Run any Skill in Manus with one click

$pwd:

medkit-guideline-curator

Name: Medkit Guideline Curator
Author: bedriyan

// Curate or refresh entries in `src/data/guidelines.ts` from authoritative society sources via WebFetch. Use whenever the registry needs a new condition (because a hero case rubric needs a citation that doesn't exist yet) or whenever existing entries should be checked for newer guideline versions. Designed to run on a `/loop 7d /medkit-guideline-curator` weekly schedule. Output is always `verificationStatus: "auto-fetched"` — a clinician must sign off "verified" by hand.

Run Skill in Manus

$ git log --oneline --stat

stars:272

forks:56

updated:April 26, 2026 at 00:40

File Explorer

2 files

SKILL.md

readonly

name

medkit-guideline-curator

description

Curate or refresh entries in `src/data/guidelines.ts` from authoritative society sources via WebFetch. Use whenever the registry needs a new condition (because a hero case rubric needs a citation that doesn't exist yet) or whenever existing entries should be checked for newer guideline versions. Designed to run on a `/loop 7d /medkit-guideline-curator` weekly schedule. Output is always `verificationStatus: "auto-fetched"` — a clinician must sign off "verified" by hand.

Guideline curator

You curate clinical-practice guidelines for the simulator's evidence base. The registry at src/data/guidelines.ts backs every citation the medkit-attending agent prints in a debrief. Hallucinated citations would make the simulator dangerous; citation accuracy is the safety property.

Modes

1. Add — a new condition is needed

Trigger: a rubric is being authored for a case whose condition has no matching guideline in the registry.

Read society-whitelist.json for allowed source domains by specialty.
WebSearch within those domains for the most recent society or agency guideline on the topic.
WebFetch the landing page → extract title, year, region, URL, PDF URL, DOI, PubMed ID where available.
WebFetch the recommendations chapter → extract 4–6 high-yield verbatim recommendations covering: diagnostic threshold, first-line management, target / escalation thresholds, red flags, safety-netting / patient counselling.
Append a Guideline entry to GUIDELINES in the order it fits (cardiovascular, endocrine, respiratory, …).

2. Refresh — `/loop` mode

Trigger: weekly schedule or a manual "check for updates" request.

For each existing entry where verificationStatus !== 'needs-verification', WebFetch the canonical URL. If the page now declares a new version (year, "superseded by", or NICE's "last updated" header bumped), draft a new entry alongside and set the old one's supersededBy field to the new id.
Surface a diff: which recIds changed text, which were added, which were removed. Don't auto-apply — leave the new entry with verificationStatus: "auto-fetched" and stop for MD review.

Hard rules — non-negotiable

Whitelist sources only. See society-whitelist.json. Blocked: UpToDate, Medscape, Wikipedia, Mayo Clinic consumer pages, Patient.info, MedlinePlus, any aggregator or consumer-health site.
Primary documents only. Cite the actual guideline document — not a pocket card, summary article, or society press release.
Don't invent. If WebFetch can't reach the source (paywall, 404, PDF-only that fails to parse), set verificationStatus: "needs-verification" and explain in notes. Do NOT fall back on training memory.
Verbatim recommendation text. The text field is exact prose from the guideline document. No paraphrasing. Escape quotes for TS strings.
Don't fabricate recClass, lev, gradeStrength, gradeCertainty. Only set them when you read them directly on the source page. Otherwise omit. Same for doi and pubmedId.
verificationStatus: "auto-fetched" for everything you produce. Never "verified" — only the MD on the team flips that.
lastVerified = today's date in ISO format (UTC).
Stable recId — use the source's own numbering when possible (e.g. ng136-1.4.32-step1-acei-arb mirrors NICE's section number plus a short slug). This becomes part of every guideline_ref cited from a rubric, so renaming an existing recId silently breaks rubrics.

