patient-level data, credentials, sponsor-confidential material, unpublished submissions, employer-proprietary content, or personal contact details to an unapproved service.

Fixtures in this repository are synthetic by construction, not anonymized from real studies. This library is not medical advice, clinical decision support, a validated GxP system, or a patient-specific decision tool.

Contribute

Requires Python 3.11 or later and the pinned dependencies:

python3 -m venv .venv && source .venv/bin/activate
python3 -m pip install --requirement requirements.lock

The requirements.lock install is required, not optional. Without it, check_all.py stops at ModuleNotFoundError: No module named 'strictyaml' and exits 1. Then:

make check          # lint, validate, test, privacy, portability, and consistency
make docs-check     # prove generated documentation matches its authorities
make release-check  # build release assets in isolation and verify them

python3 scripts/check_all.py is the same gate as make check. It validates the repository; it does not send documents anywhere.

Before contributing, read CONTRIBUTING.md and the public contributor contract. New workflows need a bounded job, an explicit human-review boundary, synthetic failure cases, objective checks, and a truthful evidence gap.

map/             167 job-model task pages. Source: catalog/job-model-167.tsv
skills/          151 independently installable Agent Skill packages
contexts/        modality and therapeutic-area context files
collections/     editable status authority
catalog/         derived registry (do not edit by hand)
evals/           synthetic suites; historical records under evals/benchmark/
shared/          tools, modules, assets, contracts
examples/        worked synthetic examples
docs/ site/      documentation and GitHub Pages source
.claude-plugin/  Claude Code marketplace manifest
plugin.json      Agent Plugins 1.0.0 manifest

Collections own package status. catalog/catalog.json and docs/CATALOG.md are generated views; the quality gate fails if they drift.

Docs

Use — Hosts · Lifecycle · Compatibility · Boundaries · Privacy

Evidence — Claim ledger · Catalog · Job tree · Roadmap

Build — Authoring · Architecture · Governance · Glossary · References

Meta — Agent Skills · Agent Plugins · Security · Citation

License

MIT — see LICENSE.