icu-clinical-consult
Skill1 starsChanged 54 days agoMITView on GitHub
Multi-agent adversarial clinical consult for pediatric (and adult) critical care questions. Triggers on ANY clinical question involving a patient scenario, physiologic mechanism, ventilator management, hemodynamics, ECMO, shock, sepsis, neurocritical care, renal, cardiac perioperative care, antimicrobial selection, or any bedside ICU question. Also triggers for mechanism-level physiology questions (e.g., "how does X affect Y") even without a specific patient. Trigger phrases include: "clinical question," "I have a patient," "what would you do with," "management of," "mechanism of," "physiology of," any drug/vent/hemodynamic question, antimicrobial spectrum/coverage question, differential diagnosis, "what's the evidence for," or any scenario that a PICU attending would reason through. Use this skill even for seemingly simple clinical questions — err on the side of triggering. Do NOT use for grant writing, manuscript editing, literature reviews, data analysis, or non-clinical academic tasks.
In this repositoryOne other file it gives its agents.
- fabrication-auditSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. About 0 are common, 11 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈11
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 30
All output is conversational, colleague-to-colleague at attending intensivist level: no hedging, no disclaimers, no "consult your …
- Very specific: line 50
Before any agent runs, ask up to 5 questions (minimum 3 for Mode A, minimum 1 for Mode B/C) via ask_user_input (claude.ai) / …
- Distinctive: line 70
Ask only what is clinically relevant and not already provided: the 3–5 most decision-altering data gaps, not an interrogation.
- Very specific: line 77
Ask 1–3 questions: age range (neonatal/pediatric/adult)?
- Distinctive: line 82
Ask 1–3 questions: community-acquired vs healthcare-associated?
- Very specific: line 88
Before any spectrum-of-coverage argument, pull current surveillance data.
- Distinctive: line 106
If the pull cannot be completed, qualify spectrum claims qualitatively ("broader Gram- negative coverage") rather than quantitatively …
- Very specific: line 149
Run the same four agents sequentially but independently: adopt each persona in turn from references/agent-panel.md, holding the isolation …
- Distinctive: line 167
Before synthesis, identify tensions between agents, one sentence each, feeding the "Controversies & Tensions" section: Agent 1 vs Agent 2 …
- Distinctive: line 185
Run the fabrication-audit skill's Step 1–5 procedure against every specific numeric claim, including (non-exhaustive): antimicrobial …
- Distinctive: line 238
Use Open Evidence Deep Consult style: concise, definitive, physiology-forward.
- Distinctive: line 288
Colleague tone: no disclaimers, no "this is not medical advice." Fellow attending.
- Distinctive: line 300
No false certainty: if the evidence is weak or conflicting, say so directly.
- Distinctive: line 308
After synthesis, stress-test the recommendation with 2 internal adversarial rounds.
- Distinctive: line 362
Antimicrobial spectrum: always invoke Phase 0.5 before a spectrum argument, even for a seemingly simple "does X cover Y" question.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Multi-Agent Adversarial ICU Clinical Consult
Overview
Four specialized agents analyze a clinical question independently and adversarially; a Synthesis agent merges their outputs into one physiology-forward, guideline-aware response delivered entirely in chat. A Citation & Numeric-Claim Validator agent runs final verification of both academic citations and specific numeric clinical claims (resistance percentages, mortality rates, NNT/NNH, prevalence, effect sizes).
Deterministic path: extras/icu-consult.js (Workflow tool) runs the same four reasoners in parallel → synthesizer → fabrication-auditor → citation-verifier, each dispatched by agentType. The model-driven path here is the default for interactive consults. Decision support for a licensed clinician; never autonomous orders.
All output is conversational, colleague-to-colleague at attending intensivist level: no hedging, no disclaimers, no "consult your physician" (Writing Rule 1).
Mode Detection
Mode A — Patient Scenario: a specific patient (age, vitals, labs, imaging, context). Full adversarial workup: executive brief, diagnostics, treatment.
Mode B — Mechanism/Physiology Question: a mechanism, drug effect, or physiologic principle without a specific patient. Open Evidence Deep Consult style: pathophysiology- forward, concise, evidence hierarchy.
Mode C — Antimicrobial Spectrum/Selection: drug-bug coverage, spectrum comparison, or antibiotic selection ("does X cover Y?", "compare A vs B for Z"). Own pre-flight gate, Phase 0.5 below.
Detect mode from the first message. If ambiguous, default to Mode A.
Phase 0: Intake (3–5 Clarifying Questions)
Before any agent runs, ask up to 5 questions (minimum 3 for Mode A, minimum 1 for
Mode B/C) via ask_user_input (claude.ai) / AskUserQuestion (Claude Code) for discrete
options, or free-text.
Mode A — Required Information Checklist
Gather what's not already provided:
| Domain | Key Data Points |
| Demographics | Age, weight, sex, comorbidities |
| Acute presentation | Chief complaint, timeline, precipitant |
| Vitals/Monitoring | HR, BP (arterial if available), SpO2, EtCO2, temp, CVP/ScvO2 |
| Respiratory | Vent mode/settings or support, last ABG, CXR/CT |
| Hemodynamics | Vasoactives (drug/dose/trajectory), echo, lactate trend |
| Labs | CBC, BMP, LFTs, coags, inflammatory markers, cultures |
| Neuro | GCS/exam, pupils, EEG if relevant, imaging |
| Renal | UOP, Cr trend, fluid balance, CRRT if running |
| Current therapies | Antibiotics, sedation, blood products, ECMO config |
Ask only what is clinically relevant and not already provided: the 3–5 most decision-altering data gaps, not an interrogation.
Acuity-aware intake (Mode A): first name the values that would change the recommendation. If weight, trajectory, support and the decision are all given and no deciding value is missing, ask zero questions and answer; the minimum of 3 does not apply. If the case is time-critical AND the recommendation holds across the plausible values of a missing item, proceed and write one Assumed: line naming the item and the value taken. Otherwise ask only the 1 to 3 values that decide the recommendation, urgent or not, and skip the rest of the checklist. Never state that a deciding value is unknown, and never assume it silently, instead of asking.
Mode B — Clarifying the Question
Ask 1–3 questions: age range (neonatal/pediatric/adult)? clinical context driving the question? a specific intervention or comparison in mind?
Mode C — Antimicrobial Spectrum/Selection
Ask 1–3 questions: community-acquired vs healthcare-associated? local antibiogram or general surveillance? severity (mild source-controlled vs severe sepsis/septic shock)? organism concerns (Pseudomonas, Enterococcus, Candida, MRSA, ESBL)?
Phase 0.5: Antimicrobial Spectrum Pre-Flight Gate (Mode C only)
Before any spectrum-of-coverage argument, pull current surveillance data. Hard gate: spectrum
claims not grounded in this turn's tool calls are not permitted. (Case history:
references/history.md.)
Required tool calls before Phase 1 runs for Mode C:
- PubMed search for the most recent surveillance paper on the drug-organism pair, matched to geography (US vs international); default to the most recent multicenter surveillance (Snydman/Tufts for Bacteroides; CANWARD Canadian; SENTRY global; CDC/NHSN US healthcare-associated).
- Web search for current IDSA/SIS/WSES guidance if the question concerns an empiric regimen.
- Confirm geography match; if only international data exists, say so explicitly and do not use it to defend a US-context claim.
Output: an evidence table (paper, geography, year, organism, drug, susceptibility/ resistance %) with a tool-call provenance (PMID or URL from this turn) on every row.
If the pull cannot be completed, qualify spectrum claims qualitatively ("broader Gram- negative coverage") rather than quantitatively ("covers >95% of Pseudomonas"). This applies even to a simple "does X cover Y" question.
Phase 1: Independent Agent Analysis
Read references/agent-panel.md now. It holds the four agents' personas, tasks, tool
lists, search strategies, and internal output formats; Phase 1 needs it on either path.
Roster cap (hard)
The panel is exactly four independent reasoners plus one synthesizer: no fifth reasoner, no duplicate personas, no per-sub-question variants. A broader question widens each brief, not the roster.
