Output quality has degraded over the past month: jargon, over-building, multiple rewrites needed for simple requests

Summary

Output quality from Claude Code has dropped noticeably over the past month. Responses use a lot of internal jargon, over-build simple requests, and need several rounds of correction to get something usable.

What happened

Session: https://claude.ai/code/session_01MJhqs296nvouoswxYqKtsU
Model: claude-fable-5, 2026-08-24

I asked for a damage calculator for a browser game (one HTML page). It took four full rewrites to get something a person could read:

1. First version: 20+ input fields, six-column tables, labels like "eff STR", "stored", "floor", "steps".
2. Asked for clearer language. Second version replaced those with "in-battle", "connects", "lands", "caps". Still not readable.
3. Loaded a plain-language writing skill and asked again. Third version still had every input visible, long explanatory paragraphs, and the actual answer was below the fold.
4. Only after I said "a human can't understand what is going on" did it open the page in a browser, look at it, and cut it down to: pick an enemy, do you win, what to train.

The model also kept explaining things with metaphors and abstract phrasing in chat until told to stop.

Expected

A simple request ("make a damage calculator", "make it easier to understand") should produce a simple result the first or second time, and the model should check its own output visually before claiming it's clearer.

Environment

• Claude Code CLI, macOS (Darwin 25.5.0)
• Model: claude-fable-5