ClinicalGuidelineDevelopmentLlm Agent

<!--
# COPYRIGHT NOTICE
# This file is part of the "Universal AI Agentic Skills" project.
# Copyright (c) 2026 MD BABU MIA, PhD <md.babu.mia@mssm.edu>
# All Rights Reserved.
#
# This code is proprietary and confidential.
# Unauthorized copying of this file, via any medium is strictly prohibited.
#
# Provenance: Authenticated by MD BABU MIA
-->



<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->

---
name: 'clinical-guideline-development-llm'
description: 'Guide LLM-assisted clinical practice guideline drafting with real-time review, evidence traceability, recommendation grading, and clinician governance.'
measurable_outcome: 'Execute skill workflow successfully with valid output within 15 minutes.'
allowed-tools:
  - read_file
  - run_shell_command
  - web_fetch
---

# Clinical Guideline Development LLM

## Overview

This skill guides LLM-assisted clinical practice guideline development from question framing through evidence-linked recommendations, review, and governance. It emphasizes transparent evidence handling, recommendation grading, consensus documentation, auditability, and human clinical oversight because guideline development affects care standards and must not be treated as ordinary drafting.

## When to Use This Skill

- Drafting, updating, or revising a clinical practice guideline, protocol, pathway, or consensus statement.
- Converting evidence summaries into guideline recommendations with explicit strength, certainty, and rationale.
- Building evidence-to-decision tables, recommendation matrices, implementation notes, or audit trails.
- Running real-time clinician or panel review of LLM-generated guideline text.
- Checking guideline drafts for evidence traceability, conflicts, scope drift, unsafe wording, or missing governance steps.
- Preparing clinician-facing review packets that separate evidence, judgment, consensus, and implementation considerations.

