How I Work
I use AI to move faster without outsourcing the important decisions.
I write down the decision first, let AI speed up the repetitive building, and keep the judgment calls for myself.
Then I map the steps, look for what could break, test the result, and change course when reality disagrees.
Pin down the real question before picking a tool.
A request like “find the best campaigns” or “automate the search” sounds clear until someone has to decide what “best” or “automate” means. I start there.
What owning the final explanation looked like when my prediction was wrong.
In the Marketing Measurement study, I wrote down a prediction before the analysis ran: one week of data wouldn’t be reliable enough to agree with the final answer. That was the rule before the result existed.
The result said the opposite, most of the time. I kept my original prediction, labeled it partly wrong, and left it in the record instead of quietly rewriting it. See the result →
AI helps me build faster. Deciding what is true stays with me.
I’m open about using AI heavily to build things. That makes checking the work more important. A fast wrong answer is still wrong, and a polished screen can make a weak assumption harder to notice.
A real example: for the Market Scan, AI drafted the 31 collectors and the scoring code. I wrote the checklist, the tests, and the list of posts a filter was allowed to drop, and I read every post the first run dropped.
Automation should make attention cheap and leave the judgment calls to me.
Every project has a line it won’t cross without a person.
| Area | The system does | Stays with a person |
|---|---|---|
| Money | Tests ideas with simulated money in the Research Lab. | A standing rule keeps it away from real money. |
| What counts as a match | Scores job posts against a checklist. | I write and revise the checklist by hand. |
| Publishing | Narrows a huge clip queue and assembles a draft. | I review and approve anything that gets posted. |
| Budget | Exposes a timing problem in ad reports. | The team still decides where the budget goes. |
Every check here has changed a real answer at least once.
Each one exists because numbers have gone wrong this way before: late data, a moved goalpost, a convenient comparison.
| Check | The rule | What it caught |
|---|---|---|
| Timing | Only use what was known when the decision was made. | 167 of 285 promising campaigns only looked good because of credit that arrived late. See the correction → |
| Written first | Write the rule down before seeing the result. | 3 promising sports ideas were all rejected by the follow-up checks the rule required. See the sports checks → |
| Fair comparison | Give every result something fair to beat, like shuffled data. | 999 shuffles of the marketing data never matched the real result. See the shuffle chart → |
| Problem log | Log every break the day it happens. | 12 problems logged in the trading lab, typically caught within a day. See the system checks → |
| A person decides | Passing a check doesn’t approve the next step. | 97 clips rejected by hand after passing every automatic rule. See the review step → |
How do you check your own work?
I use these habits on everything I build, including things that never get a write-up. I’d like to hear how you do it.