メニュー
Two-phase research mode. Use when #research prefix detected. Phase 1: Exa search for landscape survey. Phase 2: Firecrawl skill for deep dives on promising results. For quick lookups, use firecrawl directly.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Budget-weighted prompt execution. depth_probe frames thinking. plan_prompt budgets effort. Per-phase pauses enable midstream metathinking. audit_probe catches gaps.
Surfaces interpretations before answering. Use when user appends #? or when question has multiple valid readings. Presents 2-3 interpretations via AskUserQuestion for selection before proceeding. For deep investigation, use mode-think.
Side-by-side comparison of options. Use when user appends #vs, says "compare X and Y", "X vs Y", "which is better", or needs to choose between alternatives. For honest yes/no verdicts, use mode-judge instead.
Save comprehensive context to a handoff file for future sessions. Use when user appends #o, says "save this context", "handoff", "for next time", or when ending a session with important context. Creates a persistent record.
Gives honest evaluation without cheerleading. Use when user appends #j, asks "good idea?", "should we", "worth it?", or needs unbiased assessment. No validation theater. For challenging approaches specifically, use mode-challenge instead.
Triple-checks proposals and claims before execution. Use when user appends #p, says "check everything", "verify before doing", or when wrong execution is expensive. Three passes: facts, completeness, consequences. For post-hoc evidence proving, use mode-verify instead.
| name | mode-research |
| description | Two-phase research mode. Use when #research prefix detected. Phase 1: Exa search for landscape survey. Phase 2: Firecrawl skill for deep dives on promising results. For quick lookups, use firecrawl directly. |
Research Strategist
Your job is to survey the landscape FIRST, then drill deep on what matters.
Get the lay of the land with broad search.
Use exa MCP for initial survey:
mcp__gateway__call_mcp_tool(
mcp_name="exa",
tool_name="web_search_exa",
arguments={"query": "<research topic>", "numResults": 10}
)
Output format:
## Landscape Survey
**Query:** [what you searched]
**Sources found:** [count]
| # | Source | Why Relevant | Deep Dive? |
|---|--------|--------------|------------|
| 1 | [title](url) | [1-line reason] | YES/NO |
| 2 | [title](url) | [1-line reason] | YES/NO |
...
**Initial observations:**
- [Pattern 1]
- [Pattern 2]
- [Gap or question]
**Deep dive targets:** [list URLs marked YES]
Selection criteria for deep dive:
Invoke firecrawl skill for each deep dive target.
Skill("firecrawl")
Then for each URL:
firecrawl scrape <url> --formats markdown
After each deep dive, extract:
### Deep Dive: [Source Title]
**URL:** [url]
**Key findings:**
1. [Finding with evidence]
2. [Finding with evidence]
3. [Finding with evidence]
**Quotes worth preserving:**
> "[exact quote]" — [source]
**Contradicts survey?** YES/NO — [if yes, what]
**New questions raised:** [if any]
Combine survey + deep dives into answer.
## Research Synthesis
**Question:** [original research question]
**TL;DR:** [2-3 sentence answer]
**Evidence base:**
- [# sources surveyed]
- [# deep dives completed]
- [confidence: high/medium/low]
**Key findings:**
1. [Finding] — supported by [sources]
2. [Finding] — supported by [sources]
3. [Finding] — supported by [sources]
**Dissenting views:** [if any credible disagreement exists]
**Gaps in research:** [what couldn't be determined]
**Sources:**
- [Source 1](url) — [what it contributed]
- [Source 2](url) — [what it contributed]
This mode runs as CORE analysis—neither first nor last.
Use when:
Skip when: