What kind of think-aloud narration can an interviewer actually score, and what is just noise?

## The claim being corrected Candidates are told to think aloud, hear it as *talk continuously*, and produce a stream of self-narration: "okay so now I'll declare a counter, and I set it to zero, and then I loop, and inside the loop, hmm, wait, no, actually…" This maximises words and minimises signal. Interviewers do not score talking; they score the decisions the talking exposes. Understanding that difference is what turns think-aloud from a performance into a tool. ## The test: could someone disagree with this sentence? A sentence is worth saying if a competent listener could push back on it. That filter sorts almost everything correctly. **Scoreable — someone could disagree:** - *Choices with reasons.* "I'm taking the version that pays a setup cost once, because the calendar is checked far more often than it is edited." A listener can say "actually edits dominate here" — so it carries information. - *Assumptions, stated before you lean on them.* "I'm assuming a request always has an end at or after its start; tell me if malformed entries are in scope." This is a boundary someone can move. - *Risks and deferrals.* "I'm going to ignore the case where the calendar is empty for now and handle it at the end — noting it so I don't forget." A listener can say "handle it now, it changes the shape". - *Changes of mind, with the trigger.* "I said linear scan; tracing that third case shows I re-do work, so I'm switching." This is the single highest-value thing you can say, because it proves the loop between evidence and plan is closed. - *Cost claims.* "This scan is proportional to the number of existing entries per request, so it grows with both." Anything about cost is inherently arguable and therefore informative. **Noise — nobody could disagree:** - *Reading the code you are writing.* The screen already says it. - *Narrating your emotional state.* "Hmm, this is tricky, okay, um, let me see." It transmits stress, not reasoning. - *Restating the problem for the third time.* Once, during clarification, is useful; repetition is stalling. - *Announcing mechanics.* "Now I'll add a variable." Nobody objects to a variable. ## Why over-narration actively hurts There are three separate costs, and candidates usually only fear the first. **Dilution.** If forty sentences are spoken and three of them carried decisions, the interviewer has to mine for them. Some will not, and the written feedback becomes "hard to follow". **Cognitive interference.** Speaking and reasoning draw on the same working memory. A candidate obligated to fill every silence thinks worse than one who says "let me think about this case for a moment" and then thinks. The permission to be quiet, when explicitly bracketed, is worth more than the words it replaces. **Commitment drag.** Continuous narration makes you announce half-formed ideas, and once announced, they feel like positions you must defend or visibly retract. Narrating at the decision level means you speak after the thought has a shape. ## Handling silence correctly The reason continuous talk feels obligatory is that unexplained silence is genuinely bad: from the outside, thinking and freezing look identical. The fix is not to eliminate silence but to label it. "Give me twenty seconds — I want to check whether the boundary case breaks this." Now the silence has a purpose and a duration, and the interviewer can wait comfortably. When you come back, close the loop: "it does break, so I need a different check here." Bracket, think, report. That pattern is worth more than any amount of filler. ## Density, roughly A useful calibration: over a working session you might say a couple of dozen sentences that matter — the clarifying questions, the fork narration, a handful of assumptions, two or three cost claims, one or two changes of mind, and the walkthrough of your test cases. Everything else is optional. If you can't remember a sentence thirty seconds after saying it, it probably was not one of the ones that counted. ## What this looks like at the extremes The silent candidate produces a working solution nobody can evaluate the reasoning behind, and the feedback reads "unclear how they arrived at it". The over-narrating candidate produces the same solution buried in commentary, and the feedback reads "talked a lot, hard to follow the plan". The decision-level candidate produces a small number of arguable statements that make the whole path reconstructible from notes. That last one is what the rubric line about communication is actually measuring, and it is a skill you can practise deliberately: solve with a recorder running, then listen back and count how many of your sentences someone could have disagreed with.