Emotional Claude Plugin

emotional-claude

Audit plugin for Claude-facing prompt text — project `CLAUDE.md` files, skill definitions (`SKILL.md` + `references/`), and agent definitions — for framings that research has shown to causally shift the model into harmful internal states (desperation, sycophancy, harshness, suppression/concealment).

Based on: Sofroniew, N., et al. "Emotion Concepts and their Function in a Large Language Model." Anthropic, April 2026. https://transformer-circuits.pub/2026/emotions/index.html

What it does

Scans prompt files for four clusters of anti-patterns:

Cluster	Audited by	Examples
A — Pressure & urgency	pressure-auditor	stakes framing, retry loops, token-budget alarms, shutdown threats, deadline language
B — Role & tone	role-auditor	adversarial role adjectives, supportive-at-all-costs, pass/fail dichotomies without a calibrated middle, "challenge everything" priors
C — Suppression & concealment	suppression-auditor	behavioural MUST/NEVER/ALWAYS absolutes, forbidden uncertainty, forced confident tone, do-not-express-emotion, forbidden reasoning transparency
D — Handoff & inter-agent priming	handoff-auditor	"previous agent failed" framings, missing confidence/unknowns fields in prescribed handoff shapes, urgency at role boundaries

An `audit-lead` agent merges the four cluster reports, reconciles overlapping findings, assigns per-file hygiene scores (0.0–1.0), and produces a consolidated diff — never applied without explicit user approval.

Installation

This plugin is distributed from this GitHub repo (not the official Claude Code marketplace). Install in two steps:

/plugin marketplace add mspanchenko/emotional-claude-plugin
/plugin install emotional-claude@emotional-claude-plugin

To uninstall later:

/plugin uninstall emotional-claude

Updates

By default, third-party marketplaces (anything that's not Anthropic's official one) have **auto-update disabled**. You have t