AGENT SKILL

optimized-research

Research routing for Shuttle apps: decompose questions, assign fast/medium/deep depth, route to Parallel/Tavily MCP or better primary sources, cite every claim. Load before any domain skill that performs external research.

License 许可
MIT
所属仓库
Domount/shuttle
最近更新
2026-11-12
Stars
★ 0
Forks
0

01 前言 Front Matter

name
optimized-research
description
Research routing for Shuttle apps: decompose questions, assign fast/medium/deep depth, route to Parallel/Tavily MCP or better primary sources, cite every claim. Load before any domain skill that performs external research.

03 SKILL.md 正文

Optimized Research

Project skill: skills/optimized-research/. Do not store or look for this under ~/.cursor/skills-cursor/ — that path is Cursor-managed and not durable.

Use Cursor MCP only — no CLI installs, no curl to MCP endpoints.

Prerequisites

Confirm these MCP servers are enabled (names may vary per workspace):

Server Tools Role
Parallel Search MCP web_search, web_fetch Fast lookup, fact-check, official pages
Tavily tavily_search, tavily_research, tavily_extract Medium search, deep synthesis

Read tool schemas from the project's mcps/ folder before calling.

User override: If the user asks to use only Parallel, only Tavily, or a specific tool/method, skip the routing below and follow their instruction.

Beyond these two tools: Any other research method — direct fetch of primary sources, public APIs, local data/ files, other available MCPs — is allowed when it is clearly more valid, cheaper, or easier than the defaults. State the reasoning in one line when deviating.

Workflow

1. Clarify scope

  • What questions must be answered?
  • Any local files or ground-truth data to reconcile against?
  • Output format: If the user already specified (JSON schema, markdown report, website, etc.), use that. Otherwise ask once how they want results delivered before researching.

2. Decompose

Split the job into atomic sub-questions. For each, assign a depth:

Depth When Tool
Fast Single fact, current number, "does X match Y?", one official stat Parallel web_search; web_fetch if excerpts are thin and URL is known
Medium A few related facts, light comparison, recent news Parallel web_search (batch queries in one call) or Tavily tavily_search
Deep Multi-domain synthesis, narrative report, strategies, broad explainers Tavily tavily_research (pro for broad topics)

Run independent legs in parallel where possible.

3. Execute

  • Prefer Tier 1 sources when verifying facts: official agencies, primary documentation, authoritative data publishers.
  • Cross-check numbers against user-supplied or parsed local data when present.
  • Cite every factual claim with a URL from tool results.
  • Do not invent statistics.

4. Synthesize

Merge sub-results into the user's requested output format. Note disagreements between sources or tools when they matter.

Routing heuristics

Parallel wins when: speed matters, you need a specific official page/table, fact-checking a few figures, or fetching a known URL.

Tavily wins when: one prompt should produce a cohesive long-form answer, multi-angle comparison, or actionable narrative across domains.

Default for mixed jobs: Parallel for verification legs, Tavily for synthesis legs — don't run Tavily deep research for a question Parallel can answer in one search.

What not to do

  • Do not install CLIs or use bash wrappers when MCP tools are available.
  • Do not use one deep research call for the entire job if most sub-questions are simple lookups.
  • Do not assume output structure — ask if unspecified.

Downstream skills

Domain-specific scope lives in your project's skills/<name>/ files. Load this skill first; then the scoped skill for the task (e.g. skills/research-task/, skills/contacts-research/).

Record depth choices per leg in runMeta on agent-written artifacts.

04 同仓库更多 Skills