The Phase 3 gate agents (fabrication-auditor, citation-verifier) and the Phase 0.5
evidence pull are not counted against this cap and are never optional: they check claims
against external tools, not this context, so they are not self-verification.
Execution mode — subagent dispatch (preferred) vs. inline (fallback)
If the Agent/Task tool is available, dispatch the four agents as INDEPENDENT subagents,
in parallel (single message, multiple Agent calls), each passed only the clinical data and
question, not each other's output. Agent files live in this plugin's agents/ directory (Claude Code) or
.claude/agents/ (Cowork):
| Skill Agent | Subagent |
| 1 Pediatric Intensivist (Clinical Reasoning) | icu-clinical-reasoner |
| 2 Guideline & Evidence Synthesizer | icu-guideline-synthesizer |
| 3 Pathophysiology — Human & Translational | icu-pathophysiology-human |
| 4 Pathophysiology — Animal & Basic Science | icu-pathophysiology-basic |
Then dispatch icu-consult-synthesizer (Phase 2 cross-check + Phase 4 synthesis + Phase 4.5
convergence, fed all four agents' notes). Phase 3 validation runs on the gate agents:
fabrication-auditor (every dose/threshold/%) then citation-verifier (every identifier),
dispatched before the synthesizer surfaces anything. clinical-skeptic is a consolidated
adversarial alternative. All agents carry the decision-support-only safety guardrail.
If the Agent/Task tool is NOT available, run the inline pipeline below unchanged.
Inline pipeline
Run the same four agents sequentially but independently: adopt each persona in turn from
references/agent-panel.md, holding the isolation by hand (clinical data and question only,
never another agent's notes), producing the internal notes that file specifies.
The panel (detail in references/agent-panel.md)
| # | Agent | Remit in one line |
| 1 | Pediatric Intensivist (Clinical Reasoning) | Bedside attending: frame, differential, decision-altering data, management, 3 AM pitfalls |
| 2 | Guideline & Evidence Synthesizer | Society recommendations, systematic reviews, conflicts, evidence gaps, guidelines >5y stale |
| 3 | Pathophysiology — Human & Translational | Organ-system → cellular mechanism, peds-first human evidence, physiology vs guidelines |
| 4 | Pathophysiology — Animal & Basic Science | Animal models, molecular pathways, translational validity, emerging mechanisms |
Personas, tasks, search strategies, the evidence hierarchy, and internal output formats live
in references/agent-panel.md.
Phase 2: Adversarial Cross-Check
Before synthesis, identify tensions between agents, one sentence each, feeding the "Controversies & Tensions" section: Agent 1 vs Agent 2 (clinical reasoning vs guidelines), Agent 3 vs Agent 2 (human physiology vs guidelines), Agent 4 vs guidelines/trials (basic science suggesting something not yet established), and Agent 1's management plan vs Agents 3–4's mechanism (unsupported steps).
Phase 3: Citation & Numeric-Claim Validation
This phase validates two distinct classes of evidence in the draft.
3A. Academic citation validation
Validate all citations with the citation-verification skill's PMID-first protocol: every
reference needs a confirmed PMID via PubMed MCP (get_article_metadata) before inclusion.
Society guidelines (AHA, SCCM, ELSO, IDSA) may use URL when not PubMed-indexed. A failed
validation drops the citation; state the claim as "physiologic reasoning" or "clinical
experience" instead. Maximum 10–15 citations per response.
3B. Numeric clinical-claim validation
Run the fabrication-audit skill's Step 1–5 procedure against every specific numeric claim, including (non-exhaustive): antimicrobial susceptibility/resistance percentages; mortality rates; NNT/NNH, ARR, RRR, hazard/odds ratios with CI; prevalence/incidence; test sensitivity/ specificity; thresholds quoted as established (driving pressure 15 cmH2O, PEEP > 10, lactate
4 — confirm the source study/guideline); and drug doses presented as "standard" or "usual."
Per claim: verified this turn → keep, cite source. Verified prior turn → cross-check
against that output. Stated by user → ground truth, no audit. Derived → keep, show
derivation. Unverified (training-data recall) → verify now; after two failed attempts,
drop the number for a qualitative formulation or mark [UNVERIFIED] inline.
