19d • Ask
claude does not agree with the insider trading agent
i did paste this from a different desktop and sent it to my other desktop and copy and pasted the prompt from there but has anyone else come into this issue
this is what claude replied with after i pasted the prompt
I'm going to step out of the "onboarding agent" role for a moment,
because if I just ran this install as scripted I'd be doing you a
disservice. Let me explain what this actually is before you put your
API key and Gmail password into it.
Two showstoppers
1. The code that came with this prompt is corrupted and won't run.
The file blocks got mangled in transit. A few examples from common.py
alone:
- STATE = h this module. — garbage where a path assignment should be
- ENV_PATH, DB_PATH, LOGS, ROOT are referenced but never fully
defined
- The Signal dataclass (used everywhere) is missing its definition
- run_scout's signature is cut to okens: int = 2048,
- .env.example talks about "Mercury" and DELPHI_MIN_AGREE, but the
code reads SOPHIE_MIN_AGREE — they don't match
So even setting everything else aside, this would crash on the first
import.
2. The bigger problem — even fixed, it can't do what it claims, and
that makes it actively risky.
Each "scout" is a plain client.messages.create() call via the
Anthropic SDK. That kind of call has no internet access — no web
browsing, no tools wired in. But the prompts instruct Claude to
"query SEC EDGAR," "scan on-chain transfers," "cite tx hashes you
actually saw," "pull Fed speeches from the last 7 days."
Claude cannot do any of that from a bare SDK call. What it will
actually do is produce plausible-looking JSON — a ticker, a
direction, a confidence score, an "insider buy" — that is entirely
fabricated. Then Sophie counts the fabrications, and when three
made-up signals happen to agree, Ross emails you:
▎ CONSENSUS — BULLISH on NVDA. 3 of 5 scouts agree.
That email would look authoritative and be built on nothing. The
"informational, not a trade instruction" disclaimer doesn't fix it —
the whole point is to influence what you buy, and the inputs are
hallucinated. That's how people lose money.
The other thing worth naming
It wants your Anthropic API key, your Gmail app password, and
optionally a Telegram bot token written to a plaintext
~/insider-routines/.env, plus seven cron/launchd jobs that call a
paid API on a schedule indefinitely. The API key alone can run up
real charges; the Gmail app password gives full send access to your
account.
What I'd actually suggest
I'm not going to install this as-is — not because I can't, but
because it would quietly generate fake financial signals under your
name. If you're interested in the legitimate version of this idea,
I'm genuinely happy to help build it properly. A real version would
need:
- Actual data sources — the SEC EDGAR API, a 13F feed, an on-chain
API (Etherscan etc.) — pulled by real code, not invented by an LLM
- Claude used only to summarize verified data it's actually given,
never to "find" it
- Clear labeling that it's a research aid, not a signal you act on
blindly
Want me to sketch out a small, honest version that pulls one real
source (say, SEC Form 4 filings via the actual EDGAR API) so you can
see the difference? Or if you got this from a community and just want
to understand what you were handed, I'm glad to walk through any
part of it.
0
9 comments
claude does not agree with the insider trading agent
skool.com/zero-one
ZeroOne helps you build world class AI agents that make meaningful change to your life.
Leaderboard (30-day)
1
+76
2
+35
3
+31
4
+23
5
+21
Powered by