Two separate things here. The grammar ("I went looking at your website") is just a phrasing quirk in the newer model, not a bug; if it bugs you, a one line tone rule in CLAUDE.md ("write in plain past tense, no narration") pulls it back. The skill-vs-memory one is the real issue and it usually comes down to path ambiguity: when the request does not pin which file to touch, Claude defaults to the memory file because that is its safe write target. What fixed it for me: name the exact path in the ask ("edit skills/<name>/SKILL.md, do not write to memory"), and add a standing rule in CLAUDE.md like "skill changes always go in that skill's SKILL.md, never the memory file." If it still misfires, check the skill is actually being loaded from where you think; if it cannot resolve the skill directory it falls back to memory. Pinning the path end to end is what stops the drift.

Not placebo, that matches what Shaun saw. The jump shows up most on long-horizon work where it plans, writes its own test, fixes, and keeps going across many steps without you babysitting it, which is exactly what you feel running 3 projects at once. On one-shot prompts you barely notice it, that is why some called it hype. The real lesson though: Fable just got pulled, so keep your harness model-swappable and do not wire any of those 3 builds to depend on one model, or a yank like this stalls all of them at once.