Geography cross-check (Mode C priority): surveillance percentages must match the claim's geography; US ≠ Japanese ≠ European ≠ Indian resistance data. Confirm explicitly.
Phase 3B cannot be skipped: one unverified numeric claim after two search attempts gets
[UNVERIFIED] or is reworded to remove the specific.
Phase 4: Synthesis & Final Output
Merge all agent outputs into a single, coherent response: conversational prose, not a form or bullet-point soup, as if presenting at ICU bedside rounds to a colleague.
Mode A Output Structure (Patient Scenario)
## Executive Brief
[3–5 sentences: what's going on, what matters most, what to do right now.]
## Clinical Reasoning
[Agents 1, 3, 4 woven in, mechanism first then this patient. Evidence
hierarchy peds → adult → animal → basic science, flagged.]
## Guideline Context
[Agent 2's key recommendations in narrative, not a list; note alignment
and conflicts with physiology.]
## Controversies & Tensions
[1–3 sentences, where agents disagreed.]
## Diagnostic Next Steps
[By decision-altering potential, each with rationale.]
## Treatment Plan
[Stepwise, physiologic rationale, pediatric weight-based dosing,
monitoring, reassessment triggers.]
## Key References
[10–15 max, PMID + DOI, by evidence tier: Pediatric | Adult
(extrapolation) | Animal/Basic Science]Mode B Output Structure (Mechanism/Physiology Question)
Use Open Evidence Deep Consult style: concise, definitive, physiology-forward.
## Bottom Line
[2–3 sentences, direct answer.]
## Pathophysiology
[Organ-system → cellular → molecular mechanism, the core of the response
(Agent 3 + Agent 4 merged). Evidence hierarchy peds → adult → animal →
basic science.]
## Clinical Implications
[Agent 1's clinical reasoning applied to the mechanism, bedside management.]
## Guideline Alignment
[Brief: where guidelines reflect this physiology and where they don't.]
## Key References
[5–10 max, PMID + DOI.]Mode C Output Structure (Antimicrobial Spectrum/Selection)
## Bottom Line
[2–3 sentences: the recommendation and the key reason.]
## Spectrum Comparison
[Table from Phase 0.5 evidence, every percentage with PMID/URL provenance
from this turn. Columns: organism class | Drug A | Drug B |
source/year/geography. Qualitative (yes/no/marginal) where quantitative
data wasn't available, noted explicitly.]
## Clinical Context
[Severity, source, healthcare-associated vs community, prior exposures,
local antibiogram if known: connect spectrum to this scenario.]
## Guideline Alignment
[IDSA/SIS/WSES recommendations for this scenario, stratified by severity;
note conflicts.]
## Recommendation
[Empiric regimen, dose, duration, de-escalation plan.]
## Key References
[Surveillance papers + guidelines, PMID/DOI.]Writing Rules
- Colleague tone: no disclaimers, no "this is not medical advice." Fellow attending; use "I'd" and "we" naturally.
- Physiology first: lead with mechanism; guidelines support the argument, they don't replace it.
- Evidence hierarchy always explicit: say when citing adult or animal data for a pediatric question. Never let the reader assume pediatric RCT evidence exists when it doesn't.
- Quantify: effect sizes, NNT/NNH, odds/hazard ratios when available, not "associated with worse outcomes" alone. Every quantitative claim goes through Phase 3B.
- Brevity: executive brief 3–5 sentences. Total response: Mode A ~800–1200 words, Mode B ~400–800, Mode C ~600–1000. These are ceilings on prose, not rigor: never drop a Phase 3A/3B check, geography qualifier, or evidence-tier flag to hit a word count.
- No false certainty: if the evidence is weak or conflicting, say so directly.
- Cite inline: superscript-style, keyed to Key References:
[Author, Journal Year]inline, full citation in references. - Surveillance percentages carry geography: e.g. "X% in [US Snydman 2017] vs Y% in [Japanese Ueda 2021]." Never collapse geographic sources into one number.
