The three asks that began with "include the inventory" — how "total funds" became a true net worth view

2026-06-11 ／ Vol 14 ／ draft at the time of publishing

In the previous issue I wrote about building a business dashboard in a single day. This is the story of the day after. The subject isn't a new feature, though — it's how my own request deepened through three rounds of dialogue with the AI.

The short version: in the morning, all I asked for was "make each month's total funds visible, including inventory." By evening, I had six years of total assets, liabilities, and net worth on a balance-sheet basis, plus a detail page that drills down to which account holds what and which loan has what left on it. I didn't ask for that at the start. I didn't know it was something I could ask for.

The ask changed three times

Ask 1: "Make each month's total funds visible, including inventory." The dashboard I'd built the day before already showed cash over time. But to anyone running a physical-goods business, inventory is just money in a different shape — staring at cash alone doesn't tell you how much strength you actually have. One request later, I had a stacked chart of cash + inventory ("total funds") and a current-month card. So far, exactly what I expected.

Ask 2: "There are hand-entered records of past inventory in a spreadsheet I keep." The new chart only went back a few years, so I mentioned this. The timeline promptly extended back to 2020. For six years, at every stocktake, I had typed the month's inventory value into an Excel file by hand — and that habit suddenly paid off. Records I kept with no particular purpose in mind became the foundation of a 77-month trend chart. Apparently you keep records before you know what they're for.

Ask 3: "I want the true total — everything, including liabilities, receivables, prepayments." Staring at the total-funds chart made me greedy. Cash plus inventory still doesn't tell you what the business is really worth: there are loans, unpaid card balances, receivables. One more request, and the chart transformed into a balance-sheet view of total assets, liabilities, and net worth. Then I pushed once more — "more detail" — and got a drill-down page at the account and sub-account level: which bank account holds what, which loan has what remaining.

A table that checks its own math

There's one design choice in the detail data that I really like. Assets are stored as positive numbers, liabilities as negative. That way, summing every line item gives you net worth directly — the cross-check is built into the table itself. If the grand total in the matrix doesn't match the net worth computed by a separate route, some line item is broken.

The script that builds the data has consistency checks baked in too. For all 77 months, if the line items' asset total doesn't match total assets, or the liability total doesn't match liabilities, the script errors out on the spot. Better to stop loudly than to quietly publish a wrong number.

The validation failed twice — correctly

Those checks fired twice today. Both times, I was glad they did.

The first failure was a number mismatch. The script's known-value check stopped with "doesn't match the expected value." I went hunting for a code bug, but the cause wasn't code. That morning I had re-exported the latest data from the accounting software, and the entries I'd just posted had moved the numbers. In other words, the validation was telling me "the data changed." I updated the expected values and moved on. A thing I built as a bug detector turned out to work as a data-change detector too.

The second failure was an overpaid card. In one month, I had overpaid a certain credit card, flipping its balance — the liability went negative (which, sign-wise, looks like an asset). The validation logic originally classified by sign — "positive means asset, negative means liability" — and that one month broke the totals. The fix: classify by balance-sheet category, not by sign. Real data will always betray your naive assumptions somewhere.

One honest failure: the tool I left open overwrote my work

Let me be honest about this part. Midway through, I left Power BI Desktop open while editing the project files externally. The still-open Desktop saved the file with its stale contents, and part of the external edits got overwritten and lost. Redo time.

Since then, "close the tool before editing externally" is carved into the build log as an operating rule. Failures turning directly into rules is business as usual on this project.

The lesson: requirements can come last

The most interesting part, looking back: the morning version of me could not have articulated "I want to see true net worth." Only after seeing the chart with inventory did I want the history. Only after seeing six years of history did I want the net position including liabilities and receivables. Only after seeing that did I want to drill into the details. Every time I saw something working, the resolution of what I actually wanted went up.

In traditional system development this would be called a requirements failure — the requirements changed three times in one day. But when co-building with an AI, the cost of each step is small enough that you can defer requirements definition. Start with a rough ask, look, deepen. The "true net worth" page that exists tonight would probably not exist if I had tried to write a perfect spec first.

Ask roughly, look, restate. Three restatements weren't rework — they were the number of times the request grew up.