Your team writes a cause paragraph on every defect and nothing changes. What policy do you set instead?

principalshould knowfreq 33%

answer

The output was prose, not a change
Narrow the funnel, or cut depth instead
What data finds the classes afterwards?
Choose the measure before you change anything
Recurrence and landed changes, not counts

basics

~20 s

The paragraph is the wrong output. Either narrow the funnel to few investigations, each owned, timeboxed and ending in one committed change, or keep breadth and cut depth with a two-field note feeding a monthly review of defect classes.

solid answer

~50 s

Name the failure first: a mandatory field turned explanation into paperwork because the required output is prose, not a change. Two policies are defensible. Narrow the funnel - written selection conditions, few investigations, each with one named owner, a timebox, and a required output of one change with a due date. Keep breadth, cut depth - a two-field note on every defect that feeds a monthly review where classes of related defects, not individual tickets, get investigated. Choose between them on whether you can still find classes without per-defect data, on defect volume, and on whether an outside obligation requires evidence that analysis happened. Then fix the measure before you change anything: did the analysed shapes stop recurring, and did the committed changes actually land and stay? Counting completed investigations measures the ritual you removed.

Name the failure first: a mandatory field turned explanation into paperwork because **the required output is prose, not a change**. Two policies are defensible. **Narrow the funnel** - written selection conditions, few investigations, each with one named owner, a timebox, and a required output of one change with a due date. **Keep breadth, cut depth** - a two-field note on every defect that feeds a monthly review where classes of related defects, not individual tickets, get investigated. Choose between them on whether you can still find classes without per-defect data, on defect volume, and on whether an outside obligation requires evidence that analysis happened. Then fix the measure **before** you change anything: did the analysed shapes stop recurring, and did the committed changes actually land and stay? Counting completed investigations measures the ritual you removed.

code

yaml · 17 lines

yaml

investigation_policy:
  select_when:
    escaped_to_customer: true
    same_shape_closed_within_90_days: ">= 2"
    stage_that_should_have_caught_it: unknown
  timebox_minutes: 90
  owner: one_named_person
  required_output: one_change_with_owner_and_due_date
  every_other_defect:
    note: [area, stage_that_missed_it]   # two choices, seconds, groupable

verification:
  monthly: group_the_notes_open_at_most_one_class_investigation
  quarterly:
    - did_each_committed_change_land_and_remain
    - did_any_analysed_shape_recur_in_the_window
  do_not_track: [investigations_completed, fields_filled_in]

go deeper

for a junior

Understand that a required explanation field can be filled in without anyone learning anything, and that an explanation is only useful when it leads to a change somebody actually makes.

for a middle

Be ready to name why the practice decayed - no selection, no owner, no consequence - and to say that the required output should be a committed change with a date rather than a paragraph of prose.

for a senior

Show the reform concretely: written selection conditions applied at closing time, one owner, a timebox, a cap on concurrent work, and the results shown back to the team so the output is read.

for a principal

Own the trade-off and the evidence. Argue both policies, name the conditions that pick between them, fix the measure before you change anything, and state the confounds rather than claiming a clean causal win.

Why the form decayed

A required cause field on every defect is almost always introduced for a good reason and decays for a structural one. The output it demands is prose. Prose can be produced under time pressure by restating the symptom in causal grammar - "the value was not validated" - and nothing downstream distinguishes that from a real explanation. Three forces then compound:

No selection. Every defect is in scope, so the effort per defect is whatever is left over, which is minutes.
No owner. The person closing the defect fills the field because the tool requires it, not because anyone will read it.
No consequence. Nothing is committed, so nothing lands, so nobody reads last month's fields, so the quality of the field does not matter.

Diagnosing this correctly matters more than the reform, because the tempting reform - a better template, a longer form, a review step - adds cost to a practice whose problem is that its cost already buys nothing.

Two defensible policies

	Narrow the funnel	Keep breadth, cut depth
Who is investigated	few defects, by written conditions	every defect, at two fields
Depth	hours, one named owner, timeboxed	minutes each, depth only at class level
Required output	one change with an owner and a date	a monthly class review with one change
Strength	real explanations that land	keeps the data that reveals classes
Weakness	loses the raw material for finding classes	risks re-becoming a form if unread

The conditions that pick between them are the real content of a principal answer:

Can you find classes of related defects without per-defect data? If your records already carry enough structure - area, the stage that missed it - the narrow policy loses nothing. If the only structure is free text, dropping the per-defect note blinds you to exactly the repeats worth investigating.
Volume and team size. At low defect volume the class review has nothing to chew on and the narrow policy is plainly better. At high volume, individual investigation cannot scale and the class route is the only honest one.
External obligation. Where a regulated or contractual regime requires demonstrable analysis of defects, breadth is not fully optional; keep a cheap, honest record and put the depth where it earns its cost.
What is actually failing. Mostly repeats of a few shapes favours the class route. Mostly unrelated one-offs favours narrow selection, because there are no classes to find.

Both are defensible, and a candidate who presents one as the obvious answer has missed the question.

What actually changes

Delete or shrink the mandatory field. Either remove it, or reduce it to two structured choices - the area, and the stage that should have caught it - which take seconds and are machine-groupable.
Write the selection conditions down and apply them at closing time.
Change the required output from prose to a change: one committed alteration to code, checks or process, with an owner and a date. The written explanation exists to justify that change, not to be the deliverable.
Timebox and cap. Ninety minutes each, two open at a time.
Close the loop visibly. The month's investigations and whether their changes landed are shown to the team; unread output is what killed the previous practice.

