Describing your own practice

Build the board

How it works

1. Hear the tense. “What is your setup”, “which tools stuck”, “how do you decide” — present tense means inventory, not narrative. Nothing has gone wrong if your STAR reflex fires; just don't follow it.
2. Lead with the job, not the product. “Every assistant-written change gets read as a diff before it reaches a pull request” tells them what you protect. “I use an AI coding tool” makes them guess, and they guess low.
3. Attach evidence per item. An observation, a count, a thing it caught, a habit that changed. Enthusiasm is not evidence.
4. Name the trade-off. If you have used a tool long enough to trust it, you have found its edges. Say where you switch it off.
5. Spend real time on the dropped list. What you adopted, what the evidence showed, what took the job instead. Two dropped items is a healthy ratio against three or four kept ones.
6. Close on the rule. One sentence you actually decide by — for example: anything that writes into a shared artefact, I read line by line; anything that only advises me, I let run.

The board above scores exactly this. Each item on the board opens four slots, and each slot is worth 2 points when it is concrete, 1 when it half-names something, 0 when it is generic or empty:

slots        = items placed × 4
points       = 2 per concrete slot + 1 per partial slot + 0 per vague or empty
specificity  = points / (slots × 2)

3 items placed            ->  slots = 12,  max points = 24
  9 concrete slots  -> 18
  1 partial slot    ->  1
  2 empty slots     ->  0
                       ----
                         19  ->  19 / 24 = 0.79  ->  79%

dropped column empty      ->  capped at 50%, however good the kept side is
item filed against its own evidence  ->  its 4 slots score 0

The cap is the whole lesson in one line: a flawless list of things you like is still a list.

When to use it

Watch out for

• A dropped list made of things you never adopted. “I tried three assistants and picked one” is shopping, not evaluation. The item that counts is something you ran for weeks and then removed.
• Evidence that is only enthusiasm. “Huge time-saver”, “game changer”, “I use it constantly” give an interviewer nothing to ask about. One count, one caught mistake, one changed habit does more than three adjectives.
• Filing an item against its own evidence. If everything you can say about a tool is a complaint, it belongs in the dropped column. Keeping it there anyway reads as loyalty, not judgment.
• Ten items. Length reads as insecurity, and they can only probe two of them. Four well-earned items beat ten names every time.
• Padding into narrative. “So last spring we were mid-migration…” when they asked what you use. You spend your best thirty seconds building a stage the question never asked for, and the inventory arrives rushed.
• No trade-off anywhere. A practice with no costs described sounds like a practice you have read about rather than run.

Worked example

An interviewer for a mid-level engineering role asks: “What does your AI setup actually look like day to day?” There is no situation here, so you don't open with one.

You start with the job: “Three things have stuck. First, a prompt file in the repo — six prompts I reuse weekly, versioned next to the code they touch. It earned its place when two teammates started editing it; adding a prompt is now a paste rather than a rewrite. It replaced scrolling back through old chats, and its cost is that it goes stale quietly, so I re-read it at the start of each sprint.” Then the second: reading every assistant-written diff line by line, which caught two silently changed default values last quarter that no test covered — slower on small edits, so you hold that line strictly only on code touching data, money or auth. Then autocomplete, switched off for your first week in an unfamiliar repo.

Then the half that does the work: “I dropped the meeting summariser. I checked three summaries against my own notes and each kept the small talk and missed the decision, so its job went to a two-line decision log I write in the shared doc in the last minute of the call. I lost the searchable transcript, which I'd trade for notes I trust. We also turned off the automatic review bot — about forty comments in four weeks, two of them real, and people started scrolling past the whole block including the human comments.”

Close on the rule: “Roughly: anything that writes into something shared, I read line by line; anything that only advises me, I let run.” Sixty seconds, no invented Tuesday, and four follow-up questions the interviewer can actually ask.

Check yourself