5.11 Poor model reply quality

Typical symptoms: Answers drift off-topic, repeat loops, misuse tools, leak raw <think>…, or bilingual noise (Chinese riddled with English).

Sanity checks before swapping models

1. Task difficulty — hard debugging belongs in Thinking mode.
2. Thread length — start a fresh session after long debates.
3. Skill load — /skills in AI Dock lists active skills—trim noise.
4. Only after the above consider upgrading the thinking backbone.

Checklist detail

1. Wrong modality — Fast mode for micro Q&A only; Thinking for planning/tracebacks. Peek the footer model chip; tweak in AI engine if mismatched.

2. Thinking tags bleed through — some vendors wrap private reasoning inside <think>:

   • Upgrade via Settings → App & updates
   • Toggle advanced model options that hide reasoning traces.

3. Mixed languages — AI engine → Advanced ⇒ set Reply language to Follow OS or Chinese. Persisting English chunks usually means you need a stronger base model instead of UI knobs alone.

4. Wrong tools/skills — /skills; uninstall unused baggage under Skill workshop → Installed.

5. Context rot — very long transcripts confuse routing; spinning up a blank session is fastest remediation.

Make it stick

• Invest in stronger thinking models for mission-critical workflows.
• Encode repeatable procedures into Skill workshop → Create so Moss stops guessing tooling.
• Keep Studio patched—providers ship compatibility fixes continuously.