## Core Capabilities

1. **Scope and question framing**: Define population, intervention, comparator, outcomes, care setting, exclusions, and intended users before drafting recommendations.
2. **Evidence traceability**: Link every material claim and recommendation to cited evidence, evidence tables, or an explicit expert-consensus rationale.
3. **Recommendation grading**: Record certainty of evidence, balance of benefits and harms, values and preferences, resource considerations, equity, feasibility, and recommendation strength.
4. **LLM-assisted drafting controls**: Use the model for synthesis, wording, consistency checks, and gap detection while preserving source-grounded claims and clinician review.
5. **Real-time review workflow**: Track reviewer comments, disagreement, edits, unresolved issues, and rationale changes during live or iterative guideline development.
6. **Real-time LLM evaluation during drafting**: Maintain live evidence traceability, recommendation grades, clinician adjudication of model-suggested changes, bias and omission checks against source evidence, and audit logs from draft through final guideline approval.
7. **Draft-to-final evaluation workflow**: For real-time LLM guideline development, preserve change-level evidence traceability, recommendation-grade updates, conflict logs, clinician review checkpoints, and audit artifacts explaining accepted, rejected, or revised model contributions.
8. **Real-time guideline-panel safeguards**: During live LLM-assisted development, run evidence traceability and recommendation grading checks, keep human guideline-panel oversight over accepted wording, log hallucinations or errors with corrections, and escalate unsupported, unsafe, overconfident, or grading-inconsistent generated language for clinician rewrite.
9. **Clinician-in-the-loop drafting evaluation**: During recommendation drafting, require clinician review of model-generated language, live evidence traceability, recommendation grading, conflict-of-interest checks, and explicit limits that LLM text remains draft support until human approval.
10. **Real-time evaluation checkpoints**: At each panel review checkpoint, verify evidence traceability, recommendation grading, hallucination audit findings, and versioned rationale logs before moving LLM-assisted guideline text to the next draft state.
11. **Guideline text acceptance gates**: Before accepting LLM-assisted guideline text, confirm evidence traceability, recommendation grading, iterative clinician review disposition, conflict logging, and governance checkpoint signoff.
12. **Expert adjudication advance criteria**: Before LLM-generated draft text advances to clinician review, require source-linked evidence traceability, recommendation-grade consistency, hallucination and unsupported-claim checks, expert adjudication of disputed wording, and documented governance criteria for promotion or revision.
13. **Real-time LLM evaluation handoffs**: Before panel handoff, check evidence traceability, recommendation grading, conflicts, and omissions; label LLM output as drafting assistance only, with guideline authority reserved for documented panel review and approval.
14. **Reject-or-escalate governance criteria**: During real-time LLM evaluation for guideline development, reject or escalate generated text when evidence traceability is missing, recommendation grading is inconsistent, reviewer checkpoint disposition is unresolved, contradictions are logged against source evidence or prior recommendations, or required governance approval is absent.
15. **Real-time LLM-assisted guideline development pattern**: Maintain evidence traceability and recommendation grading through iterative expert review, require conflict-of-interest disposition, retain audit logs of model suggestions and panel decisions, and reject unsupported recommendations lacking source or consensus justification.
16. **Real-time recommendation drafting audits**: Audit recommendation drafts for source-linked evidence traceability, recommendation-grade consistency, panel review checkpoint disposition, and unsupported or overconfident language; do not allow unsupervised LLM-generated guideline language to advance without documented panel review and approval.
17. **Real-time LLM-assisted guideline governance**: In live drafting, require evidence traceability, recommendation grading, clinician editorial control, audit logs, conflict handling, and explicit reject-or-revise gates for generated guideline text that is unsupported, grading-inconsistent, unsafe, unresolved after review, or lacking required approval.
18. **Draft adoption evaluation checkpoints**: Before any draft guideline is adopted, confirm evidence traceability, recommendation grading, live expert review disposition, disagreement logging, and governance gate approval.
19. **Consensus and governance**: Document panel composition, voting or consensus method, conflict-of-interest handling, signoff authority, and escalation criteria.
20. **Safety and implementation checks**: Identify ambiguous directives, unsafe absolutes, missing contraindications, patient subgroup concerns, monitoring needs, and update triggers.
21. **Audit-ready output**: Produce dated outputs with version history, evidence links, reviewer decisions, unresolved assumptions, and required human approvals.
22. **Real-time LLM evaluation record**: Across draft iterations, verify evidence traceability and recommendation grading, require expert adjudication of model-suggested recommendations or disputed edits, perform risk-of-harm review for unsafe or overconfident wording, and retain audit logs of prompts, outputs, edits, reviewer decisions, and rationale changes.
23. **Prospective guideline review checkpoints**: During real-time LLM-assisted guideline development, use prospective panel review checkpoints to confirm evidence traceability, adjudicate generated text, verify recommendation grading, and log failure modes such as unsupported claims, omissions, unsafe wording, or grading conflicts before accepting draft language.
24. **Live expert review acceptance criteria**: Use live expert review loops to compare LLM-suggested guideline language against source evidence, recommendation grades, and panel judgment; log discrepancies and treat LLM assistance as acceptable only when evidence traceability, grading consistency, reviewer disposition, and governance signoff are complete, escalating unsafe or unsupported output.
25. **Documented acceptance checkpoints**: Before guideline text is accepted, verify evidence traceability, recommendation grading, expert adjudication, bias and omission review, and documented human sign-off.
26. **Limits on autonomous guideline drafting**: At defined expert adjudication checkpoints, verify evidence traceability and recommendation grading, review generated content for bias, hallucinations, and unsupported claims, and prohibit the LLM from autonomously drafting, approving, or issuing guideline recommendations without documented expert review and sign-off.
27. **Concurrent expert evaluation protocol**: Review LLM-generated text concurrently with guideline drafting, trace claims to evidence, classify identified errors, verify recommendation grading, and adjudicate disagreements; prevent any unreviewed LLM text from entering the guideline or advancing toward approval.
28. **Worked real-time evaluation protocol**: Prospectively log prompts, outputs, edits, elapsed drafting and review time, evidence-verification results, recommendation grades, reviewer disagreements and adjudications, and errors by defined taxonomy; compare recorded time with the selected non-LLM workflow without presuming savings, escalate unsupported evidence, grading conflicts, unresolved disagreement, unsafe language, or high-impact errors, and keep LLM drafting assistance separate from final recommendation and approval decisions reserved for the guideline panel.
29. **Prospective real-time evaluation design**: During live guideline development, retrieve evidence as questions arise, preserve recommendation-to-source traceability, require expert adjudication of generated drafts, log reviewer disagreements and their disposition, measure retrieval, drafting, and review latency, and stop the workflow when a draft is unsafe, unsupported by retrieved evidence, or cannot be resolved through expert review.

## Inputs / Outputs

**Inputs**

- Guideline topic, scope, target population, setting, intended users, and clinical questions.
- Source evidence, citations, systematic reviews, trial summaries, existing guideline text, or evidence tables.
- Preferred grading framework, consensus method, review timeline, and required institutional governance steps.
- Reviewer identities or roles, conflict-of-interest declarations, comments, decisions, and approval requirements.

**Outputs**

- Structured guideline draft with scope, definitions, recommendations, rationale, implementation notes, and review status.
- Evidence-to-recommendation table linking each recommendation to evidence, certainty, strength, and rationale.
- Traceability matrix mapping claims to sources or consensus justification.
- Reviewer decision log with comments, changes, unresolved issues, and escalation items.
- Final human-review checklist covering evidence support, grading, conflicts, safety language, governance, and update plan.

## References

- Erstad BL. Real-Time Evaluation of a Large Language Model for Clinical Practice Guideline Development. *Crit Care Explor*. 2026 May 1. https://pubmed.ncbi.nlm.nih.gov/42042855/