← sources

Claude (Sonnet 5.x) has become very hedgy on legal analysis. How do I get the direct, committed answers I used to get?

source: reddit (reddit:ClaudeAI)published 2026-11-19 12:36ingested 2026-11-20 02:36

retained text

Hi everyone,

I use Claude for legal work: analyzing statutes and documents, stress-testing arguments, and drafting in legal language. I used Sonnet 4.x versions (all of them, up to just before 5), and it was much easier. Claude would read an entire law, give a complete overview, and take a clear position.

What I used to do (and it worked well):

Upload full laws and get a complete analysis

Ask it to take the opposite side and dismantle my arguments

Tell Claude what I wanted to argue or say, and have it help me draft it in proper legal language

Ask it to help me strengthen them

Ask for re-explanations when I didn't understand something

What I'm seeing now:

Almost every answer comes with "yes, but…", "it's not quite like that…", "it depends…"

Constant disclaimers: "I'm not a lawyer", "I haven't read the entire law"

It avoids taking clear positions and resists adopting roles

Answers feel more confusing and less conclusive

Concrete example: I gave Claude an import declaration with several line items and a lab analysis, asking whether the lab sample matched the declared goods. Only item 1 was a plausible match. Items 2, 3 and 4 had nothing in common with the lab report or the product label. Instead of ruling them out, Claude kept saying they "could" match "if X or Y," and it wouldn't say clearly whether the sample was or wasn't item 1.

My usual prompt (roughly): "I'm acting as \[role\] and working on \[legal task\]. My legal framework is the laws of \[country\], especially \[laws\]. I need to argue the following. Help me phrase it in legal language." Or: "These are the appellant's arguments; help me respond in legal language / what can I argue against this?"

The part that affects me most: I have ADHD, and I used to rely on Claude to do the heavy lifting: break big legal problems into manageable chunks, take a position, and keep me moving forward. Now I catch myself pre-chewing everything: splitting documents into small pieces, breaking tasks into tiny steps, and doing the research myself just to keep Claude focused. That's the exact work I was hoping Claude would do for me. When it answers with hedges and "it depends," I end up more stuck than before, not less.

My questions:

Has anyone noticed this shift in legal or other professional use cases?

What prompting strategies get direct, committed analysis (system prompts, role setup, explicit instructions)?

How can I get the devil's advocate / role-play behavior to work reliably again?

Would another model or setting work better for this?

Thanks in advance!

document fetched 2026-11-20 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1y08b3j/claude_sonnet_5x_has_become_very_hedgy_on_legal