If you want to try this on your own project, I am happy to spend 20 minutes with you on a video call. Friends and labmates first.
What you work on, one task you would like help with, and how you would know that task worked.
You type; I watch. First write down what you think the files do and one output that would worry you. Then ask the agent the same thing, and find one thing in its answer that is wrong. Then try your task with approval on, and give its check an empty or shuffled input to see whether it says no.
Write your five-line rules file, with the mistake you found as its first entry. Then what to try on your own. Last, one thing that helped and one that confused you; I use those to fix this page.
You leave with a rules file for your project, one task you have already tried, and one check you have watched say no. What you will copy: the rules file, the status file, and the prompts.
Claude Code installed and logged in. It needs a paid Claude plan or API access; the free plan does not include it. If installing fails, say so when you ask and we use the 20 minutes for that. A folder you understand: an analysis, a pipeline, or a set of notes. One real task.
Leave out data that cannot leave your institution: whatever the agent reads is sent to the company that runs the model. We look at code and file names, not protected data.
If we share a Slack workspace, message me there. Otherwise open an issue on the repository; issues are public, so keep the project line general. Three lines is enough:
I work on [project, one line]. I would like to try [one task]. I have Claude Code installed / need help installing it.
I reply when I can. No obligation either way, and nothing to prepare beyond the three things above.
Read the essay (about 15 minutes). Note which of the seven failure patterns you think you would have missed. We can start there.
Source ยท text CC BY 4.0