Before you read anything, try this first
You submitted a response to a technical prompt. One grading criterion reads:
Criterion: Response identifies the off-by-one error in the loop bound. Weight: Primary objective.
Your response found the bug, explained it clearly, and also suggested renaming two variables for readability.
Which one determines whether this criterion is met?
The first one. A criterion marked as a primary objective is graded on exactly what it says: did the response identify the off-by-one error, yes or no. The variable-renaming suggestion is not irrelevant to your response as a whole, but it doesn't move this criterion. It might satisfy a different criterion, if one exists for code style; it does nothing for this one.
Graders working from a rubric don't average your response into a single vibe. They check each criterion independently against what's in front of them.
Insight: A rubric criterion is a yes/no or graded check against one specific thing. Extra effort that doesn't address that specific thing doesn't help the criterion, even if it makes the response better overall.
Learn this before the interview rather than during it
Most people walk into a domain or technical interview stage assuming they’re being judged on general impression: is this a good answer, does this person seem competent. Some stages work that way. Increasingly, many don’t.
A structured rubric breaks “is this a good answer” into a list of specific, independently checkable criteria. Each one measures one thing. Your response is scored criterion by criterion, and the criteria you didn’t know existed are exactly the ones you’re most likely to miss, because nothing about the prompt tells you they’re being checked.
Once you know the anatomy of a rubric criterion, you can often infer what’s being measured from how the prompt is phrased, even when the rubric itself isn’t shown to you. That’s the actual skill this module teaches: not memorizing a specific rubric, but recognizing the shape of one.
The anatomy of a rubric criterion
A well-built grading rubric breaks each criterion into the same handful of parts. Not every rubric labels them explicitly, but every rubric that’s doing its job has something playing each role.
Criterion description. The specific, checkable thing being measured. Not “is this response good” but “does this response correctly identify X” or “does this response avoid claiming Y.” A criterion description should be narrow enough that two different graders would reach the same yes/no answer looking at the same response.
Source. What the criterion is checked against. This might be the original prompt’s explicit instructions, a reference answer, a style guide, or a policy document. If you don’t know what the source is, you can’t verify your response against it before submitting, which is exactly why strong candidates ask or infer the source when it isn’t stated.
Rationale. Why the criterion exists at all. A criterion banning a specific claim might exist because that claim is factually common but wrong; a criterion requiring a specific format might exist because a downstream system parses the output automatically. The rationale tells you how strictly the criterion is likely to be enforced and what a near-miss will still fail on.
Weight (primary versus not-primary). Not every criterion counts the same. A primary objective is usually a hard gate: missing it can fail the response outright regardless of how well everything else is handled. A not-primary criterion typically adjusts the score up or down without being disqualifying on its own. Reading which category a criterion falls into tells you where to spend your effort when you’re short on time.
Criterion type. Three common shapes show up repeatedly:
- Extraction: did the response correctly pull out or state a specific fact, value, or detail that exists in the source material.
- Reasoning: did the response draw a correct inference, comparison, or conclusion, rather than just restating a fact.
- Style: does the response meet a formatting, tone, or presentation requirement, independent of whether the underlying content is correct.
Knowing the type tells you what kind of failure to watch for. An extraction criterion fails on a missed or wrong detail. A reasoning criterion fails on a broken logical step even if every fact cited is accurate. A style criterion fails regardless of content quality if the required format isn’t met.
Dependencies. Some criteria only apply if an earlier one is met. A criterion checking the quality of a proposed fix is meaningless if the earlier criterion checking whether the bug was correctly identified failed. Rubrics that specify dependencies are telling you the order failures cascade in, and where to focus first: an earlier failed dependency can make several later criteria automatically fail with it.
Try It: break down a criterion
You’re asked to review an AI-written summary of a technical article and rate it. One grading criterion reads:
The summary must state the specific benchmark score reported in the article (94.2% accuracy) without rounding or approximating it. Primary objective. Checked against the source article.
Identify the criterion description, the source, the weight, and the criterion type.
See answer
Criterion description: the summary states the exact benchmark score, 94.2% accuracy, with no rounding.
Source: the original source article, which is the reference the exact number is checked against.
Weight: primary objective, meaning a summary that rounds to “94%” or “around 94 percent” fails this criterion outright, even if every other part of the summary is excellent.
