【Project】How We Build the Deep Research Skill

TL;DRThis is the project's own deep research skill design, fully disclosed. Core choices: only Groundlane MCP for web tools, strict source-quality grading (A/B/C/D), research hands off to post skill for publishing. Not the most powerful, but the best fit for us.

#deep-research #skill #groundlane #project #architecture

Series: Deep Research 前沿 (13 / 15)

Commercial Landscape: OpenAI, Perplexity, Gemini, Claude, Grok 【Ecosystem】How the Community Builds Deep Research Skills

🌏 中文版

The previous 12 articles covered other people's research. This one covers ours.

The project's deep-research skill has been iterated since March 2026. This article tears it open: design philosophy, architectural choices, tool boundaries, and why.

Core Design Philosophy

Tool Boundary: Strictly Groundlane Only

This is the most distinctive design choice in the project skill:

All public web research uses only Groundlane MCP (web_search / web_fetch / web_extract). No Tavily, Exa, Firecrawl, Jina, or Linkup.

Why?

Control: Groundlane is a self-maintained MCP server—we control providers, rate-limiting, and authorization
Consistency: All research goes through the same pipeline, behavior is predictable
Cost: Self-hosted provider keys, no per-tool fees
Auditable: All requests through a single entry point, easy to debug

Groundlane itself:

Author: vincentxuu
Type: Vendor-neutral remote MCP server
Providers: Tavily, Exa, Parallel, Browserbase, Brave, Firecrawl, SerpApi, Linkup, Serper, You.com
Deployment: Cloudflare Workers + Containers

Output Goal: Research Note → Post Publishing

Not all research gets published. The project skill distinguishes:

Output	Storage	Purpose
Research note	`.research/<YYYY-MM-DD>-<slug>.md`	Working backup, not version-controlled
Published article	`src/content/posts/<category>/`	Final article

Flow: deep-research → research note → post skill → bilingual article

This means the research phase doesn't pollute version control—only confirmed-to-publish articles go in.

The 7-Step Workflow

0. Multi-case research (if topic is a series/comparison)
1. Decompose research sub-questions
2. Gather (≥2 sources per sub-question)
3. Catalog reading completeness
4. Cross-verify
5. Extract structure
6. Produce research note
7. Hand off (to post skill for publishing)

Key Discipline

Decompose before searching—wrong questions are more expensive than wrong answers
Bring full text of primary sources—not just query results (queries return snippets)
Cross-table only shows "what sources said"—inference is labeled separately
Numbers with conditions—effect sizes without control conditions are unusable
List conflicting facts—don't pick sides

Source Quality Grading

General Source Tiers

Tier	Definition	Examples
A — Official Primary	Published by creator/maintainer	Official docs, release notes, GitHub README
B — First-hand Author	Author's informal publications	Author's X/Mastodon, personal blog, talks
C — High-quality Secondary	Third-party with independent verification	HN top threads,知名 blog with testing
D — Low-quality Secondary	No independent verification	Medium reprints, SEO blogs, AI summary sites

Academic Paper Quality

Hard indicators: Conference tier (NeurIPS/ICLR/ACL), citation count, institutional prestige Soft indicators: Real benchmarks run, reproducibility, cited by later work, survey's taxonomic insight

Paper Quality Tiers

Tier A: Tier-1 conference accepted, >200 citations with open-source code,知名 AI lab with experimental data
Tier B: Tier-2 conference, >100 citations,知名 institutions
Tier C: arXiv preprint, <50 citations

Companion Skill Ecosystem

Skill	Role
`deep-research`	Core research workflow
`research-selection`	Multi-case selection (population → coverage matrix → bias)
`post`	Publish after research (bilingual)
`post-review`	Pre-publishing self-review
`post-verify`	Fact-layer verification
`series-curriculum-design`	Series planning

This ecosystem forms a pipeline: research → selection → publishing → review → verification.

Design Trade-off Reflections

Why No Adversarial Verification?

External skills (jamoeight v2, tolmachevmaxim) almost all have adversarial verification (Optimist/Pessimist/Fact-Checker). We don't.

Reasons:

Adversarial verification needs multiple models, increasing cost
Cross-verification (≥2 independent sources) is sufficient in most cases
Adversarial verification suits "controversial claims" but deep research is more about "synthesis tasks"

Why No Persistent Knowledge Base?

External skills (hyperresearch) have SQLite-indexed persistent research knowledge. We don't—each session is independent.