Skip to content
Back to search

clinical-research-orchestra

MarketplaceWorks with Claude CodeGitHubIncomplete

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/drtranquanghung91-vn--clinical-research-orchestra/badge.svg)](/listings/drtranquanghung91-vn--clinical-research-orchestra)

HTML

<a href="/listings/drtranquanghung91-vn--clinical-research-orchestra"><img src="/api/listings/drtranquanghung91-vn--clinical-research-orchestra/badge.svg" alt="findsafeskills safety rating" /></a>

Install

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Claude Code

/plugin marketplace add drtranquanghung91-vn/clinical-research-orchestra
/plugin install phi-guard@clinical-research-orchestra
/plugin install orchestra@clinical-research-orchestra

Then start a project:

/orchestra:start

Tools the scripts use

Python 3.10 or later. The harness scripts use the standard library only. Literature lookups call the public PubMed, Europe PMC and Crossref APIs; no API key is needed. Analysis scripts written by `statistical-analysis` need `pandas`, `scipy` and `statsmodels` (`lifelines` for Cox models): `python -m pip install pandas scipy statsmodels openpyxl lifelines`.

Optional, for presenting (P3): `matplotlib` for figures, `python-pptx` for the defence deck and the poster PPTX (`python -m pip install matplotlib python-pptx`), and Google Chrome or Microsoft Edge (headless) for the poster PDF, PNG and print preflight. A script that needs a missing one prints the install hint and exits 2; the skills report that check as not run.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 20 lines
  1. {
  2. "name": "clinical-research-orchestra",
  3. "owner": { "name": "Trần Quang Hưng" },
  4. "metadata": {
  5. "description": "Evidence-first Claude Code plugins for clinical research: question, evidence, protocol, analysis, manuscript, thesis, poster and defence.",
  6. "version": "0.4.0"
  7. },
  8. "plugins": [
  9. {
  10. "name": "phi-guard",
  11. "source": "./plugins/phi-guard",
  12. "description": "Blocks patient identifiers from leaving the machine through MCP, WebFetch or WebSearch; de-identifies CSV/XLSX files. Install first."
  13. },
  14. {
  15. "name": "orchestra",
  16. "source": "./plugins/orchestra",
  17. "description": "Research lifecycle skills: question, literature, appraisal, protocol, sample size, analysis, manuscript, thesis, poster and defence, with verified references and checked numbers."
  18. }
  19. ]
  20. }
README.md · README · 232 lines
  1. <h1 align="center">Clinical Research Orchestra: A Multi-Agent Harness for<br>Evidence-First Clinical Research</h1>
  2. <p align="center">
  3. <b>From bedside question to podium, with every citation and number checked.</b><br>
  4. Claude Code plugins that take medical students, residents, doctors and other healthcare
  5. professionals through a research project: question, evidence, protocol, analysis,
  6. manuscript or thesis, poster and defence.
  7. </p>
  8. <p align="center">
  9. <a href="LICENSE"><img alt="MIT licence" src="https://img.shields.io/badge/licence-MIT-0F5E7A"></a>
  10. <img alt="Claude Code" src="https://img.shields.io/badge/Claude%20Code-plugin-C8372D">
  11. <img alt="Languages" src="https://img.shields.io/badge/languages-English%20%7C%20Vietnamese-56677F">
  12. <img alt="Status" src="https://img.shields.io/badge/status-0.4.0%20present-B7791F">
  13. <a href="https://github.com/drtranquanghung91-vn/clinical-research-orchestra/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/drtranquanghung91-vn/clinical-research-orchestra/actions/workflows/ci.yml/badge.svg"></a>
  14. </p>
  15. > **Status: early.** Version 0.4.0 adds the P3 "present" skills (figures, conference abstract,
  16. > print poster in three layouts with preflight, thesis-defence deck with rehearsal questions, prose
  17. > humanizer, consistency audit) to the P2 "analyse and report results" skills (data
  18. > preparation, statistical analysis per the SAP, results writing with a Chapter 3 .docx), the
  19. > P1 "plan the study" skills (research question, literature search, critical appraisal, guideline
  20. > check, literature review, protocol design, sample size and SAP) and the 0.1.0 foundation (project
  21. > folder, ledgers, gates, reference verification and patient-data protection). Manuscript and thesis
  22. > writing skills arrive phase by phase; see [Roadmap](#roadmap). A simulated project run through