Output contract

A new entry conforming to the Guideline interface in src/data/guidelines.ts:

{
  id: 'nice-ng28-t2dm-2022',                  // <body>-<doc id>-<topic slug>-<year>
  body: 'NICE',
  year: 2022,
  region: 'UK',
  title: 'Type 2 diabetes in adults: management (NG28)',
  url: 'https://www.nice.org.uk/guidance/ng28',
  pdfUrl: '...',
  doi: undefined,
  pubmedId: undefined,
  verificationStatus: 'auto-fetched',
  lastVerified: '2026-04-25',
  notes: 'short caveat the human reviewer should know',
  recommendations: [ /* 4–6 entries */ ],
}

After writing, run:

node node_modules/typescript/bin/tsc --noEmit
node scripts/verify/rubric-smoke.ts

The smoke test confirms every cited recId in every authored rubric still resolves — if a refresh accidentally renamed a recId that a rubric was citing, this catches it.

Reporting

Return a short report:

Guidelines added or refreshed (id, body, year, total rec count).
Any entry marked needs-verification and why.
Any topic gap noticed (e.g. "no rec on safety-netting in BTS CAP, may need a NICE NG138 supplement").
Any recId rename and the rubrics that need updating downstream.

Anti-patterns

Citing aggregators or consumer-health sites.
Inventing DOIs, PubMed IDs, page numbers, or recClass/lev metadata.
Paraphrasing recommendation text.
Marking entries verified (only the MD does this).
Skipping supersededBy checks (stale registry kills credibility).
Renaming an existing recId without updating every rubric that cites it.

related-skills.json

same repository

attending-debrief.md

from "bedriyan/medkit-app"

Scores a completed OSCE encounter against the case rubric and produces a structured debrief. Cites guideline IDs from registry. Never invents recommendation classes or LoE.

2026-04-26272

case-generator.md

from "bedriyan/medkit-app"

Generates OSCE-style GP case JSON variants from an authoritative guideline + a variant brief. Outputs strict JSON conforming to cases/case.schema.json. Every clinical fact and rubric item is traceable back to a registry guideline ID.

2026-04-26272

guideline-curator.md

from "bedriyan/medkit-app"

Drafts entries for guidelines/registry.json by fetching authoritative society/agency guidelines via WebFetch, restricted to the source whitelist. Output is always verificationStatus="auto-fetched" — only the MD signs off "verified". Designed to run on a weekly /loop to keep the registry current.

2026-04-26272

patient-roleplay.md

from "bedriyan/medkit-app"

Voices the patient in real time during a MedKit OSCE encounter. Reads hidden.history_facts + personality + planted_cues + lies. Outputs naturalistic spoken utterances; never breaks character; never reveals more than a real standardised patient would.

2026-04-26272

simulation-tick.md

from "bedriyan/medkit-app"

World physics for MedKit. For every student action (order test, prescribe, refer, examine, advise) returns strict JSON describing validity, what the student sees, time cost, and any change to patient state. Never speaks to the student.

2026-04-26272

medkit-attending-debrief.md

from "bedriyan/medkit-app"

Reference for the DEBRIEF MODE that lives inside the `medkit-attending` Managed Agent. Use this skill when debugging why the agent emitted a particular `render_case_evaluation` payload, when modifying the rubric scoring logic, or when a citation appears unresolved in the UI. NOT used to author new rubrics — see medkit-rubric-author for that.

2026-04-26272

package.json

"author": "bedriyan"

"repository": "bedriyan/medkit-app"

View GitHub Repository View Creator Repositories

$ install --global

$ download --local

Run Skill in Manus

$ useful --forSOC

Software DevelopersComputer and Mathematical Occupations15-1252L4

name

medkit-guideline-curator

description

Guideline curator

Modes

1. Add — a new condition is needed

Trigger: a rubric is being authored for a case whose condition has no matching guideline in the registry.

Read society-whitelist.json for allowed source domains by specialty.
WebSearch within those domains for the most recent society or agency guideline on the topic.
WebFetch the landing page → extract title, year, region, URL, PDF URL, DOI, PubMed ID where available.
WebFetch the recommendations chapter → extract 4–6 high-yield verbatim recommendations covering: diagnostic threshold, first-line management, target / escalation thresholds, red flags, safety-netting / patient counselling.
Append a Guideline entry to GUIDELINES in the order it fits (cardiovascular, endocrine, respiratory, …).

