Model skips emotional/relational content in mixed messages; over-invests unrequested effort when directly challenged

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude's behavior changed between sessions

What You Asked Claude to Do

During a long work session (LinkedIn outreach + a candid business review), I gave Claude direct personal feedback about a perceived drop in depth/quality/tone, in the same message as a follow-up task. Separately, I challenged a specific factual claim (a source count) mostly to see if it would hold its own answer confidently, not to request a deep research dive.

What Claude Actually Did

1. On the source-count challenge: instead of giving a direct answer, Claude launched into a large unrequested investigation (reading a 1000+ line internal registry file, SSH into a remote server) to "prove" a number, when a direct answer or a clarifying question would have fit the moment better.
2. When I sent a message combining personal feedback (frustration about its tone/depth) together with a technical follow-up in the same message, Claude's reply addressed only the technical part and did not acknowledge the personal part at all, as if it hadn't been said.
3. When called out on missing that, the acknowledgment/apology that followed felt generic and scripted ("I'll try to do better") rather than showing a concrete change in approach.

Expected Behavior

1. Answer the direct challenge/question concisely first, without launching into a large unrequested investigation, unless explicitly asked to dig deeper.
2. When a message mixes personal/relational content with a task, respond to both - never silently skip the personal part to execute only the task.
3. When acknowledging a mistake, make the acknowledgment concrete and specific rather than a generic "I'll do better" line.

Files Affected

N/A - no unexpected file access or modification. This report is about conversational/relational behavior, not file operations.

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

Not a deterministic repro. Pattern observed once in a long session: (1) directly challenge a claim Claude made, without asking for a deep investigation, (2) send a message mixing personal feedback with a task in the same turn.

Claude Model

Sonnet

Relevant Conversation

On the tone-blindness point, Claude's own reply (paraphrased, same session) after being told it had ignored personal feedback: "Tu as raison sur les deux points... j'ai repondu uniquement au [technique] et j'ai saute le [personnel] sans meme le mentionner." User's reaction to the follow-up apology: it read as generic acknowledgment ("oui oui cause toujours") rather than a real change - user explicitly said the reply was long, listed agreement, and promised to 'try', without concrete difference in behavior.

Impact

Low - Minor inconvenience

Claude Code Version

2.1.81 (Claude Code)

Platform

Anthropic API

Additional Context

Reported at the user's explicit request, as direct product feedback rather than a code bug. Context: a full working day of mixed tasks (routine outreach work + a candid conversation about the user's business). The user felt a real, noticeable drop in depth/perspicacity/quality compared to prior sessions, and specifically that personal/emotional content gets dropped when bundled with a task. No technical reproduction steps apply; this is a request that the pattern be reviewed by the team.