convez.ai
Agente
az9713

deep-research-demo

az9713/deep-research-demo

Ver no GitHub
InicialPluginScore · de 0 a 100

A worked showcase of Claude Code's /deep-research multi-agent harness: two verification passes turn a persuasive talk into a fact-checked, valuation-aware investment view.

Instalação

$npx convez-ai az9713/deep-research-demo

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 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

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

Deep Research, Showcased — From a Persuasive Talk to a Verified Investment View

A worked, end-to-end example of the /deep-research multi-agent harness (Claude Code): how to turn a vague but compelling talk into a fact-checked, valuation-aware investment view — with the method's own limitations honestly labeled.

The source is Dan Dreyfus's critical-minerals / commodity-supercycle presentation (All-In). Two /deep-research runs verified the thesis, corrected the source, priced the opportunities, and — just as importantly — refused to certify the data it couldn't corroborate.

This repo is a methodology demonstration, NOT financial advice. See LICENSE. Do your own due diligence.

Last updated: 2026-06-10


Start here

If you want to…Read
Understand how it was built (the point of the repo)DEVELOPMENT_JOURNEY.md
Understand how /deep-research worksdeep_research_explained.md
See the investment outputinvestment_opportunity_analysis.md → investment_valuation_pass2.md
Reuse the prompt on your own topicinvestment_opportunity_prompt.md

The documents

#DocumentWhat it isRead it for
0DEVELOPMENT_JOURNEY.mdThe build process, phase by phaseHow the brief was engineered, how the workflow was chosen, what each run taught
1investment_opportunity_prompt.mdThe engineered research briefThe scenario → bottleneck → beneficiary → instrument design + adversarial kill-risks
2investment_opportunity_analysis.mdRun 1 — thesis & opportunitiesBottleneck map, claims audit, 5-tier matrix (~25 names), top-5 ranked, what-to-avoid
3investment_valuation_pass2.mdRun 2 — valuation & entryValuation matrix, per-name entry framework, best-entry ranking, priced-for-perfection callout
4deep_research_explained.mdMethod write-upPhases, agent math, both runs documented, when-to-use guide

Source material: Dan Dreyfus — "The Future of Critical Minerals" (All-In) (the talk's transcript was used locally during research but is not redistributed in this repo)


What /deep-research is (in one diagram)

YouTube talk → transcript  ──►  [1] engineered prompt
                                     │
                    /deep-research    ▼
        ┌──────────── Run 1: "Is the thesis true?" ───────────┐
        │  5 phases · 109 agents · 27 sources · 123 claims     │
        │  → 25/25 verified  →  [2] analysis (opportunities)   │
        └──────────────────────────────┬──────────────────────┘
                                        ▼
        ┌──────────── Run 2: "Is the price right?" ───────────┐
        │  5 phases · 109 agents · 27 sources · 123 claims     │
        │  → 22/25 (3 refuted, 5 names uncovered)              │
        │  → [3] valuation (entry levels)                      │
        └──────────────────────────────┬──────────────────────┘
                                        ▼
                          [4] deep_research_explained

Each run = Scope → Search → Fetch → Verify → Synthesize. The agent budget breaks down as 1 + 5 + 27 + 75 + 1 = 109, with ~69% of agents (75 of 109) spent trying to disprove the findings before any reach the report.


Headline findings

Run 1 — thesis (25/25 claims verified):

  • ✅ Copper, rare-earth-processing, grid-equipment, and uranium bottlenecks are real and primary-sourced.
  • ⚠️ Two of Dreyfus's headline numbers don't verify (the "50,000 t/GW" copper figure; the silver "3 years left").
  • 🔁 "Silver = solar" is backwards — PV silver demand is falling (−19%); the deficit is investment-led.
  • 🚨 Highest-value catch: China's rare-earth controls are SUSPENDED until ~Nov 2026 — a fact the talk never mentioned.

Run 2 — valuation (the margin-of-safety check):

  • The best-thesis names are the worst-value entries. PWR (~95x P/E), GEV (~70x EV/EBITDA), ETN (~60% above 10-yr avg) are priced for perfection; FCX (~8x) and FSLR offer real margin of safety.
  • The engine refused to certify 3 conflicting price claims and left 5 of 10 names unranked — failing loudly instead of fabricating.

Best risk-adjusted entries (verified names): FCX → FSLR → ETN → GEV → PWR.


The method lesson

/deep-research is a truth filter — only as good as how stably "true" the data is.

Question typeExampleResult
Durable facts (use it)forecasts, laws, filings, fundamentalsRun 1: 25/25 verified
Live values (don't)today's price, current P/ERun 2: 3 refuted, 5 names uncovered

Use it when being wrong is expensive and the truth is spread across many sources that disagree. For live single-source numbers, look them up directly and feed them in.


Repository layout

.
├── README.md                          ← you are here
├── DEVELOPMENT_JOURNEY.md             ← how it was built (phase by phase)
├── deep_research_explained.md         ← the method, demystified
├── investment_opportunity_prompt.md   ← the engineered brief
├── investment_opportunity_analysis.md ← Run 1: thesis & opportunities
├── investment_valuation_pass2.md      ← Run 2: valuation & entry
├── LICENSE                            ← MIT (original work in this repo)
└── .gitignore

The source talk's transcript, raw page captures, and the session transcript are git-ignored (third-party / scratch / personal). The source is the linked YouTube video.


Reproduce it yourself

  1. Provide your own source(s) — a transcript, PDF, article, or URL.
  2. Adapt investment_opportunity_prompt.md to your domain.
  3. In Claude Code, run /deep-research pointed at the brief (pin your universe/horizon/constraints).
  4. Chain a second pass for whatever Run 1 leaves open (here: valuation).
  5. Trust the "verified" label because the engine refuses to certify what it can't corroborate.

See DEVELOPMENT_JOURNEY.md for the full reusable pattern.


License & sources

Original work (prompt, analyses, write-ups, docs) is MIT-licensed — see LICENSE. The underlying source is a third-party talk — Dan Dreyfus, "The Future of Critical Minerals" (All-In) — whose transcript is not redistributed in this repository. Nothing here is financial advice.

Ver o original no GitHub

Bom para

Relacionados