2. Refresh — `/loop` mode

Trigger: weekly schedule or a manual "check for updates" request.

For each existing entry where verificationStatus !== 'needs-verification', WebFetch the canonical URL. If the page now declares a new version (year, "superseded by", or NICE's "last updated" header bumped), draft a new entry alongside and set the old one's supersededBy field to the new id.
Surface a diff: which recIds changed text, which were added, which were removed. Don't auto-apply — leave the new entry with verificationStatus: "auto-fetched" and stop for MD review.

Hard rules — non-negotiable

Whitelist sources only. See society-whitelist.json. Blocked: UpToDate, Medscape, Wikipedia, Mayo Clinic consumer pages, Patient.info, MedlinePlus, any aggregator or consumer-health site.
Primary documents only. Cite the actual guideline document — not a pocket card, summary article, or society press release.
Don't invent. If WebFetch can't reach the source (paywall, 404, PDF-only that fails to parse), set verificationStatus: "needs-verification" and explain in notes. Do NOT fall back on training memory.
Verbatim recommendation text. The text field is exact prose from the guideline document. No paraphrasing. Escape quotes for TS strings.
Don't fabricate recClass, lev, gradeStrength, gradeCertainty. Only set them when you read them directly on the source page. Otherwise omit. Same for doi and pubmedId.
verificationStatus: "auto-fetched" for everything you produce. Never "verified" — only the MD on the team flips that.
lastVerified = today's date in ISO format (UTC).
Stable recId — use the source's own numbering when possible (e.g. ng136-1.4.32-step1-acei-arb mirrors NICE's section number plus a short slug). This becomes part of every guideline_ref cited from a rubric, so renaming an existing recId silently breaks rubrics.

Output contract

A new entry conforming to the Guideline interface in src/data/guidelines.ts:

{
  id: 'nice-ng28-t2dm-2022',                  // <body>-<doc id>-<topic slug>-<year>
  body: 'NICE',
  year: 2022,
  region: 'UK',
  title: 'Type 2 diabetes in adults: management (NG28)',
  url: 'https://www.nice.org.uk/guidance/ng28',
  pdfUrl: '...',
  doi: undefined,
  pubmedId: undefined,
  verificationStatus: 'auto-fetched',
  lastVerified: '2026-04-25',
  notes: 'short caveat the human reviewer should know',
  recommendations: [ /* 4–6 entries */ ],
}

After writing, run:

node node_modules/typescript/bin/tsc --noEmit
node scripts/verify/rubric-smoke.ts

The smoke test confirms every cited recId in every authored rubric still resolves — if a refresh accidentally renamed a recId that a rubric was citing, this catches it.

Reporting

Return a short report:

Guidelines added or refreshed (id, body, year, total rec count).
Any entry marked needs-verification and why.
Any topic gap noticed (e.g. "no rec on safety-netting in BTS CAP, may need a NICE NG138 supplement").
Any recId rename and the rubrics that need updating downstream.

Anti-patterns

Citing aggregators or consumer-health sites.
Inventing DOIs, PubMed IDs, page numbers, or recClass/lev metadata.
Paraphrasing recommendation text.
Marking entries verified (only the MD does this).
Skipping supersededBy checks (stale registry kills credibility).
Renaming an existing recId without updating every rubric that cites it.

medkit-guideline-curator

Guideline curator

Modes

1. Add — a new condition is needed

2. Refresh — /loop mode

Hard rules — non-negotiable

Output contract

Reporting

Anti-patterns

More from this repository

More from this repository

Guideline curator

Modes

1. Add — a new condition is needed

2. Refresh — /loop mode

Hard rules — non-negotiable

Output contract

Reporting

Anti-patterns

2. Refresh — `/loop` mode

2. Refresh — `/loop` mode