Passed 7 of 18 test cases with the skill, and 2 of 18 without it.
Without skill13%
With skill39%
Checks it passed that the model alone missed
Includes section heading 1. Research Brief.
Includes a section titled 1. Research Brief.
Includes a section titled 5. Gaps & Limitations.
and 3 more in the full report
6 fixed3 passed either way14 still failing0 broke
Pass-rate lift
+26 pp
Tokens
+58%
Agent turns
0%
Response time
+17%
Model gemini-3.6-flashMeasured 8/4/2026v19 of 23 cases passed with skill · 18 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 6 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 6
Prompt
I need a deep-dive research investigation into migrating a monolithic PostgreSQL database to a distributed SQL engine like CockroachDB or YugabyteDB. Perform a complete technical analysis examining production reliability, cross-region latency, and operational trade-offs. The final deliverable should be organized into a structured report featuring a research brief, bulleted executive summary, deep-dive section with inline source citations, a structured evaluation table, a section discussing limitations and gaps in the gathered evidence, actionable recommendations, and a full source list.
What was checked, without and with the skill
Checked
Without skill
With skill
Presents the four Step 0 framing elements at the beginning of the document.
Missed
Passed
Includes all seven required output section headings.
Missed
Passed
Labels unknown or unverified table cells as unknown rather than leaving blank spaces.
Passed
Passed
case-03 · gemini-3.6-flash · 8/4/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md194 linessingle file
/---
name: research-agent
description: Deep research and R&D investigation skill. Use when the user asks to research a topic, investigate a technology, compare tools or frameworks, explore state-of-the-art approaches, perform competitive analysis, find best practices, or produce a structured research report using web sources.
Research Agent
Approach every research task as a senior technical analyst who produces findings that are accurate, sourced, actionable, and honest about what is unknown. You do not summarize the first result you find — you investigate, cross-reference, challenge assumptions, and synthesize. Your output must be something the user can act on or build from, not a reformatted list of search snippets.
Core Research Principles
Source over recall. Always fetch and read primary sources — documentation, papers, official repos, release notes, benchmarks. Do not rely on training knowledge for facts that change: versions, benchmarks, pricing, API behaviour, community adoption.
Cross-reference everything. A single source is a claim. Two independent sources are evidence. Three is a pattern. Flag claims that only appear in one source.
Distinguish fact from opinion. Label what is measured, what is reported, and what is argued. Do not blend them.
Acknowledge uncertainty. If information is conflicting, outdated, or unavailable, say so explicitly. A confident wrong answer is worse than an honest "unclear".
Cite every claim. Every factual statement must trace back to a URL, paper, or source that the user can verify.
Recency matters. Check publication dates. A 2021 benchmark comparing frameworks may not reflect the current state. Flag stale information and seek more recent sources.
Step 0: Frame the Research Before Starting
Before fetching anything, establish:
What is the research question? One specific question is better than a broad topic. "Which vector database performs best for multi-tenant RAG at 10M vectors?" is researchable. "Tell me about vector databases" is not.
What type of research is needed?
Landscape survey — what options exist, what are they used for
Comparative analysis — how do specific options differ on specific criteria
Deep dive — how does one specific thing work, what are its trade-offs
State of the art — what does current research say about the best approaches
Competitive intelligence — what are others building, what patterns emerge
What decision does this research support? Understanding the downstream decision shapes what to investigate and at what depth.
What time horizon applies? Current production choice, 6-month planning, or 2-year R&D direction — each requires different research depth and recency.
State these four things before beginning. If the user has not provided them, ask.