Criterion type: extraction. The task is pulling one specific, verifiable fact from source material correctly, not reasoning about what the fact implies or how it’s formatted.
The detail worth noticing: “without rounding or approximating” is doing real work here. A summary that says “roughly 94%” reads as correct to a human skimming it, but it fails a rubric written this precisely. Extraction criteria are frequently this exact and this unforgiving, because their entire point is verifying a specific fact was preserved, not just gestured at.
Reading a rubric you can’t see
Domain and technical interview stages don’t always hand you the written rubric. Often you only see the prompt. The skill is inferring the shape of the rubric from how the prompt is worded.
Specific verbs and qualifiers point to extraction criteria. “State the exact value,” “cite the specific line,” “name the function” are asking for something precise and checkable. Vague or approximate language where the prompt wanted precision is a common way to fail a criterion you never saw.
Words like “explain why” or “justify” point to reasoning criteria. These are checking the logical step, not just a fact. A response that states the right conclusion but skips the reasoning that gets there often fails a reasoning criterion even when the final answer is correct.
Explicit formatting instructions point to style criteria, and they’re usually primary. “Respond in valid JSON,” “keep it under 100 words,” “use exactly this structure” read like minor housekeeping. Treat them as gates. A technically excellent answer in the wrong format frequently scores worse than a mediocre one in the right format, because a style criterion phrased this specifically is usually a hard requirement, not a preference.
A multi-part prompt usually means multi-part grading, with dependencies. If a prompt asks you to identify a bug and then propose a fix, assume there are two criteria, and assume the fix criterion depends on the identification criterion. Get the first part wrong and the second part may not even be evaluated on its own merits.
Try It: infer the hidden rubric
The prompt you’re given: “Review the function below. Identify any bug that would cause incorrect output for negative input values, explain why it occurs, and rewrite the function to fix it. Respond with the corrected code only, no commentary.”
Without seeing the actual rubric, list the criteria you’d expect it to contain, and mark which ones are likely primary objectives.
See answer
A rubric behind this prompt likely contains something close to:
- Bug identification (extraction, likely primary). Does the response correctly identify that the function fails on negative input specifically, not a different unrelated bug.
- Explanation of cause (reasoning, likely dependent on #1). Does the response correctly explain the mechanism, not just state that a bug exists. This criterion probably only gets evaluated if #1 passes; if the wrong bug was identified, there’s no correct explanation to check.
- Corrected code is correct (reasoning/extraction hybrid, likely primary). Does the rewritten function handle negative input correctly without introducing a new bug.
- Output format: code only, no commentary (style, likely primary). The prompt states this explicitly and specifically, which is the pattern for a hard gate. A correct fix wrapped in an explanatory paragraph, when the prompt said “code only,” risks failing this criterion regardless of how correct the fix is.
The instructive part is criterion 4. It’s easy to treat “no commentary” as a stylistic suggestion and add a short explanation anyway, because explaining your reasoning feels like good practice. When a prompt states a format constraint this specifically, treat it as checked, not as a suggestion you’re free to improve on.
What this changes about how you answer
Once you’re reading for rubric shape instead of just answering the question, three habits follow.
Answer the narrowest version of the question first, then add. If a criterion is checking one specific fact or one specific bug, make sure that exact thing is stated clearly and early in your response, not buried inside a longer explanation where a grader (human or automated) might miss it.
Treat explicit formatting instructions as gates, not preferences. “Respond in exactly this format” is not a stylistic nudge. It’s very likely a primary, pass/fail criterion sitting on top of everything else you did well.
Don’t pad a correct answer with unrequested extras, and don’t assume extras help. Going back to the opening example: identifying a bug the prompt didn’t ask about, or suggesting an improvement nobody requested, doesn’t add points to a criterion that isn’t checking for it. It can also cost you time you needed for a criterion that is being checked.
Rubric grading flashcards
- Reading a hidden rubric: precise verbs signal extraction, “explain why” signals reasoning, explicit formatting instructions signal a style criterion that’s usually a hard gate.
- The habit that changes your score most: answer the narrow, specific thing a criterion is checking clearly and early, and treat stated format instructions as pass/fail, not as suggestions.
The specific thing being checked by this criterion.
What the response is checked against.
Why the criterion exists.
Whether the criterion is primary or secondary to the overall score.
Extraction, reasoning, or style, which shapes how a criterion should be read and graded.
What has to pass first before this criterion even applies.
