Working notes

Most of what's on this site was built with AI doing the typing. Not assisted, running: agent pipelines that research, write code, check each other's output and file the results while I'm at school. The ticketing site, the planning packs, the trading system's test harness. One person, but only in the sense that one person decides. The output is a small team's.

The interesting problem stopped being how to prompt a model a while ago. It's deciding what to hand over. Code, research sweeps, first drafts of routine documents: delegated, always. The reflection in a report, a message to someone who trusts me, the call on whether a result is real: mine, always. I learned where that line sits the uncomfortable way. An agent once wrote up a backtest in language that made it sound like live results. It read fine. Confident, clean, wrong. Nothing caught it except the habit of tracing numbers back to their source, which by now is closer to a reflex.

The reason I care about getting this right is where I want to end up. Finance is turning into exactly this shape of work: the models produce the analysis, and the person is paid for knowing which outputs to trust and what to do when they disagree. Running these pipelines at seventeen feels like early practice for that job: choosing what to delegate, catching what's wrong, owning what ships. More useful practice, honestly, than most things with a certificate attached.

What I can't measure is what the delegation costs me. Some of the code that runs my own systems I have never read line by line. That would have horrified me two years ago. Whether that's the right price for the output, or a debt that comes due the day someone asks me to explain my own stack, I don't know yet. Keeping the question in the file.

No badminton since mid-September, none until mid-October, which is a strange sentence from someone who got into this school on a badminton DSA. I miss it more than I expected, in a physical way, like my body is annoyed with me. But the freeze was decided back in August, written into the project trackers next to everyone else's dates, so when it actually started nobody had to discuss it, including me.

That works for the big collisions. Rank points against everything else is not a hard question in October, because I answered it a month ago on a spreadsheet, and I'm not reopening it in the middle of a maths paper. The part I haven't solved is ordinary weeks. On a normal Tuesday, when a tutorial is due and the Elaris group chat is going, I usually answer the chat first, because a person waiting feels more urgent than homework even when the homework matters more. No rule for that yet. Closest thing is the Sunday night tracker reshuffle, where at least the trade-off gets decided looking at the whole week instead of at 11pm with my phone in my hand.

Handing things over is the other half. LeadSGP went to someone else even though I founded it, so I could do Project Hope properly. Chin Hao runs Elaris sessions now, off a run sheet and an operations manual, and this month is the test of whether those documents were any good. If the programme wobbles while I'm buried in revision, I built the structures wrong. Some nights I'm honestly not sure whether all this stepping back is discipline or a slower way of quitting things I said mattered. Ask me again in November.

Submitted the YFC final report this week. Fundraising line: $0, against a target of $3,100. I typed the number myself and sat there a bit before sending.

There was a softer way to put it, something like fundraising was deprioritised due to operational constraints, which is technically true and also not the story. The story is our first partner pulled out before a single session ran, we spent three weeks rebuilding the whole programme with Ren Ci, and fundraising never got picked back up. The report says that in plain words.

The zero stung less than I expected because everything else in the report holds up. We photographed every receipt on the day we spent the money, so the totals actually add up, and volunteer hours get logged the same day with a link to proof, then checked by someone who isn't me. That last rule exists partly because I know myself. If logging waited until the end of the month I would have rounded up somewhere.

AI is the other thing on the school's integrity slide, and it's a live question for me because I use it every day for code and first drafts of routine things. Numbers were the easy part: nothing gets submitted that can't be traced back to a receipt or a log. Prose took longer. I wrote the reflection section of that report myself, before letting anything else near it, because the assignment was the thinking, and outsourcing that would have left the report with no reflection in it. So the rule I've landed on: where the thinking is the point, the first draft is mine, and the tools only get to argue with it afterwards.

There's a slide from term one with Holistic Thinker on it, and I copied it into my notes without knowing what I meant by it. Best guess after a few months: the different things you do are supposed to connect somewhere. Mine mostly didn't, until PhysicsLab.

I kept getting SHM questions wrong. Could do the algebra, couldn't picture what the phase was doing. So I built a simulation site to force myself to understand, twelve labs now. There's a specific moment when you've coded the equations in and the pendulum on screen does exactly what the maths says, the period stretching as you drag the length slider, and something clicked there that three readings of the notes never managed. The slide also says Wonder, capital W. If I've felt that this year, it was then. After that I put the site up free for the cohort, because sitting on it would have felt weird, and now people I've never spoken to are revising off it.

That changed things. The site can't be wrong about the physics anymore, because it isn't only mine. Same with the Elaris run sheets, which have to work when Chin Hao is running a session and I'm not in the room. If I'm honest, I hold those two things to a higher standard than my own tutorial work, which is a bit backwards for a student. Do I chase understanding, or do I just like building things people use? My tutorial file, currently three weeks behind on annotations, has a view on that.

Two months in. In secondary school I kept most things in my head and it mostly worked. Here I missed a class admin deadline in week two because I trusted my memory, and I'm the chairperson, so it was mine to miss. The tutor said it was fine. It was not fine, it was the one thing I was supposed to catch.

I came in on DSA, so training a few times a week was always part of the deal. What I underestimated was everything stacking on top of it: lectures moving faster than I can digest, class chair admin, tuition, LeadSGP still running because I said yes to things before I knew what a JC week feels like. Most of February I was sleeping past 12.30 and feeling it at training the next day. I've seen my secondary school friends once since orientation. My mum pointed out I'd stopped eating breakfast, and she was right, which was annoying.

So, new systems. Everything goes into the calendar the moment it exists. Deadlines, trainings, even "reply Ms Yim" gets a slot. A tracker for class stuff so I stop rediscovering tasks at midnight. And on training nights I stop work at 11 whatever's unfinished, because a 1am tutorial done badly helps nobody and my knees complain the next day either way.

I haven't dropped anything since. But some Wednesday nights there is still nothing left in me to do the work with, calendar or not. Probably the real fix is saying yes to fewer things. I'm not ready to do that yet, or maybe I just don't want to.