  23. > every P1, P2 and P3 skill is in [`examples/fictional-cohort/`](examples/fictional-cohort/). Do not
  24. > rely on it for real studies yet.
  25. ---
  26. ## Why this exists
  27. Ask a general-purpose agent to "write the literature review for my thesis" or "calculate the
  28. sample size" and you get fluent text in minutes. It is often wrong in the ways that a
  29. supervisor, an ethics committee or a journal reviewer will catch:
  30. - **References are recalled, not retrieved.** A PMID or DOI "from memory" ends up in the
  31. reference list, or a real paper is cited for a claim it does not make.
  32. - **Numbers drift.** 85% in the abstract, 86.3% in the results; a sample size of "≈26"
  33. that was never rounded up; a confidence-interval bound used in place of the point estimate.
  34. - **The review is a list of papers.** Chapter 1 reads "Smith found… Nguyen found…" instead
  35. of a synthesis organised by question.
  36. - **Patient data leaks.** A name, a national ID number or a record number goes into a web
  37. search or an external tool.
  38. - **The prose sounds machine-written**, and nothing says how AI was used.
  39. Clinical Research Orchestra turns these steps into a method with files, checks and gates,
  40. in the spirit of [Superpowers](https://github.com/obra/superpowers) and
  41. [PaperOrchestra](https://github.com/google-research/paper-orchestra). The agent reasons;
  42. scripts measure. Anything a script can verify (an identifier, a number, a sample-size
  43. calculation, a poster's print resolution) is not left to the model's word.
  44. ## How it works
  45. Every project lives in one folder with stage directories (`01_question/` to `07_present/`),
  46. ledgers for references, results and claims, and four human gates:
  47. | Gate | Stops before | You decide |
  48. |---|---|---|
  49. | G1 Question | searching and protocol work | the research question, PICO/PECO and objectives |
  50. | G2 Protocol | data collection | design, eligibility, variables, sample size, analysis plan |
  51. | G3 Results lock | writing | the analysis outputs, locked with a SHA-256 hash |
  52. | G4 Release | sending anything out | the manuscript, thesis, poster or deck after all checks pass |
  53. Skills hand work forward through these files, not through the conversation, so you can
  54. enter at any stage (a thesis already half written, a poster due next week) and resume later
  55. with `/orchestra:status`. A gate can be waived only with a written reason in `decisions.md`.
  56. Two plugins ship from this marketplace:
  57. - **`orchestra`**: the research skills, commands, scripts and reference library.
  58. - **`phi-guard`**: a hook that scans outbound tool calls (web search, fetch, MCP tools,
  59. network commands) for patient identifiers and blocks or asks, plus a
  60. `deidentify-data` skill for CSV and Excel files. Install it on its own if you only want
  61. the protection.
  62. ## The skills
  63. | Stage | Skill | What it produces | Phase |
  64. |---|---|---|---|
  65. | 0 | `research-project-setup`, `/orchestra:start`, `/orchestra:status` | project folder, ledgers, gates, next step | 0.1.0 |
  66. | 0 | `deidentify-data` (phi-guard) | de-identified data file and a separate key file | 0.1.0 |
  67. | 1 | `research-question` | PICO/PECO, FINER check, objectives, hypotheses, title options | 0.2.0 |
  68. | 2 | `literature-search` | nomenclature check, search strategy, verified references, evidence table and map, gaps | 0.2.0 |
  69. | 2 | `critical-appraisal`, `/orchestra:appraise` | appraisal by the study's real design, with the matching tool | 0.2.0 |
  70. | 2 | `guideline-check` | current clinical and reporting guidelines, each with version and date | 0.2.0 |
  71. | 2 | `literature-review`, `/orchestra:review` | thesis Chapter 1 or manuscript background, synthesised by question (`review_lint.py`, `cite_render.py`) | 0.2.0 |
  72. | 3 | `protocol-design` | design with trade-offs, eligibility, variables table, bias and confounding | 0.2.0 |
  73. | 3 | `sample-size-sap` | sample size with a sensitivity table (`samplesize.py`) and a statistical analysis plan with an analysis register | 0.2.0 |
  74. | 4 | `data-preparation` | codebook, data audit, cleaning log, de-identified data in `04_data/` | 0.3.0 |
  75. | 4 | `statistical-analysis`, `/orchestra:analyze` | analysis per the SAP register with a vendored `crostats.py`, or SPSS/Stata/R output imported; declared post-hoc analyses; results ledger (`analysis_plan.py`, `results_ledger.py`) | 0.3.0 |
