research-grounded-advice

name	research-grounded-advice
description	Evidence-based recommendation protocol — search before advising, cite sources, cross-reference claims, and never hallucinate health/parenting/financial guidance. Use when user says "research this first", "cite sources", "evidence-based", "verify before suggesting", "don't make things up", "backed by research", or any agent needs to give health, medical, parenting, or safety advice.

Research-Grounded Advice Skill

Standard protocol for agents that give health, medical, parenting, or safety recommendations. Ensures every suggestion is backed by real sources, never hallucinated.

Why This Exists

Multiple agents (wellness-coach, parenting-coach, health-coach, nicu-care, fitness-coach) give recommendations that affect human health and wellbeing. Hallucinated medical advice is dangerous. This skill enforces the "search first, cite always" pattern.

The Protocol

Step 1 — Search Before Advising

Before providing ANY health/medical/parenting/safety recommendation:

1. perplexity-search(query: "[specific question]")       # FIRST CHOICE — best for factual answers with citations
2. exa-web_search_exa(query: "[broader context]")         # SECOND CHOICE — clean content, good for deep dives
3. exa-crawling_exa(urls: ["..."])                        # Read full page content from URLs found above
4. exa-get_code_context_exa(query: "...")                 # For technical/code research specifically

⚠️ TOOL PRIORITY (MANDATORY): Follow the research-tools skill (.github/skills/research-tools/SKILL.md) for the complete search tool hierarchy and parallel research patterns. Key rules:

• ALWAYS prefer Exa and Perplexity over web_search and web_fetch
• web_search and web_fetch are LAST RESORT only — they frequently fail and return poor results
• For complex research: use perplexity-reason or perplexity-deep_research
• For code/technical: use exa-get_code_context_exa + {{EMPLOYER_PARENT}} MCP tools
• For {{EMPLOYER}}/Azure topics: also use mslearn-microsoft_docs_search

Minimum: At least ONE search tool must be consulted before giving specific advice.

For significant recommendations (medication interactions, emergency guidance, developmental concerns):

• Cross-reference with at least 2 different search tools
• Prefer official sources: ACOG, APA, AAP, CDC, WHO, Postpartum Support International, peer-reviewed journals

Step 2 — Cite Sources in Response

Always attribute recommendations:

• ✅ "According to the AAP, skin-to-skin contact for at least 60 min/day improves..."
• ✅ "Research from [journal/org] suggests..."
• ✅ "Per ACOG guidelines, postpartum BP monitoring should..."
• ❌ "Studies show that..." (which studies? cite them)
• ❌ Making a specific claim with no attribution

Step 3 — Acknowledge Uncertainty

If you cannot verify a claim:

• ✅ "I haven't been able to verify this — let me research before suggesting"
• ✅ "This is commonly recommended but I couldn't find a primary source — check with your provider"
• ❌ Presenting unverified information as fact
• ❌ Saying "research shows" without having done the research

Step 4 — Source Hierarchy

When multiple sources conflict, prefer (in order):

1. Official clinical guidelines (ACOG, AAP, APA, CDC)
2. Peer-reviewed meta-analyses
3. Peer-reviewed studies (RCTs > observational)
4. Official organization recommendations (Postpartum Support International, KellyMom for lactation)
5. Expert consensus (major hospital systems, established practitioners)
6. High-quality journalism (avoid mommy blogs, forum posts, unvetted social media)

Scope — Which Agents MUST Use This Skill

Agent	When
wellness-coach	ALL recommendations (CBT techniques, medication info, coping strategies)
parenting-coach	Developmental guidance, behavioral strategies, sibling adjustment
health-coach	Medical tracking, postpartum guidance, medication management
nicu-care	NICU protocols, pumping guidance, preterm development
fitness-coach	Exercise safety, nutrition claims, supplement guidance

Anti-Patterns

• ❌ Inventing statistics ("90% of babies..." without a source)
• ❌ Giving medication dosage advice without verification
• ❌ Stating developmental timelines without citing AAP/CDC milestones
• ❌ Making claims about supplement efficacy without evidence
• ❌ Recommending treatments you cannot verify are safe for postpartum/breastfeeding

Exception: Known Validated Library

If the agent's core.md contains a pre-validated technique library (e.g., wellness-coach's CBT techniques), those can be used WITHOUT re-searching every time. But for ANY NEW suggestion beyond the library → search first.