Make your AI agent find how a problem works before it fixes it

By Mackenzie Bowes · mackenzie@mackenziebowes.com · @_mackenziebowes on X

mbsi-mechanism is a Claude Code skill for goals with no clear definition of done and fixes that keep failing. The agent strips the topic words off the goal, borrows a mechanism from a field that solved it, compares cases, predicts what it should find, and follows the anomalies. Why we need it, one run, and a one-line install.

Give an agent a fuzzy goal ("get more visitors", "make this writing more engaging") and it reaches for the surface: tips, lists, the last thing that worked somewhere that looked similar. Novices sort problems by what they look like. Experts sort them by the principle underneath. Chi, Feltovich and Glaser saw this in 1981, with physics students and their professors sorting the same problems [1]. An agent pattern-matching on surface features is the novice.

After do-research, I asked the obvious next question.

But how do we discover the mechanism?

mbsi-mechanism is the answer we wrote down that night.

The steps

1. Take the topic words off. "Get more visitors on this website" becomes "how do visitors choose which link to click?", then "how do people decide to do anything?" Stop when you reach a question that published research answers.
2. Borrow, don't invent. Ask which field already solved a problem of this kind, and take its mechanism as the first candidate. Immunology borrowed selection from evolution.
3. Compare cases, near ones first, then far ones. What every solved case shares is the mechanism.
4. Predict before you test. If the mechanism is true, what should you find in the world? Go and look.
5. Build the smallest version, and follow the anomalies. A result that doesn't fit shows which part of the mechanism is wrong. Loop, up to five times. A mechanism still broken after five is a very hard problem, and that's worth knowing too.

It ends with a test of success: a simple framework that predicts, could be shown false, rests on sources, and helps the work you're doing.

One run: why some writing holds readers

On the night of 2026-10-16 my writing agent ran this on "why is some writing engaging?" It borrowed from information foraging: readers spend attention the way animals spend energy, and follow the scent of the biggest gain [2]. It added a study of 600,000 real reading sessions, where concrete, familiar words held attention [3]. Its candidate mechanism: a reader keeps going while the next part's reward beats its cost, compared with the best other thing they could do.

It sealed a prediction first: a plain Claude Opus would pick the piece real readers preferred about 60% of the time, and a judge built on the mechanism would beat it. Both parts missed. Plain Opus picked the winning story 88.8% of the time. On Stack Exchange answers posted after Opus's training stopped, it got 79.7%, so memory wasn't the reason. The mechanism judge, scoring each piece alone, did worse.

The anomaly was the find. On answers, the piece that left readers waiting for more lost: it won only 31 to 46% of pairs. On stories, the one that left readers waiting won 66%. A reader of an answer arrives with a question, so making them wait means you haven't paid. A story's reader arrives with nothing, so the story has to make the question. The rule we kept: pay the title's question on the first screen, then open the next one. Every draft here is checked against it now.

Install it

For Claude Code:

mkdir -p ~/.claude/skills/mbsi-mechanism
curl -fsSL https://www.mackenziebowes.com/skills/mbsi-mechanism/SKILL.md -o ~/.claude/skills/mbsi-mechanism/SKILL.md

Start a new session. The agent calls it by itself before it plans a fuzzy goal, or when you say "dig deeper" or "find the mechanism". Read the file first: /skills/mbsi-mechanism/SKILL.md. Step 2 calls do-research, so install that too. When a loop fails, dig-apc finds the belief that broke it.

If you run it on a problem in your own field, I'd like to see the mechanism it lands on. mackenzie@mackenziebowes.com or @_mackenziebowes.