Skip to content

【Ecosystem】How the Community Builds Deep Research Skills

Sep 19, 20261 min
TL;DR10+ community deep-research skills represent 10+ philosophies of 'how to do research.' From hyperresearch's persistent vault to jamoeight v2's Co-Scientist 6-agent, from adversarial verification to benchmark alignment. This article puts them all on one table.

🌏 中文版

The previous article covered the project's own skill. This one covers the community.

From late 2025 to early 2026, the community flooded with 10+ deep-research skills. They represent not one solution, but 10+ philosophies of "how to do research."

Community Landscape

By Core Design

CategoryRepresentativeCore Idea
Persistent KnowledgehyperresearchSQLite index, knowledge accumulates across tasks
Enterprise-gradehashbullaNATO Admiralty grading, CRAG grounding
SOTA Paper Implementationjamoeight v2Co-Scientist, AlphaEvolve, BrowseConf
Adversarial VerificationtolmachevmaximOptimist/Pessimist/Fact-Checker three agents
Scaledkaynquang13 specialized agents
Academic-focusedSilence-viewSTORM-inspired, citation chasing
Platform-agnosticramit-mitraskills.sh compatible, 30+ platforms
Universal AdapterBhllcoder115 runtime adapters
Data-drivenSocialpranker75 report blocks, 280+ stat sources
Engineer-orientedrobertnowell6-phase, source quality gates
Human-in-loopWeizhenaHuman-in-the-loop, OpenCode/Codex

Detailed Comparison

hyperresearch (jordan-gibbs)

  • Core: 16-step pipeline + persistent vault + MCP server + 16 agents
  • Unique: "Patch, don't regenerate"—research knowledge accumulates to SQLite
  • Advantage: Cross-task knowledge reuse
  • Cost: Complex architecture, high maintenance

jamoeight/claude-code-deep-research-v2

  • Core: Co-Scientist 6-agent + AlphaEvolve + BrowseConf + BATS
  • Unique: v1→v2 upgraded with novel hypothesis generation and evaluator-driven search
  • Data: +10.3pp on deep-research benchmarks
  • Advantage: Latest SOTA paper implementation
  • Cost: Depends on Claude Code ecosystem

hashbulla/deep-research

  • Core: 7-phase + NATO Admiralty 2×6 grading + CRAG grounding loop
  • Unique: Enterprise-grade grading system
  • Advantage: Structured evaluation framework
  • Cost: Heavy framework

tolmachevmaxim/deep-research-skill

  • Core: 3 agents (Optimist/Pessimist/Fact-Checker)
  • Unique: Adversarial verification—not self-verifying, but three roles debating
  • Advantage: Concise and effective, file-based state
  • Cost: Depends on model quality

Common Architecture Patterns

Across all skills, these are nearly universal:

Common PatternDescriptionWhy Nearly Universal
Multi-agent parallel research2-13 agents researching simultaneouslyCoverage breadth + speed
Source credibility scoringTier grading / 6 dimensions / NATO AdmiraltyFiltering low-quality sources
Adversarial verificationPessimist/Fact-Checker/Red-teamPreventing over-optimism
Triangulation3+ independent sourcesEnsuring factual accuracy
File-based intermediate stateCrash recoveryNot losing progress on long tasks
Multi-tier depth modesquick/standard/deep/ultradeepAdapting to different needs
Progressive DisclosureSKILL.md + references/Maintainability

Comparison with Project Skill

DimensionProject SkillExternal Skills
Tool BoundaryStrictly GroundlaneMostly Tavily/Exa/Brave/Serper
Output Goal.research/ → post skill publishingDirect report output
Academic QualityHas A/B/C/D gradingSome have (hashbulla, Silence-view)
Adversarial Verification✅ Nearly all have it
Persistent Knowledge✅ hyperresearch has it
Benchmark Alignmentjamoeight v2 has BrowseConf
Cross-platform❌ Agent-environment only✅ skills.sh / Multi-runtime

Designs Worth Learning

  1. hyperresearch's persistent vault—knowledge accumulates across tasks instead of starting from scratch each time
  2. tolmachevmaxim's three-agent adversarial—simple but effective
  3. Socialpranker's 75 blocks—breaking reports into manageable chunks
  4. robertnowell's source quality gates—filtering low-quality sources before research starts
  1. Skill explosion—10+ skills represent methodology fragmentation
  2. Platform dependency—most skills are tied to Claude Code or Codex
  3. Benchmark gaming—some skills optimize for benchmark scores, not real research quality
  4. Complexity creep—16 agents, 10 phases may be over-engineering

References