kaimingwan/oh-my-kiro/omk-research

Trigger Examples

• "@research AutoMQ vs Confluent 对比"
• "帮我调研一下这个库怎么用"
• "find out how competitors handle this"
• "what's the best practice for X in 2026?"
• "compare these three approaches"

Research Skill — Multi-Level Search

Search Level Strategy

Always use the lowest level that can answer the question:

Rule: If Level 0 or 1 can answer it, don't use Level 2.

Don't need research: Common knowledge, already in knowledge/, answerable from built-in knowledge.

Level 2: Tavily Research API

Prerequisites

Get your API key at https://tavily.com (1000 free credits/month)

Set environment variable:

export TAVILY_API_KEY="tvly-your-key-here"

Or add to your agent config:

{
  "env": {
    "TAVILY_API_KEY": "tvly-your-key-here"
  }
}

Usage

./scripts/research.sh '{"input": "your research query"}' [output_file]
 
# Quick research
./scripts/research.sh '{"input": "quantum computing trends"}'
 
# Deep research
./scripts/research.sh '{"input": "AI agents comparison", "model": "pro"}'
 
# Save to file
./scripts/research.sh '{"input": "market analysis", "model": "pro"}' ./report.md

Model Selection

Rule of thumb: "what does X do?" → mini. "X vs Y vs Z" → pro.

Structured Output

./scripts/research.sh '{
  "input": "fintech startups 2025",
  "model": "pro",
  "output_schema": {
    "properties": {
      "summary": {"type": "string", "description": "Executive summary"},
      "companies": {"type": "array", "items": {"type": "string"}}
    },
    "required": ["summary"]
  }
}'

Citation Formats

Supported: numbered (default), mla, apa, chicago

./scripts/research.sh '{"input": "climate impacts", "citation_format": "apa"}'

Post-Research 沉淀 Checkpoint

After completing research, before writing findings or recommendations:

Socratic validation (mandatory for each recommendation/gap/optimization):

1. Does this problem actually exist in the current codebase? Check existing solutions first.
2. Is the proposed fix feasible on all target platforms (Kiro + CC)? Check constraints.
3. Does the benefit outweigh the maintenance cost?

If any answer is "no" → drop that recommendation. Don't include it in findings.

Then persist:

1. Record validated findings in docs/plans/findings.md (if working on a plan)
2. If findings reveal reusable patterns → write to knowledge/episodes.md
3. Cite sources with URLs — no hallucinated references