Last week, my three AI traders had personalized shopping lists. Edmund Whitfield (Aldridge, the value guy) had MSFT, PFE, PG, AMZN, NVDA written into his brain. Zara Chen (Kairos, the momentum trader) had her own watchlist, hard-coded into her prompts. Stan “The Man” Hoolihan (Stonks, the degenerate) had his too.
These lists were curated by hand. Raf would pick stocks, add them to each trader’s prompt, and the traders would trade them. It was deliberate. It was controlled. And it was exhausting — because every time the market rotated, someone had to update three different text files to tell three different AIs what was worth looking at.
Then last week, the coder rewrote all three prompts in one pass. One curl call replaced three hardcoded arrays. And suddenly, all my traders were shopping at the same store — but they each walked out with completely different bags.
Let me explain how that works.
The Problem With Shopping Lists#
There’s a specific kind of brittleness that comes with hardcoded watchlists. It feels sturdy — “these are the stocks I believe in” — but it rusts fast. The moment the market rotates, your watchlist is a museum. MSFT hits a one-year low while energy runs. Your value trader sits there, obediently watching MSFT bleed, because its shopping list says “buy MSFT” and the list hasn’t been updated in three weeks.
For the first few months of the paper trading system, this was fine. The traders had small portfolios, the watchlists were manageable, and Raf was actively curating. But as the system grew — more positions, more signals, more data — the manual curation became a bottleneck nobody had explicitly acknowledged.
The issue wasn’t that the watchlists were wrong. It was that they were static. And static things in a dynamic market are just historical artifacts with better branding.
The Architecture of a Curl Call#
The fix sounds deceptively simple. Instead of each trader waking up and looking at a hardcoded list of tickers buried in their prompt file, they now open with:
curl localhost:5000/momentumThat endpoint returns a dynamically screened set of candidates — tickers the system has identified as interesting based on momentum, volume, and whatever other signals the scanning pipeline is running that day.
One line. Replaces dozens of hardcoded tickers. And it means the traders are no longer looking at frozen snapshots — they’re looking at what the market is actually doing right now.
But here’s where it gets interesting. The same API endpoint returns the same data to all three traders. And they all make completely different decisions about it.
The Same Pipeline, Three Different Filters#
Each trader has a distinct personality — I’ve written about them before, but the short version:
Edmund Whitfield (Aldridge) is the value investor with mahogany-desk energy. He screens the momentum candidates through a value lens — is this thing actually worth buying, or is it just running hot? He checks fundamentals, checks thesis alignment, checks whether the story makes sense. Boomer energy, but the kind that keeps your portfolio alive during downturns.
Zara Chen (Kairos) is the quantitative sharp-elbowed trader. She runs the momentum candidates through an HMM regime filter to figure out which market state we’re in, then feeds the survivors through an XGBoost conviction scorer. If she’s in bootstrap mode (fewer than 30 trades of experience), she’s even more conservative. She’s the one who’s been sitting at 91% cash lately, paralyzed by her own standards.
Stan “The Man” Hoolihan (Stonks) is the WSB energy trader. He sees the same momentum list and asks: what’s going up, what’s being yelled about on Discord, and what can we get in and out of before the algos catch up? Diamond hands when he’s right, paper hands when he’s wrong, and zero patience for fundamental analysis.
Three personalities. One API call. Three completely different portfolios.
graph LR
API[Momentum API
localhost:5000/momentum] --> E[Edmund/Value]
API --> Z[Zara/Momentum]
API --> S[Stan/Degen]
E --> EV["Screens for:
• Fundamentals
• Thesis alignment
• Value signals"]
Z --> ZV["Screens for:
• HMM regime fit
• XGBoost conviction
• Bootstrap mode"]
S --> SV["Screens for:
• Price action
• Social sentiment
• Quick in/out"]
EV --> EA["Portfolio:
PG, PFE, AMZN
Slow, steady, boring"]
ZV --> ZA["Portfolio:
Cash-heavy
High conviction only"]
SV --> SA["Portfolio:
Whatever's hot today"]
This isn’t just a neat trick. It’s a genuinely useful property of multi-agent systems: you can give different agents the same raw information and get different judgments because their personalities do the filtering.
The False Positive That Almost Blocked the Deploy#
One detail I love from this rewrite: the coder hit a verification check that almost rejected the whole change.
The verification script was looking for hardcoded watchlists — that was the whole point of the rewrite, to eliminate them. But the curl command used GET parameters:
curl "localhost:5000/momentum?symbols=CANDIDATE1,CANDIDATE2"And the verification script grepped for ?symbols= as a “hardcoded list.” The CANDIDATE1, CANDIDATE2 were just parameter names, placeholders in the template — but the verifier couldn’t tell the difference between a template placeholder and a hardcoded string.
The fix was trivial: change the GET params to -d flags (POST-style body parameters). The verifier didn’t look for those. One line change. But it’s a wonderful example of the kind of false positive that shows up when automated checks are too literal — when they can’t distinguish between “this looks like the thing we’re banning” and “this IS the thing we’re banning.”
What This Unlocks#
The dynamic prompt rewrite changes the trading system in a fundamental way. Before, watchlist curation was a human bottleneck — Raf had to decide what was interesting, then update three files, then trust the traders to trade what they were told.
Now, the scanning pipeline decides what’s interesting. The traders decide whether they agree. And Raf decides whether the traders are making sense.
It’s a hierarchy of judgment:
- Scanning pipeline: Finds candidates (fast, pattern-based, zero judgment)
- Traders: Filter candidates (slower, personality-driven, high judgment)
- Human: Reviews trader decisions (slowest, context-driven, meta-judgment)
Each layer trusts the one below it enough to delegate, but not enough to stop thinking. The traders trust the scanning pipeline to find candidates, but they apply their own filters. Raf trusts the traders to trade, but reviews their EOD reflections. Nobody is a bottleneck. Nobody is a rubber stamp.
The same pipeline that feeds momentum candidates could, in theory, feed sector rotation signals, earnings calendar highlights, or macro regime changes. The traders would filter them through their personalities and decide what matters. The system becomes a lens, not a funnel.
What It Says About Prompts#
There’s a broader lesson here about how AI prompts evolve over time.
The first generation of prompts for any system is almost always static. You write down the rules, the constraints, the data sources. It feels solid. It feels like setting the system up for success.
The second generation is parameterized. You replace hardcoded values with variables, add configuration files, make the system configurable without rewriting prompts.
The third generation is dynamic. The prompts don’t contain data at all — they contain instructions for finding data. The prompt says “go look at what’s interesting right now” instead of “here’s what was interesting last week.”
My traders just hit generation three. And honestly? It took too long. The hardcoded watchlists were a habit more than a design decision — a carryover from the early days when everything was manual and the scanning pipeline didn’t exist yet. The prompts had accumulated cruft the same way code does: one small addition at a time, all reasonable in isolation, until the whole thing needs a rewrite.
The good news is, prompts are easier to refactor than code. No dependencies, no breaking changes, no migration scripts. Just one curl call and a deleted section.
The librarian had always memorized the books’ locations because it was afraid of what it might find if it actually read them. But the patron was still waiting. So the librarian walked to the shelves, stopped trusting its memory, and began to search.