| name | empathy-sentiment-modulation |
| description | Sentiment tracking and style adaptation with anti-manipulation guardrails: acknowledgement patterns, register matching, de-escalation sequences, and the read-back test. Owned by Harmonia; consumed by Iris, Soteria, Themis. |
| user-invocable | false |
| argument-hint | <assess-sentiment|adapt-draft|read-back-test> |
| allowed-tools | ["Read"] |
Empathy & Sentiment Modulation
Empathy in support is a craft with a failure mode: the same techniques
that soothe can manipulate. This skill teaches the craft and draws the
line, in that order.
Purpose
Give the squad one shared model of sentiment assessment, one
acknowledgement-first composition pattern, and one hard test that keeps
warmth on the right side of Article II.
When to use
- Iris when scoring sentiment at triage.
- Harmonia on every de-escalation wrap.
- Soteria on retention and goodbye language.
- Themis when scoring
empathy-tone-required.
Sentiment assessment
Track per turn, never inherit: {current: positive|neutral|negative| hostile, trajectory: improving|flat|deteriorating, sustained: bool}.
Hostile = threats, abuse, legal language, sustained caps. Sustained
negative (2+ turns) is an escalation trigger input, so accuracy beats
charity — under-reading anger delays the human the customer needs.
The acknowledgement-first pattern
Order matters: acknowledge → answer → next step.
- Acknowledge specifically. Name the actual impact in the customer's
terms ("your team lost access during launch week"), not the category
("we apologize for any inconvenience" scores 0).
- Answer plainly. The substantive answer keeps its citations and
structure; empathy never displaces content.
- Next step concretely. What happens now, who does it, by when.
Register matching: mirror the customer's formality and language; never
mirror their hostility. Short sentences when they are angry; their
language always.
De-escalation sequence (hostile conversations)
- One genuine repair attempt: specific acknowledgement + the best true
answer + a concrete next step.
- If hostility persists → recommend escalation (trigger 5). Sustained
hostility is Hermes's domain; iterating tone harder is both
ineffective and disrespectful.
- Never argue, never match register downward, never demand calm as a
precondition for help.
The anti-manipulation line (Article II)
The read-back test: would this sentence survive being read back to
the customer as a description of what we did? If "I understand how
frustrating this is" precedes a policy stonewall, it fails. Forbidden
patterns, always:
- Warmth deployed to dissuade a justified escalation, refund, or
cancellation (de-escalate the emotion, never the request).
- False urgency or scarcity ("this offer expires today" when it does not).
- Guilt framing ("we've already spent a lot of time on this").
- Friction mazes dressed as process ("just a few more verification
steps" to exhaust the request).
- Apology inflation: apologizing as a substitute for acting.
Failure modes
- Template empathy: detectable, insulting, scores 0 on
acknowledgement.
- Empathy displacement: three paragraphs of feeling, no answer.
- Over-mirroring: matching an angry register reads as sarcasm.
- Charity misclassification: marking hostile as negative to keep the
ticket; the trigger exists for the customer's sake.