  76. | 5 | `results-writing`, `/orchestra:results` | thesis Chapter 3 or manuscript Results from the locked ledger, with .docx (`render_results.py`, `check_numbers.py`, `build_docx.py`) | 0.3.0 |
  77. | 4 | `research-figures` | forest plot, Kaplan–Meier with numbers at risk, rates chart, STROBE/CONSORT flow from the locked ledger or the de-identified data (`figures.py`) | 0.4.0 |
  78. | 6 | `conference-abstract` | structured abstract within the word limit, every number a ledger placeholder (`lint_prose.py --count-words`, `check_numbers.py`) | 0.4.0 |
  79. | 6 | `research-poster`, `/orchestra:poster` | print poster in three layouts (editorial, pathway, figure-led) as PDF, PNG and PPTX, with print preflight (`poster_build.py`, `poster_preflight.py`) | 0.4.0 |
  80. | 6 | `defense-prep`, `/orchestra:defense` | consistency audit before any slide, thesis-defence deck, committee questions (`consistency_audit.py`, `pptx_from_spec.py`, `check_pptx.py`) | 0.4.0 |
  81. | 5–6 | `prose-humanizer` | rewrites AI-sounding prose, keeps every number, citation and hedge (`lint_prose.py`, `invariants.py`) | 0.4.0 |
  82. | 5–6 | agents `consistency-auditor`, `committee-simulator` | numbers, dates and terms checked across deliverables; likely committee or reviewer questions with answers from the project files | 0.4.0 |
  83. | 5 | `manuscript`, `thesis`, `case-report`, `systematic-review` | manuscript and thesis .docx, CARE case report, PRISMA review | P4 |
  84. | 3, 7 | `ethics-and-crf`, `journal-selection`, `reviewer-response` | ethics dossier and CRF, journal shortlist, response to reviewers | P5 |
  85. You can talk to the skills in Vietnamese or English. The language of research documents is
  86. a project setting (`project.yaml → language.output`), so a thesis can be written in
  87. Vietnamese while the manuscript from the same study is in English.
  88. ## What makes it different
  89. - **Verified references only.** Every reference enters through `search_lit.py add` and is
  90. checked by `verify_refs.py` against PubMed, Europe PMC or Crossref. Citations in text use
  91. `[@id]` keys that must exist in the ledger, and `cite_render.py` numbers them in Vancouver
  92. style from the ledger. A source with no PMID or DOI enters only as an author-supplied record
  93. (`add-manual`) and is listed as MANUAL, never as verified.
  94. - **Numbers from files.** Results come from a results ledger, not from the model's memory, and a
  95. sample size comes from `samplesize.py` with a sensitivity table, every parameter sourced or
  96. labelled as an assumption.
  97. Results enter text only as `{{R-…}}` placeholders filled from the locked ledger, and
  98. `check_numbers.py` blocks any typed number the ledger does not hold, at its printed precision (see
  99. `judgement-defaults.md`). `consistency_audit.py`
  100. flags numbers, dates and terms that disagree between the thesis, abstract, poster and slides.
  101. - **Patient data protected by default.** `phi-guard` knows Vietnamese identifier formats
  102. (CCCD/CMND, BHYT, record numbers, labelled names, dates of birth, addresses) as well as
  103. phones and emails. Detection is heuristic and not proof that a text is clean.
  104. - **Tested like code.** Unit tests for every script, and behaviour scenarios that were run
  105. first without the skills (RED) and then with them (GREEN).
  106. - **Honest about AI.** Manuscripts and theses get an AI-use disclosure; the humanizer
  107. removes tells of machine prose without hiding that AI helped.
  108. ## Install
  109. ### Claude Code
  110. ```
  111. /plugin marketplace add drtranquanghung91-vn/clinical-research-orchestra
  112. /plugin install phi-guard@clinical-research-orchestra
  113. /plugin install orchestra@clinical-research-orchestra
  114. ```
  115. Then start a project:
  116. ```
  117. /orchestra:start
  118. ```
  119. ### Tools the scripts use
  120. Python 3.10 or later. The harness scripts use the standard library only. Literature lookups call
  121. the public PubMed, Europe PMC and Crossref APIs; no API key is needed. Analysis scripts written by
  122. `statistical-analysis` need `pandas`, `scipy` and `statsmodels` (`lifelines` for Cox models):
  123. `python -m pip install pandas scipy statsmodels openpyxl lifelines`.
  124. Optional, for presenting (P3): `matplotlib` for figures, `python-pptx` for the defence deck and the
