Claude Code plugins for academic and clinical work. Each one also lives in its own repository and installs independently.
/plugin marketplace add neelshah4/claude-plugins
Then install what you want:
/plugin install citation-verification@neel-plugins
/plugin install icu-clinical-consult@neel-plugins
Every entry is pinned to a release tag, so an install is reproducible.
| Plugin | Version | Does |
|---|---|---|
| icu-clinical-consult | 1.2.0 | Four independent critical-care specialists reconciled by a synthesizer |
| citation-verification | 2026.8.29 | Checks every PMID, DOI, PMC, arXiv, and NCT identifier against live sources |
| fabrication-audit | 2026.8.29 | Verifies doses, thresholds, percentages, and other retrievable values |
| prompt-optimizer | 2026.8.22 | Structures a request before Claude produces content, with enforcement hooks |
Two versioning schemes, deliberately. icu-clinical-consult carries a semantic version because the underlying skill does. The other three are calendar-versioned, because those skills are dated rather than numbered. Each plugin's version matches its skill's own declared version — nothing is renumbered for cosmetic tidiness.
These share one assumption: a language model's confident output looks identical whether it retrieved a fact or invented one. There is no tell in the prose. So verification has to be structural rather than a matter of asking the model to be careful.
That assumption produces two families.
The verification layer. fabrication-audit and citation-verification split the claim surface by type — academic identifiers in one, every other retrievable value in the other. They compose deliberately, and neither covers the other's territory. Running one alone leaves half the surface unchecked.
The adversarial panel. icu-clinical-consult solves a different problem: a single reasoner finds a frame early and defends it, so asking for alternatives returns variations rather than genuine alternatives. It dispatches four specialists who cannot see each other's findings, then runs a cross-check whose actual output is the disagreements. Where two reviewers contradict each other is where the real question is.
prompt-optimizer works the other end — specifying the task before the model guesses at it.
Writing anything with citations → citation-verification and fabrication-audit. Together.
Critical-care questions → icu-clinical-consult. Read its scope statement first: decision support for licensed clinicians, not an autonomous system, and not a medical device.
Everything else → prompt-optimizer is the one with no domain assumptions.
| Needed for | |
|---|---|
| Claude Code with the Agent tool | The multi-agent panel. Without it it degrades to a single pass and says so. |
| Web access | Every verification path, and guideline retrieval. |
| PubMed MCP (optional) | Improves retrieval in citation-verification and icu-clinical-consult. |
All four work on any current Claude model. They were developed and tuned on Opus 5; nothing pins a model ID.
icu-clinical-consult and citation-verification each ship a copy of fabrication-audit so they work standalone on install.
If you install both, also install fabrication-audit itself — that repo is canonical, and if the copies drift, it wins.
For maintainers, scripts/sync-fabrication-audit.sh checks and repairs drift across the bundles. Clone the repos into one directory, then:
./scripts/sync-fabrication-audit.sh check # report drift, exit 1 if any
./scripts/sync-fabrication-audit.sh apply # copy canonical over the bundlesIt fails loudly rather than silently skipping a repo it cannot find, which is the behavior you want from something that runs before a release.
Each plugin versions independently and tracks its own skill's declared version. Marketplace entries pin ref to a release tag rather than main, so:
- an install today and an install next month give the same code
- a version bump requires editing
marketplace.jsonin the same change - you can pin an older version by installing from the repo at that tag
Bug reports and pull requests belong on the individual plugin repository, not here. This repo holds only the listing.
MIT throughout. Author: Neel Shah, MD, MSc.