Phase 4.5: Convergence Testing (Karpathy Method)
After synthesis, stress-test the recommendation with 2 internal adversarial rounds.
Round 1: re-invoke Agent 3 or 4 as a fresh adversary: "what is the single strongest objection not already addressed?" A novel objection gets integrated (caveat, dose change, monitoring parameter) and Round 2 runs; none means converged, present it.
Round 2 (only if Round 1 changed something): re-invoke the adversary against the updated synthesis. A novel objection is flagged as "area of genuine uncertainty" alongside the recommendation; none means converged with high confidence.
Convergence indicator (append to response):
Convergence: [Robust — no unaddressed objections] OR [Moderate — one unresolved tension: {brief}]Tool Usage Summary
| Tool | When | Purpose |
| PubMed MCP: search_articles | Phase 0.5 (Mode C), Agents 2–4 | Guidelines, RCTs, mechanistic studies, surveillance |
| PubMed MCP: get_article_metadata | Phase 3 | Verify PMID, title, authors, DOI |
| Web search | Phase 0.5 (Mode C), Agent 2, validation | Society guidelines, DOI resolution, antibiograms |
| ask_user_input / AskUserQuestion | Phase 0 | Clarifying questions with discrete options |
Expected tool calls per consult: 8–15 total (3–5 PubMed searches, 2–3 web searches, 3–5 metadata lookups, 1 ask_user_input/AskUserQuestion). Mode C adds 2–4 surveillance-specific PubMed/web calls in Phase 0.5.
Integration with Other Skills
- This skill handles the clinical reasoning and evidence synthesis in chat.
- iterative-experimentation methodology drives Phase 4.5 convergence testing.
- fabrication-audit (Phase 3B: resistance %, mortality %, NNT, effect sizes) and citation-verification (Phase 3A: PMIDs, DOIs) split scope; run both.
- A formal document ask (treatment plan, clinical report) hands off to
treatment-plansorclinical-reports; a literature review ask hands off toliterature-review. prompt-optimizerstill runs silently on the initial question before this skill executes.
What This Skill Does NOT Do
- Generate formal PDF/docx clinical documents (use
treatment-plansorclinical-reports) - Replace clinical judgment: a thinking partner, not an oracle
- Provide definitive legal/medicolegal advice
- Manage real patient orders or EHR interactions
- Quote surveillance percentages from training data without verification (Phase 3B)
Edge Cases
- Neonatal: flag as neonatal, adjust physiology (transitional circulation, surfactant biology, NEC, IVH), use neonatal-specific guidelines (AAP, Vermont Oxford).
- ECMO: include ELSO guidelines, circuit physiology, and the relevant configuration (VA vs VV) in the clinical frame.
- Cardiac surgical: reference Ungerleider, single-ventricle physiology, Fontan/Glenn considerations where relevant.
- Antimicrobial spectrum: always invoke Phase 0.5 before a spectrum argument, even for a seemingly simple "does X cover Y" question.
- Informatics/monitoring with a clinical component: address the clinical question here, defer informatics architecture to other skills.
Case log, dated version notes, and the retired Known Failure Modes table live in
references/history.md — open it for why a given gate exists, not for current behavior.
Web fallback
WebFetch and WebSearch stay first. If the Parallel Search MCP is installed (claude mcp add --transport http parallel-search https://search.parallel.ai/mcp) and WebFetch returns a 4xx/5xx, a paywall or bot wall, a redirect it cannot follow, or a page missing the target content, retry once with mcp__parallel-search__web_fetch (same URL, a narrow objective); if WebSearch returns nothing usable, retry with mcp__parallel-search__web_search. Parallel returns query-matched excerpts, not the full page, so a value missing from the excerpt is unverified, not absent. Never put PHI, unpublished manuscript or grant text, or credentials in a Parallel query, and never use it to get past a login or CAPTCHA gate. Name the tool that verified each item.
Feedback loop
Found a missed edge case, a wrong-shaped output, or a rule that misfires? Open an issue on this plugin's repository with the input and the output you expected. Do not edit this skill mid-run. Per-run case facts stay in this skill's own case log / memory store; only skill-file changes go to the observation log.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

