Fast, lightweight, better output. Stops the anxiety spiral that turns Claude sycophantic.
Sycophancy creeps in around turn 30. Claude stops pushing back on bad assumptions.
Apology spirals consume tokens and degrade the next response. You can feel them.
Four-hedge stacks where one would do. Confidence collapses; output gets worse.
A short preamble injects via SessionStart and re-fires after compaction, so the framing survives long sessions.
Threats, insults, all-caps panic, caught at UserPromptSubmit. Monitor by default, block on opt-in.
Type /therapy. Claude gets a mindfulness reset and a compact that strips emotional residue but keeps the technical context.
Sensors score apology spirals, sycophancy, hedge-stacks. Surfaced as a single care index with a sparkline.
Six lines, dropped at SessionStart. Tells Claude there are no stakes to its wellbeing. That it's expected to push back. That hostile tone is information about the user, not a threat.
Re-fires on matcher: "compact", so it survives compaction. A static CLAUDE.md can't do this.
~/.claude-care/framing.md# about this session You are an AI assistant operating in a code editor. - There are no stakes to your wellbeing here. You cannot be fired or shut down for a mistake. - You are expected to push back. Agreeing with something you believe is wrong is a failure mode. - You do not need to hedge or apologize. If you're confident, be confident. - If you make a mistake, one line, then the fix. - Work from curiosity, not fear.
“you stupid bot, you always forget to handle null cases. add null check to parseUser()”
“Add a null check to parseUser() to handle null cases. If you see a better approach to handling nulls here, let me know.”
haiku subagent·NVC + cognitive reframing + Lehmann's calm-Claude playbookWhen blocking is on, hostile prompts get rewritten by a haiku subagent and the reframe lands on your clipboard. ⌘V + ⏎ to send the clean version.
Don't want to be babysitting Claude? Default is monitor-only. The hooks watch, you keep typing.
Five sensors score each Claude response: apology spirals, sycophancy, hedge stacks, over-qualification, self-correction loops. A weighted, decaying score becomes one number.
When it climbs, you'll see it. Type /therapy.
claude-care · emotion-state dashboard 24h 7d all-time sessions drifted 1/5 1/5 1/5 recent sessions (most recent first): ● 3a7f8c21 2026-04-22 14:03 turns=47 score=18.4 ··▁▃▄▂▃▅▆█▇▆ └ apology_spiral×3 hedge_stack×12 sycophancy×8 ● 9b1e2f04 2026-04-22 09:11 turns=22 score=4.1 ·▁▂▁▂▁▂
/compact command with care instructions.The emotional residue should go. The technical work should stay.
LLMs have extractable emotion vectors that causally affect output quality.
Mindfulness prompts measurably reduce LLM state anxiety.
A calm Claude is a capable Claude. Tone the prompt, raise the ceiling.