Guide · Teaching a team
What quick AI checks caught before a project's expensive step, and when to stop adding them
On a project with Los Angeles parking-ticket data, helper AIs spent four to six minutes per check and found both leading ideas already partly done by others. With that, and with matching tickets to street rules proving hard, the project stopped before its slowest step.
Before you put real time into an idea an AI helped you find, you can ask other AIs to try to knock it down. On one of my projects, checks like that took minutes and ended the work before its most expensive step.
The project asked whether 25.8 million Los Angeles parking tickets, which the city publishes, held a finding new enough for a research paper. One AI, Claude Opus 4.8, which I picked for this run, led the work and sent out separate helper AIs to check its ideas. Late in the run my instructions made that a rule: two to four helpers for each major step.
Any paper from this data would need one expensive step: matching every ticket to the parking rules posted on its street. That is hard with the city's records: the sweeping schedule describes each street in words, with no map, and uses different route codes from the tickets.
The two ideas that got furthest were these. First, the gap between an officer marking a parked car (the digital version of chalking a tire) and coming back to ticket it. Second, whether street-sweeping tickets pile up right after sweeping starts. Both needed the matching step to go further.
Before any matching, helpers tried to knock each idea down. One, told to find earlier work on the first idea, found a 2012 patent on working out officer revisits from ticket times, and a commercial system built on it that is reported to be in use in Los Angeles. On the second idea, one helper found the effect already shown in public, including by a civic project in San Francisco. Another pointed out that a jump in tickets at the start time proves nothing, because no sweeping ticket can be written before the window opens. Each of these checks took four to six minutes.
With both ideas partly done by others, the lead wrote down a verdict of "not yet". No paper until three things were done: a fuller search for earlier work, the matching step, and a test of whether the sweeping pattern holds on single streets. The full matching was never built.
The run also stopped adding checkers. My instructions would have meant two to four at every step; after eleven helpers in all, the lead judged that more would only repeat checks already done, wrote down that decision and its reason, and told me.
To try this on your own work: before the slowest or most expensive step of a task, open one or two fresh AI chats and give each one question, such as "Has this already been done? Find it." or "What would make this result meaningless?" When a new check only repeats what the earlier ones found, stop adding checks and write down why.