| name | full-research |
| description | Comprehensive grounded research with rigorous source verification. Use when user needs thorough research with verbatim citations, fact-checking verification, or academic-level source grounding on complex topics. |
| argument-hint | ["topic"] |
| allowed-tools | WebFetch, Read, Task |
| user-invocable | true |
| context | fork |
| agent | Explore |
Grounded Research: $ARGUMENTS
Grounding Rules
- Use only information explicitly present in discovered sources
- Include inline citation for every factual claim: [Source: X, Quote: "..."]
- When information is absent from sources, state "Not found in provided sources"
- Classify each claim:
- 📄 STATED: Direct from source (with quote)
- 🔗 INFERRED: Logical connection between sources (explain reasoning)
- ❓ UNCERTAIN: Reasonable but unverified (flag explicitly)
Avoid: unsourced claims, meaning-altering paraphrases, "general knowledge" gap-filling, unstated assumptions.
Phase 1: Source Processing
Search for and read 3-5 high-quality sources.
Prioritize: official docs > meta-analyses > highly-cited papers > tutorials.
For each source, extract:
### Source [N]: [Title/URL]
Key Claims (verbatim quotes):
- "[Exact quote]" → Supports: [what this evidences]
- "[Exact quote]" → Supports: [what this evidences]
Source Limitations:
- Type: [docs/paper/blog/etc.]
- Potential bias: [if any]
Funding & Sponsorship:
- 💰 Funder/Sponsor: [Organization/company funding the study, or "Not disclosed"]
- 🏢 Author Affiliations: [Key affiliations that may indicate conflicts]
- ⚠️ Conflict of Interest: [Declared COIs, or "None declared", or "Not stated"]
- 🎯 Sponsor Alignment: [Does the conclusion align with sponsor's commercial interests? Yes/No/N/A]
Phase 2: Chain-of-Thought Synthesis
Re-anchor: every claim needs a verbatim quote and source number.
Connect sources systematically:
- Claim: [Synthesized claim]
- Evidence: [Source N, Quote: "..."]
- Confidence: High/Medium/Low
- Sponsorship Bias Risk: 🟢 None / 🟡 Possible / 🔴 Likely — [brief explanation if applicable]
- Reasoning chain: Source A states X → Source B confirms Y → Therefore Z
When multiple sources agree, verify they do not share the same funder — because shared funding can create correlated bias. Weight industry-funded studies with sponsor-aligned conclusions lower. Prefer independently funded or publicly funded research.
Phase 3: Grounded Verification (Subagent)
Use Task tool to spawn a verification subagent with clean context.
Pass to subagent:
- Synthesis claims (Phase 2 output)
- All source URLs and their extracted quotes
Subagent prompt:
You are a fact-checking agent. Verify these claims against provided sources.
CLAIMS TO VERIFY:
[paste synthesis claims]
SOURCES WITH QUOTES:
[paste source extracts]
For each claim:
1. Find supporting quote in sources
2. Check if quote actually supports claim (exact match vs stretched interpretation)
3. Flag any claim without direct source backing
Output format:
| Claim | Source Quote | Match Quality | Issue |
|-------|--------------|---------------|-------|
Flag as UNGROUNDED anything not directly traceable to a source quote.
SPONSORSHIP BIAS CHECK:
For each claim backed primarily by industry-funded sources:
1. Is there independent corroboration from non-industry sources?
2. Does the claim directly benefit the funding organization?
3. Rate sponsorship bias risk: 🟢 None / 🟡 Possible / 🔴 Likely
Integrate subagent findings back into your output.
Phase 4: Self-Critique
Re-anchor grounding rules: every claim needs a citation, no gap-filling, uncertainties flagged.
| Check | Pass/Fail |
|---|
| Every claim has citation? | |
| Quotes are verbatim? | |
| No "common knowledge" gap-filling? | |
| Uncertainties flagged? | |
| Funding/sponsorship identified per source? | |
| Industry-funded claims independently corroborated? | |
If any check fails, revise the relevant section before producing output.
Output Format
Output TL;DR first, before any phases or detailed analysis.
TL;DR
[2-3 sentence bottom-line answer. Plain language, no citations. What does the user need to know?]
Detailed analysis follows below.
Summary
[Key findings - every sentence with [Source: N] citation]
### Evidence Table
| Claim | Source | Verbatim Quote | Confidence | Bias Risk |
|---|
| [claim] | [N] | "[exact words]" | High/Med/Low | 🟢/🟡/🔴 |
Sponsorship & Bias Map
| Source | Funder/Sponsor | Author Affiliations | Declared COI | Conclusion Favors Sponsor? |
|---|
| [N] | [org] | [affiliations] | [yes/no/undisclosed] | [yes/no/N/A] |
Bias Warnings: [Flag any claims that rely primarily on industry-funded sources without independent corroboration]
What Sources Don't Cover
[Gaps - what wasn't found]
Sources
[List all sources with URLs]
Edge Cases
- Fewer than 3 sources found: State the gap explicitly, reduce confidence ratings, note limited evidence in TL;DR
- All sources share a funder: Flag correlated bias risk prominently, search for independent corroboration before synthesizing
- Subagent flags claims as UNGROUNDED: Remove or reclassify as ❓ UNCERTAIN with explanation — do not keep ungrounded claims in the final output
- Contradictory sources: Present both positions with quotes, do not pick a winner without evidence-based reasoning
Input: "Research the health effects of intermittent fasting"
Output: TL;DR first, then 3-5 source extracts with funding/bias metadata, synthesis claims with evidence chains, verification subagent table, self-critique checklist, and final structured output with Evidence Table and Sponsorship Map.
