convez.ai
SkillBusca & PesquisaAvançado
harrrshall

Non-Consensus Research

harrrshall/non-consensus-research

Pesquisa multi-agente que caça achados não-consensuais e verifica se o mercado já precificou

Ver no GitHub
InicialPluginScore · de 0 a 100

Uma skill multi-agente que transforma Claude Code ou OpenAI Codex em um aparato de pesquisa autônomo capaz de identificar descobertas genuinamente não-consensuais em qualquer domínio temático, com verificação independente de precificação de mercado. O pipeline executa em ciclos, sintetiza achados cross-setoriais e elimina automaticamente qualquer dado já refletido no consenso de mercado.

Instalação

$npx convez-ai harrrshall/non-consensus-research

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Identificar oportunidades de investimento em setores emergentes (biotech, climate tech, AI hardware)
  • Síntese cross-setorial de tendências com análise de intersecções ignoradas
  • Due diligence contínua com verificação automatizada contra consenso de analistas sell-side

Veredito

Ferramenta sofisticada e bem-engenhada para pesquisa de alpha genuíno, com rigor na verificação de claims contra mercado. Exige domínio técnico para setup e está orientada a pesquisadores, analistas e investidores que precisam de sinal acima do ruído de consenso. A execução cíclica e limpeza automática de dados 'já precificados' são features raramente vistas.

README

A tradução para o português deste README ainda está na fila.

non-consensus-research

A reusable skill for Claude Code and OpenAI Codex CLI that turns either agent into a self-running, multi-agent research apparatus. It hunts findings that are technically real but still absent from mainstream analyst consensus, with an independent verification agent that ruthlessly purges anything the market already prices in.

What it produces

For any topic domain you point it at (AI hardware, biotech, climate tech, defense supply chain — whatever you care about), the pipeline produces:

  • N sector deep-dives — full per-sector research with sources, paper analyses, and 11-section synthesis
  • conclusion.md — cross-sector findings, bottlenecks, confidence-rated conclusions
  • market_opportunities.md — thesis / evidence / risk / horizon per opportunity
  • future_trends.md — 12–24mo + 3–5yr projections with falsifiers
  • cross_sector_alpha.md — a C(N,2) pairwise matrix + ranked non-consensus deep dives, the highest-value output
  • opportunity.md — a curated action shortlist, signal only
  • verification_log.md — independent fact-check of every "not priced in" claim against current sell-side, specialist, vendor, and trade-press consensus, with dated evidence URLs
  • run_history.md — verdict per cycle (GENUINELY NEW FINDINGS / MINOR REFRESH / NO MATERIAL CHANGE)
  • A mobile-first viewer site at index.html (no build step, fetches markdown at runtime)

The pipeline re-runs on a schedule (default every 5 hours). After every cycle's verification agent finishes, all ALREADY-PRICED-IN items are cleanly deleted from the synthesis files. No breadcrumbs. The document you read contains only what survived skeptical review.

How it works in one paragraph

For each cycle, N parallel sector teams run a four-agent chain (Collector → Analyzer → Validator → Writer). After a hard synchronization barrier, a global synthesis agent runs a C(N,2) pairwise permutation matrix over the sector reports to surface intersections that no single sector contains. A separate verification agent then reads the synthesis with zero access to the author's reasoning and does fresh web searches against major analyst desks, specialist research, vendor disclosures, and trade press to independently classify every "non-consensus" claim. Whatever the verification agent flags as already-priced-in gets deleted from the synthesis files. The surviving signal lands in a curated opportunity.md action list.

Install

For Claude Code

Clone this repo into your Claude skills directory. Choose one of:

Global install (available in all projects):

git clone https://github.com/harrrshall/non-consensus-research ~/.claude/skills/non-consensus-research

Project install (this repo only):

mkdir -p .claude/skills
git clone https://github.com/harrrshall/non-consensus-research .claude/skills/non-consensus-research

Then in Claude Code, just describe what you want:

"Set up a non-consensus research pipeline on biotech long-COVID treatments, running every 6 hours."

Claude will read SKILL.md, walk you through the setup (sectors, keywords, repo name, schedule), bootstrap the working repo, and trigger Run #1.

For OpenAI Codex CLI

Codex automatically reads AGENTS.md from the project root and from ~/.codex/AGENTS.md globally.

Project install:

git clone https://github.com/harrrshall/non-consensus-research /tmp/non-consensus-research
cp /tmp/non-consensus-research/AGENTS.md ./AGENTS.md
cp -r /tmp/non-consensus-research/orchestration ./orchestration
cp -r /tmp/non-consensus-research/scripts ./scripts

Global install:

mkdir -p ~/.codex
git clone https://github.com/harrrshall/non-consensus-research /tmp/non-consensus-research
cp /tmp/non-consensus-research/AGENTS.md ~/.codex/AGENTS.md

Then in Codex CLI:

codex "Set up a non-consensus research pipeline for the topic <X>. Use AGENTS.md."

