"Not your keys, not your coins" is a genuine, technically accurate slogan, not just a meme. A digital asset held in self-custody exists as an entry on a public ledger controlled by whoever holds the private key (typically generated from and recoverable via a seed phrase), no intermediary can freeze, seize, or lose it on your behalf, but you also have no recourse if you lose the key yourself, and no fraud department to call. A digital asset held with a custodial provider (an exchange, a broker) is instead a claim against that provider's own books, convenient, but only as safe as that provider's solvency, internal controls, and honesty about what it actually holds.
What did the FTX collapse teach about custody?
The primary-sourced case study for why that distinction is not academic is FTX. FTX Trading Ltd. filed for Chapter 11 bankruptcy protection in the District of Delaware on 11 November 2022, after reporting emerged that customer assets deposited on the exchange had been improperly commingled with and lent to Alameda Research, a proprietary trading firm under common ownership, a direct violation of the basic custodial promise that customer funds are segregated and available on demand. Sam Bankman-Fried, FTX's founder, was convicted on federal fraud charges in November 2023 and sentenced in March 2024. The lesson for this module is narrow and mechanical, not a moral one: an exchange's own claimed proof-of-reserves, a cryptographic demonstration that it holds a certain quantity of assets, proves exactly that and nothing more; it does not prove the absence of matching or exceeding liabilities, and it does not prove those reserves were not simply borrowed for the day the snapshot was taken. A custodian can pass a proof-of-reserves check and still be insolvent. Ask for evidence of the liabilities side, not just the assets side, before treating any custodial arrangement as verified, and where the amount matters, prefer self-custody or a regulated, audited (not merely attested) qualified custodian.
Advertisement
{
"skill": "Hype-filtering a new project's pitch before it enters any analysis, portfolio note, or client conversation",
"use": "Have an AI tool draft a list of the specific, checkable claims embedded in a project's marketing materials (an APY figure, a 'backed by X' claim, an 'audited by Y' claim, a team-credential claim) -- then verify every single one against its actual primary source yourself, treating the AI's list as a checklist generator, never as a verdict.",
"tools": ["Chat assistants -- Claude, ChatGPT, to extract and itemize claims from a document", "Research/agentic tools -- a browsing-enabled assistant to locate the actual audit report or team registration (still verify what it finds)"],
"prompt": "Read this project's pitch/whitepaper/landing page. Extract every specific, checkable factual claim it makes (a yield number, a reserve or collateral claim, a named audit or auditor, a named team member or backer, a regulatory-status claim). For each one, tell me exactly what primary document I would need to open to verify it, and flag any claim that sounds checkable but is not (e.g. 'audited' with no named auditor or report).",
"verify": "This is the highest-hype-risk moment in the whole corpus: an AI tool asked 'is this legit' will often produce a plausible, reassuring-sounding answer built from the same marketing language it was asked to evaluate, because that language is exactly what a plausible answer about a real, legitimate project would also contain. Never accept an AI verdict on legitimacy. Open the named audit report yourself; check the named auditor actually exists and published it; check the reserve attestation against the issuer's own transparency page; check custody and registration claims against the actual regulator's public register (Guardrail 1-5 in full, especially #1 Trace -- to the primary document, never to the AI's summary of it).",
"diy": "You must already know the three stablecoin designs, the Howey elements, and what an attestation is versus an audit -- everything this section built -- to know which claims even need checking and what a real answer looks like. AI can extract a checklist fast; it cannot tell you a project is safe.",
"market": "IN"
}