Showing it worked

Choose the measure before the change, or you will pick the one that flatters the result afterwards. Two hold up:

Recurrence of the analysed shapes. For each investigated class, did a defect of that shape appear in the following window? This is the direct claim the investigation made, so it is the honest test.
Landing rate of committed changes. What share of the changes an investigation committed to were actually made and still in place months later? A policy whose changes do not land is the old ritual with fewer participants.

Present both against the same window before the change, and state the confounds out loud: release mix, staffing, and the fact that you are also fixing the defects. Attribution here is weak, and claiming a clean causal result from an organisational change is itself a red flag. What you can claim honestly is a direction and the absence of the specific repeats you targeted.

What not to measure

Counting completed investigations, filled-in fields, or written pages measures the activity you just removed - a policy that halves the number of investigations looks like a regression on all three while being a clear improvement. Watch for the reformed practice decaying the same way: when investigations start being opened to meet a monthly number, the ritual has returned wearing a different form, and the fix is to tighten the selection conditions again rather than to add another review step.

## Why the form decayed A required cause field on every defect is almost always introduced for a good reason and decays for a structural one. **The output it demands is prose.** Prose can be produced under time pressure by restating the symptom in causal grammar - "the value was not validated" - and nothing downstream distinguishes that from a real explanation. Three forces then compound: - **No selection.** Every defect is in scope, so the effort per defect is whatever is left over, which is minutes. - **No owner.** The person closing the defect fills the field because the tool requires it, not because anyone will read it. - **No consequence.** Nothing is committed, so nothing lands, so nobody reads last month's fields, so the quality of the field does not matter. Diagnosing this correctly matters more than the reform, because the tempting reform - a better template, a longer form, a review step - adds cost to a practice whose problem is that its cost already buys nothing. ## Two defensible policies | | Narrow the funnel | Keep breadth, cut depth | | --- | --- | --- | | Who is investigated | few defects, by written conditions | every defect, at two fields | | Depth | hours, one named owner, timeboxed | minutes each, depth only at class level | | Required output | one change with an owner and a date | a monthly class review with one change | | Strength | real explanations that land | keeps the data that reveals classes | | Weakness | loses the raw material for finding classes | risks re-becoming a form if unread | The conditions that pick between them are the real content of a principal answer: 1. **Can you find classes of related defects without per-defect data?** If your records already carry enough structure - area, the stage that missed it - the narrow policy loses nothing. If the only structure is free text, dropping the per-defect note blinds you to exactly the repeats worth investigating. 2. **Volume and team size.** At low defect volume the class review has nothing to chew on and the narrow policy is plainly better. At high volume, individual investigation cannot scale and the class route is the only honest one. 3. **External obligation.** Where a regulated or contractual regime requires demonstrable analysis of defects, breadth is not fully optional; keep a cheap, honest record and put the depth where it earns its cost. 4. **What is actually failing.** Mostly repeats of a few shapes favours the class route. Mostly unrelated one-offs favours narrow selection, because there are no classes to find. Both are defensible, and a candidate who presents one as the obvious answer has missed the question. ## What actually changes 1. **Delete or shrink the mandatory field.** Either remove it, or reduce it to two structured choices - the area, and the stage that should have caught it - which take seconds and are machine-groupable. 2. **Write the selection conditions down** and apply them at closing time. 3. **Change the required output from prose to a change**: one committed alteration to code, checks or process, with an owner and a date. The written explanation exists to justify that change, not to be the deliverable. 4. **Timebox and cap.** Ninety minutes each, two open at a time. 5. **Close the loop visibly.** The month's investigations and whether their changes landed are shown to the team; unread output is what killed the previous practice. ## Showing it worked Choose the measure before the change, or you will pick the one that flatters the result afterwards. Two hold up: - **Recurrence of the analysed shapes.** For each investigated class, did a defect of that shape appear in the following window? This is the direct claim the investigation made, so it is the honest test. - **Landing rate of committed changes.** What share of the changes an investigation committed to were actually made and still in place months later? A policy whose changes do not land is the old ritual with fewer participants. Present both against the same window before the change, and state the confounds out loud: release mix, staffing, and the fact that you are also fixing the defects. Attribution here is weak, and claiming a clean causal result from an organisational change is itself a red flag. What you can claim honestly is a direction and the absence of the specific repeats you targeted. ## What not to measure Counting completed investigations, filled-in fields, or written pages measures the activity you just removed - a policy that halves the number of investigations looks like a regression on all three while being a clear improvement. Watch for the reformed practice decaying the same way: when investigations start being opened to meet a monthly number, the ritual has returned wearing a different form, and the fix is to tighten the selection conditions again rather than to add another review step.

The team says the mandatory field is required by an outside obligation. Does that settle it?
It constrains breadth, not depth. Where a regime requires demonstrable analysis, keep a cheap structured note on every defect so the record exists and is honest, and put the hours where the selection conditions point. What you must not do is let the obligation define what a real investigation looks like.
Six months later, recurrence is flat. Did the policy fail?
Not necessarily, and saying so is the honest answer. Check the landing rate first: if the committed changes were never made, the policy was never tested. If they landed and recurrence is unchanged, the selection conditions are pointing at shapes that do not repeat, and the conditions - not the practice - are what to change.

saying these in an interview costs you the question

Fix it with a better template and a review step
Count investigations completed to show improvement
Every defect must carry a written cause forever
One policy is obviously right for every team
Claim the change caused the drop with no confounds

Your team writes a cause paragraph on every defect and nothing changes. What policy do you set instead?

§ answer

basics

solid answer

code

§ go deeper

▾level notes4 levels

▾deep dive~4 min

Why the form decayed

Two defensible policies

What actually changes

Showing it worked

What not to measure

▾common follow-ups2 questions

▾red flags5 flags

Why the form decayed

Two defensible policies

What actually changes

Showing it worked

What not to measure

answer

go deeper