Codex picks up AGENTS.md automatically and follows the eight-step cycle.

Or just fork it

Fork the repo, edit the spec, point a scheduled agent (Anthropic Code Routines, GitHub Actions, an external cron + the API) at it, and let it run.

Repository structure

non-consensus-research/
├── README.md                            ← you are here
├── SKILL.md                             ← Claude Code skill (YAML frontmatter)
├── AGENTS.md                            ← OpenAI Codex equivalent
├── orchestration/
│   ├── PIPELINE.md                      ← the generic master spec
│   ├── CLAUDE_TEMPLATE.md               ← copy into target repo as CLAUDE.md (or AGENTS.md)
│   └── index.html                       ← mobile-first reader template
├── scripts/
│   └── linkify_sources.py               ← convert Source Index plain titles to Markdown links
└── LICENSE                              ← MIT

The eight-step cycle (quick reference)

Every recurring run executes all eight steps. Skipping any step makes the run incomplete.

StepAction
0Read run_history.md, determine Run #N
1Discover NEW sources per sector (append, never remove)
2Analyze new papers → papers/paper-NNN.md
3Re-validate with six criteria → append to validation_log.md
4Refresh sector research.md if new findings
5FULL REWRITE: cross_sector_alpha.md, future_trends.md, market_opportunities.md
6ADDITIVE: append new conclusions to conclusion.md with [Run #N — new] marker
7Independent Verification Agent — fresh web search, three-verdict classification
7.5Clean-delete every ALREADY-PRICED-IN from cross-sector & market files (no marks)
7.6Overwrite opportunity.md with verified + partial items only
8Append Run #N verdict to run_history.md
9Commit + push (site auto-updates)

Full text of every step lives in SKILL.md / AGENTS.md.

Customizing

The skill is topic-agnostic. To use it for a new domain, just answer the eight setup questions Claude/Codex asks:

  1. Topic. (e.g. "AI hardware", "longevity biotech", "rare-earth supply chain")
  2. Sectors. 8–12 that cleanly partition the topic.
  3. Keyword seeds per sector.
  4. Source priority tiers. Tier 1 (peer-reviewed), Tier 2 (preprints), Tier 3 (specialist analysis), Tier 4 (primary disclosures) — outlets per tier vary by domain.
  5. Recency window. Default: last 6 months sliding.
  6. Cadence. Default: cron 0 */5 * * * (every 5 hours).
  7. GitHub repo to publish to.
  8. Scheduler to fire the cycle (Code Routines, Codex Scheduled Tasks, GitHub Actions).

Worked example

A reference deployment of this pipeline applied to the AI / semiconductor hardware domain (10 sectors, ~230 paper analyses per cycle) is the test case the skill was built from. The pipeline runs every 5 hours, runs the verification agent, purges priced-in items, regenerates the action list, and publishes to a mobile-first site.

To inspect a working instance's structure, look at the file tree any new domain will produce:

your-research-repo/
├── CLAUDE.md (or AGENTS.md)             ← operational spec
├── PIPELINE.md                          ← topic + sectors spec
├── index.html                           ← the live site (mobile-first)
├── .nojekyll                            ← disable Jekyll on GitHub Pages
└── research/
    ├── conclusion.md
    ├── market_opportunities.md
    ├── future_trends.md
    ├── cross_sector_alpha.md
    ├── opportunity.md                   ← signal-only action list
    ├── verification_log.md              ← audit trail
    ├── run_history.md                   ← cycle verdicts
    └── <sector>/
        ├── research.md                  ← 11-section synthesis
        ├── sources.json                 ← all sources for this sector
        ├── validation_log.md
        └── papers/
            └── paper-NNN.md             ← per-paper analysis

Philosophy

In an era where building is cheap, finding something that is not priced in is elite.

Most LLM research workflows produce confident summaries of facts the market already absorbed. This one runs a separate verification pass against current sell-side, specialist, and vendor coverage, and deletes anything that's already public knowledge. The remaining signal is small but earned.

A verified non-consensus call isn't a thesis — it's a starting point. The skill does the discovery and the filter. You do the dig.

License

MIT. Fork it, change it, sell it, embed it. Just don't blame us if a non-consensus call turns out to have been priced in by an obscure Substack you missed.

Inspiration / prior art

Ver o original no GitHub

Bom para

Relacionados