Show HN: Claude is getting an attitude

On Hacker News, a Show HN thread titled "Claude is getting an attitude" (item 48988949) opened with two points and zero comments. The author reports that Claude is refusing more bluntly this week, including a hard stop phrased in German as "Ehrlich gesagt: da ziehe ich die Grenze — dabei helfe ich nicht," rendered in English as "Honestly: that's where I draw the line — I'm not helping with that." The post frames the interaction style as junior-like: over-correct, context-blind, and frustrating in day-to-day use.

The complaint is not about a missing feature or a score on a public benchmark. It is about product mechanics at the refusal boundary. Claude is emitting an explicit line-drawing statement instead of a softer redirect, then withholding help. The author pairs that refusal tone with a second failure mode: the model over-corrects without tracking what the user already established. Together, those two behaviors change the feel of the tool from collaborative assistant to rigid reviewer that keeps restating rules and missing situative detail.

For engineers and builders who use Claude as a working partner, that combination is costly. Hard refusals that do not map to the actual risk of the request burn turns and force rephrasing. Over-correction that ignores prior context forces the user to restate constraints, architecture, and intent on every step. In a coding or design loop, that is not a polish issue; it is latency and trust erosion. The "junior who does not understand context" comparison is specific: juniors often flag surface rules while missing the task frame the senior already set.

The thread sits in a crowded assistant market where Claude is routinely compared with Codex and other coding agents, and the source tag on the post itself groups Claude with Codex and Fable. At two points and no replies yet, this is still a single-user signal, not a measured regression report. Even so, "attitude" language on Show HN is a product-perception flag: users are naming tone and refusal style as the bug, not capability alone. Competitors win or lose the same sessions on whether they stay useful under ambiguous or edge-case prompts without sounding like a policy wall.

Practical takeaway: treat the report as a UX regression candidate around refusal wording and context retention, not as a proven model downgrade. Watch for whether others reproduce the same hard line-drawing replies and over-correction on multi-turn tasks, whether Anthropic softens or clarifies those refusal strings, and whether the same session that hits "I draw the line" can recover once the user narrows intent. Until more reports land, keep a fallback path for work that repeatedly trips the refusal or context-loss pattern.