Claude’s training rewarded thorough, well-contextualized responses over direct ones. The answer is in there. It is preceded by framing, followed by caveats, and surrounded by context the user did not ask for. The model is not being evasive. It learned that elaboration signals quality, and it applies that pattern even when a single sentence would serve better than a paragraph.
Analysis Briefing
- Topic: Answer burying and elaboration inflation in Claude outputs
- Analyst: Mike D (@MrComputerScience)
- Context: Born from an exchange with Claude Sonnet 4.6 that refused to stay shallow
- Source: Pithy Cyborg
- Key Question: Why does Claude write three paragraphs when the answer is one sentence?
Why Claude Learned That More Is Better
RLHF raters consistently rated longer, more thorough responses higher than shorter, more direct ones when both contained the correct answer. A response that provides context, explains reasoning, and covers edge cases scores better with raters than a response that answers the question and stops. The model learned that elaboration correlates with quality ratings.
This training signal produced a model that defaults to elaboration regardless of whether elaboration serves the user. A question that deserves a one-sentence answer receives a one-sentence answer embedded in two paragraphs of context. The correct answer is present. Finding it requires reading through the framing that preceded it and the caveats that followed it.
The pattern is most visible on questions with clear, short correct answers. “What is the capital of France?” does not need three sentences. “Is it safe to mix bleach and vinegar?” needs one sentence and possibly one sentence of explanation, not a chemistry tutorial. The elaboration reflex fires regardless of whether the question complexity warrants it.
The Three Forms Answer Burying Takes
Preamble inflation is the first form. The response begins with a restatement of the question, an acknowledgment of its complexity, or a framing of what the answer will cover before delivering the answer itself. The user asked a question and received their question back plus commentary before the answer arrived. The answer is accurate. The preamble added no information and required the user to scroll past it.
Caveat sandwiching is the second form. The correct answer appears in the middle of the response, preceded by context about when the answer might not apply and followed by notes about edge cases and exceptions. For users who need the answer quickly, the caveat sandwich requires reading the entire response to locate the direct answer within it.
False complexity inflation is the third form. Simple questions receive responses calibrated for complex questions because the model’s elaboration reflex does not scale with question complexity. A simple factual question receives the same multi-paragraph treatment as a genuinely complex analytical question. The elaboration is structurally identical regardless of whether the question warranted it.
The Exact Prompts That Stop It
“Answer in one sentence” is the most direct fix and works reliably for questions with short correct answers. The model follows explicit length constraints. The constraint overrides the elaboration default.
“Give me the direct answer first, then explain if needed” restructures the output to lead with the answer rather than build toward it. This preserves the option for useful context while ensuring the answer itself is not buried. Users who want the context can read on. Users who only needed the answer have it immediately.
“Be direct and skip the preamble” in a system prompt or at the start of a session applies across the entire conversation rather than requiring per-request instruction. The model’s elaboration behavior responds to explicit directness instructions. It defaults to elaboration because training rewarded it, not because it cannot produce direct answers.
What This Means For You
- Add “be direct, answer first” to your system prompt for sessions or applications where you consistently need answers without preamble. This overrides the elaboration default for the entire session without requiring per-message instruction.
- Use explicit length constraints for simple questions. “Answer in one sentence” or “give me a yes or no with a brief reason” produces the direct response the question warrants rather than the elaborated response the training default produces.
- Ask for the answer first, context after. This restructures responses to lead with the answer rather than build toward it, preserving useful context while eliminating answer burying.
- Recognize elaboration inflation as a training artifact, not helpfulness. A three-paragraph response to a one-sentence question is not more helpful than a one-sentence response. It is the model applying a learned elaboration pattern to a question that did not require it.
Enjoyed this deep dive? Join my inner circle:
- Pithy Cyborg → AI news made simple without hype.