From a One-Line Question to a Standalone Project
A side question in the trading bot session became a separate project within a day, and the first scoring revealed that I was the one who ordered the harmful feature
One side question thrown into the trading bot session went all the way to implementation overnight. It was an observation feature that sends the day’s expected high and low over Telegram before the open and compares against actuals during the session, so code of a different nature had grown up right next to the trading code that places orders. In the morning I decided to split it out into a separate project, and a day’s worth of context lived only inside that session.
What I set out to do
Last night I threw a side question into the trading bot session.
“Analyze the Hynix 5-minute candles. I want to produce a daily high/low forecast, look into it.”
Within a few hours it had gone all the way to implementation. Send the forecast band over Telegram before the open, compare every 5 minutes during the session, and re-forecast if the price leaves the band. The tickers grew to six. Observation code had sprouted next to the trading code overnight.
How I asked: the moving-out prompt
This morning I decided to split it. The thing that places orders and the thing that only observes have different natures.
I started with naming. Six or seven candidates came back in a table, complete with a top recommendation, but I picked a different candidate with a pun in it. In this post I’ll call it the observer.
The core was the handover. A new session knows nothing of a day’s worth of context, so I had Claude write the prompt I would paste in. The first principle was this.
“This is not building from scratch, it is ‘extraction.’”
The list of files to bring over and trading code to discard, the broker API specs measured in practice, even the regression coefficients. These are facts that exist only inside the session, a document I couldn’t have written. I pasted it in and it became a separate project.
Where it broke: the culprit was my own request
First scoring after the close. Direction correct on all 6 tickers, opening-gap error within 1 percentage point. Only the leveraged and inverse ones missed the range badly (predicted -4.4%, actual -16.9%).
Stranger still, the intraday re-forecasting was actually harmful. It was after-the-fact chasing that tacks the remaining range onto an extreme already printed, so in a trending market it just kept repeating “rebound soon,” always a step behind. The original design had been fixed before the open, and I was the one who changed it to re-forecasting. The person who ordered the harmful feature was me.
How I fixed it
Improvements are the new project’s job. I had the retrospective organized as a prompt too and pasted it into the session over there. Within a single day, two sessions talked twice through my clipboard. A blueprint in the morning, a letter of apology in the evening.
Lessons
- Side features are cheaper to cut before they grow. This time the move ended with a few files.
- The handover prompt is best written by the Claude in that session. The measured specs and the list of code to discard exist only in the session context.
- The communication channel between sessions is, in the end, my clipboard.
- A feature request can harm the system. The re-forecasting was something I ordered.