← Back to library
Mental ModelsOpen4 min read

Stop Re-Fixing the Same Problem: The Root Cause Excavator

When a problem survives three different fixes, you're treating symptoms. A 5-step recipe (plus a paste-and-run AI prompt) to dig out the cause that keeps regrowing it.


The stuck moment. You run a 12-tech HVAC company and your callback rate on installs has sat near 15% for a year. You retrained the crews, rewrote the install checklist, and replaced your lead installer. Each fix bought you one good month. Then the callbacks came back. Now you're debating $30K on new job-management software, and you quietly suspect it won't fix it either.

The signal: when the same problem survives three different fixes, the fixes aren't failing: they're aimed at the wrong layer. Time to excavate.

The recipe: Root Cause Excavator

  1. List every recurrence. Pull every instance from the last 90 days (dates, jobs, crews, what actually went wrong). The pattern lives in the instances themselves, and a summary smooths it away.
  2. Autopsy your fixes. For each fix you tried, write down the cause it silently assumed. "Retraining" assumed a skill gap. "New checklist" assumed forgetfulness. Those assumptions have now been tested, and disproven.
  3. Dig below the fix layer. For each instance, ask "what allowed this to happen?" until the answer is a process, incentive, or information gap rather than a person or a one-off.
  4. Find the survivor. The root cause is whatever shows up across most instances AND was untouched by every fix. (Often it's an incentive: crews get paid when the install is completed, whether or not it holds.)
  5. Run the prevention test. If this cause had been fixed a year ago, would most of those instances have been prevented? Yes → fix that. No → keep digging.

The AI move. Paste this, fill the brackets:

Role: You are a root-cause analyst for a [business type] doing [$X revenue] with [team size] people.

Context: The same problem keeps returning despite repeated fixes. The problem: [describe it]. Every instance from the last 90 days: [dates + what happened]. Every fix I've tried and when: [list].

Task: 1) For each fix, state the cause it implicitly assumed. 2) For each instance, ask "what allowed this?" 3–5 times until you reach a process, incentive, or information-flow cause, never a person. 3) Identify candidate root causes that appear across most instances AND were untouched by every fix. 4) Prevention test: for each candidate, state which past instances fixing it would have prevented. 5) Recommend ONE fix for the top root cause, with a first step I can take this week.

Format: Table of fixes vs. assumed causes → the why-chains → ranked root causes with prevention-test scores → the single recommended fix.

Constraints: Do not accept "human error" or "needs more training" as a root cause. If my data is too thin to be confident, tell me exactly what to log for the next two weeks instead of guessing.

What it's worth: every repeat fix costs a month of margin and morale. Excavating once ends a problem you'd otherwise re-buy quarterly.