health

IRIS Health Research

Healthcare research skill. This is the domain-specific skill for investigating health and medical topics: clinical evidence, pharmaceuticals, epidemiology, regulatory frameworks, public health, bioethics, health informatics, and any health-related subject. It guides users through topic definition, parameter configuration, and investigation launch via conversational interaction.

All outputs include mandatory disclaimers and evidence grading.

Workflow Routing

Workflow	Trigger	File
Investigate	"health research", "investigate disease", "analyze treatment"	workflows/investigate.md
Orchestrate	(internal — invoked by investigate)	workflows/orchestrate.md
Continue	"continue", "deepen"	../research/workflows/continue.md (shared)
Status	"health status", "check investigation"	../research/workflows/status.md (shared)
Archive	"archive health research"	../research/workflows/archive.md (shared)

How It Works

1. User invokes IRIS health research with a healthcare topic
2. Investigate workflow activates:
   • Checks for existing investigations (offers to continue or start new)
   • Gathers topic through conversational refinement
   • Verifies the topic is within healthcare scope
   • Confirms understanding with the user
   • Generates a request file in .requests/ with GUID-based ID
   • Creates a UUID run folder in .research/{uuid}/agents/
   • Transitions to the orchestrate workflow
3. Orchestrate workflow takes over:
   • Spawns health researcher agents in parallel
   • Monitors completion (mandatory wait gate — never skips agents)
   • Spawns the health-synthesizer agent
   • Status tracked in request file (researching → synthesizing → completed)
4. Results appear in .research/{uuid}/SUMMARY.md with disclaimers and evidence grading

Cumulative Research

Investigations support multiple runs for iterative deepening:

• Each run creates a new UUID folder under .research/
• The request file tracks all runs in research-runs
• Use /iris:research continue to deepen an existing investigation
• The continue workflow reads previous SUMMARY.md and injects context into new agents

Principles

Conversational Refinement

Never assume you understand the topic. Always:

1. Restate what you understood
2. Ask for confirmation using ask_user
3. Only proceed after explicit user approval

Progressive Disclosure

Start with the broad topic, then drill into specifics:

1. What health topic do you want to investigate?
2. What specific aspects interest you most? (clinical, regulatory, economic, etc.)
3. What questions should the investigation answer?
4. Is there a specific geographic/regulatory jurisdiction?
5. What language should the final report be in?

Evidence-Based Research

All health investigations must:

• Source information ONLY from official health organizations and peer-reviewed literature
• Grade every finding by evidence level (systematic review > RCT > cohort > case > expert opinion)
• Note publication dates and flag sources older than 3 years
• Include mandatory disclaimer in every output

Request File Contract

Every investigation produces a request file in .requests/ with:

• Full frontmatter (name, topic, language, status, domain: health)
• Detailed body describing the investigation scope
• Updated by the orchestrator as the investigation progresses

Context Files

• See workflows/investigate.md for the main investigation workflow
• See workflows/orchestrate.md for agent orchestration
• See workflows/archive.md for the archival workflow
• See workflows/status.md for checking investigation status

Healthcare Guardrails

This skill enforces strict healthcare research guardrails in addition to standard IRIS guardrails:

Standard IRIS Guardrails

• The ONLY writable directory is .research/ and .requests/
• No code execution that modifies the system
• No privilege escalation
• No network requests to external APIs on behalf of targets
• No destructive filesystem operations

Healthcare-Specific Guardrails

1. Verified sources only — ONLY official health organizations (WHO, FDA, EMA, NIH, CDC, PAHO) + peer-reviewed research (PubMed, Cochrane, NEJM, Lancet) + university research with DOI
2. Mandatory disclaimer — Every output must include: "This is research output, NOT medical advice. Consult a healthcare professional."
3. Evidence grading — Classify every finding: systematic review > RCT > cohort study > case study > expert opinion
4. Date sensitivity — Note publication date of every source; flag if older than 3 years
5. No prescriptions — Never recommend dosages, treatments, or diagnoses
6. No personal health advice — Research is informational, never prescriptive
7. Explicit jurisdiction — Always clarify geographic and regulatory context
8. Conflict of interest — Note if studies were industry-funded
9. Retraction awareness — Verify cited papers have not been retracted or corrected
10. Bias detection — Flag methodological limitations (sample size, study design, population)
11. Zero patient data — Never include patient-identifiable information