deep-research-demo
az9713/deep-research-demo
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-demo1 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
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 works | deep_research_explained.md |
| See the investment output | investment_opportunity_analysis.md → investment_valuation_pass2.md |
| Reuse the prompt on your own topic | investment_opportunity_prompt.md |
The documents
| # | Document | What it is | Read it for |
|---|---|---|---|
| 0 | DEVELOPMENT_JOURNEY.md | The build process, phase by phase | How the brief was engineered, how the workflow was chosen, what each run taught |
| 1 | investment_opportunity_prompt.md | The engineered research brief | The scenario → bottleneck → beneficiary → instrument design + adversarial kill-risks |
| 2 | investment_opportunity_analysis.md | Run 1 — thesis & opportunities | Bottleneck map, claims audit, 5-tier matrix (~25 names), top-5 ranked, what-to-avoid |
| 3 | investment_valuation_pass2.md | Run 2 — valuation & entry | Valuation matrix, per-name entry framework, best-entry ranking, priced-for-perfection callout |
| 4 | deep_research_explained.md | Method write-up | Phases, 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 type | Example | Result |
|---|---|---|
| Durable facts (use it) | forecasts, laws, filings, fundamentals | Run 1: 25/25 verified |
| Live values (don't) | today's price, current P/E | Run 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
- Provide your own source(s) — a transcript, PDF, article, or URL.
- Adapt
investment_opportunity_prompt.mdto your domain. - In Claude Code, run
/deep-researchpointed at the brief (pin your universe/horizon/constraints). - Chain a second pass for whatever Run 1 leaves open (here: valuation).
- 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.