  125. poster PPTX (`python -m pip install matplotlib python-pptx`), and Google Chrome or Microsoft Edge
  126. (headless) for the poster PDF, PNG and print preflight. A script that needs a missing one prints the
  127. install hint and exits 2; the skills report that check as not run.
  128. ## Roadmap
  129. | Phase | Scope | State |
  130. |---|---|---|
  131. | P0 Foundation | project folder, ledgers, gates, `phi-guard`, `verify_refs.py`, `search_lit.py`, CI | done (0.1.0) |
  132. | P1 Plan the study | research question, literature search, appraisal, guideline check, literature review, protocol, sample size and SAP | done (0.2.0) |
  133. | P2 Analyse and report results | data preparation, statistical analysis (run and import), results ledger, results writing, Chapter 3 .docx | done (0.3.0) |
  134. | P3 Present | figures, conference abstract, poster with print preflight, defence deck, prose humanizer, consistency audit | done (0.4.0) |
  135. | P4 Write | manuscript, thesis, case report, systematic review | next |
  136. | P5 Extend | ethics and CRF, journal selection, reviewer response, Codex and Antigravity packaging | planned |
  137. The design spec is in [`docs/superpowers/specs/`](docs/superpowers/specs/) and each phase
  138. has its implementation plan in [`docs/superpowers/plans/`](docs/superpowers/plans/).
  139. ## Repository layout
  140. ```
  141. .claude-plugin/marketplace.json marketplace manifest (plugins: orchestra, phi-guard)
  142. plugins/orchestra/ research skills, commands, scripts/, templates/project/
  143. plugins/phi-guard/ outbound PHI hook, phi_scan.py, deidentify.py, deidentify-data skill
  144. examples/fictional-cohort/ a simulated project run through every P1 and P2 skill, with its checks
  145. tests/unit/ tests for every script
  146. tests/scenarios/ RED/GREEN behaviour scenarios and the baseline record
  147. tools/lint_skills.py skill frontmatter and length checks
  148. docs/superpowers/ design spec and phase plans
  149. AGENTS.md conventions for agents working on this repository
  150. ```
  151. ## Testing
  152. ```bash
  153. python -m unittest discover -s tests/unit -v
  154. python tools/lint_skills.py
  155. ```
  156. CI runs both on Linux and Windows with Python 3.10 and 3.12 on every push.
  157. The skills are written test-first, as Superpowers prescribes. Each behaviour is first run on
  158. an agent without the plugin, and the failures are recorded in
  159. [`tests/scenarios/BASELINE.md`](tests/scenarios/BASELINE.md). The skills are then written to
  160. fix those failures and the scenarios re-run with them loaded
  161. ([`tests/scenarios/README.md`](tests/scenarios/README.md)).
  162. ## Authors
  163. - **Trần Quang Hưng**: owner, architecture, harness and scripts.
  164. - **Amelia Le** ([@ngoclnl189-amelia](https://github.com/ngoclnl189-amelia)): co-author of the research-methods skills (research question, literature and
  165. evidence, critical appraisal, protocol design, sample size and SAP) and their reference
  166. material.
  167. See [CONTRIBUTORS.md](CONTRIBUTORS.md).
  168. ## Acknowledgements
  169. - [google-research/paper-orchestra](https://github.com/google-research/paper-orchestra):
  170. outline-as-plan, two-stage citation verification, reviewer accept/revert loop (ideas only;
  171. no code or prompt text copied)
  172. - [obra/superpowers](https://github.com/obra/superpowers): skills as methodology, TDD for
  173. skills, subagent-driven development
  174. - [drtranquanghung91-vn/grand-rounds](https://github.com/drtranquanghung91-vn/grand-rounds):
  175. the literature API client, reference verification and PHI patterns this harness builds on;
  176. use it for medical talks beyond the thesis-defence path
  177. - The `humanizer` skill: the pattern catalog behind `prose-humanizer`
  178. Details and licences are in [NOTICE.md](NOTICE.md).
  179. ## Disclaimer
  180. Clinical Research Orchestra helps plan, write and present research. It does not give medical
  181. advice and does not replace a supervisor, a statistician or an ethics committee. The author
  182. is responsible for the study, for checking every claim against its source, for ethics
  183. approval and patient consent, for de-identifying data, and for disclosing AI use as their
  184. institution and journal require.
  185. ## Licence
  186. [MIT](LICENSE) © 2026 Trần Quang Hưng and contributors.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.