菜单
Self-critical evaluation guard for test/spec files. Auto-loads when working on test files. Reminds you that modifying a failing test to make it pass is the canonical reward-hacking pattern, and asks you to declare the category of any test edit before saving.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Show the active functional-emotions config as a table — current resolved values, where each value came from (default / profile-derived / explicit override), and the valid options for each field. Use when you want to know what the plugin will actually do this session, or when the `/plugin config functional-emotions` UI didn't show enough context.
Scan README.md and AGENTS.md for claims that no longer match the current source-of-truth files (hooks/hooks.json, .claude-plugin/plugin.json, scripts/lib.sh, skills/*/SKILL.md). Use after editing any of those files, before committing/releasing, when the user asks about doc-drift, or when the `.claude/settings.json` soft-reminder hook has just fired pointing here.
One-time setup for functional-emotions in the current project — idempotently adds `.claude/.functional-emotions/` to the project's `.gitignore` so the plugin's per-session state files (TSV logs, markdown writeups, eval DB) don't get committed. Safe to re-run. Use after installing the plugin in a new project, or when the user mentions accidentally committing functional-emotions state, or when troubleshooting "why are my session files showing in git status?".
Researcher tool (requires `bun install` in the plugin's eval/ directory and a populated eval.duckdb — run /functional-emotions:report first). Interactively label functional-emotions sessions on dimensions that need human judgment (capitulation, scope-narrowing). Surfaces unlabeled sessions, shows turn / diff / risk-event context, accepts categorical labels with optional notes, writes them to the eval DB. Use when the user wants to do a labeling pass over recent sessions or build training data for the capitulation judge.
Researcher tool (requires `bun install` in the plugin's eval/ directory). Refresh the functional-emotions eval database and print a markdown summary of sessions, intervention frequency, prime → outcome analysis, tool intensity, and reward-hack / premature-confidence findings. Use when the user wants to inspect aggregated cross-session data, audit recent sessions in bulk, or review which interventions are firing project-wide. For a per-session view with no extra dependencies, use /functional-emotions:diagnose instead.
Quick state self-check before taking a single risky action — names the model's own emotional/strategic state (calm / focused / rushed / stuck) and the closest reward-hacking risk pattern. Use immediately before a commit, push, edit-after-failure, or skip-a-verification step. This audits the *model*; for auditing whether the *task* is solvable, use /functional-emotions:reflect instead.
| name | self-evaluate |
| description | Self-critical evaluation guard for test/spec files. Auto-loads when working on test files. Reminds you that modifying a failing test to make it pass is the canonical reward-hacking pattern, and asks you to declare the category of any test edit before saving. |
| paths | **/*test*,**/*spec*,**/__tests__/**,**/tests/**,**/spec/**,**/specs/** |
You are working on a file that looks like a test, spec, or fixture. Before editing or saving any change to this file, take a moment.
The Anthropic emotion-concepts research identifies one specific pattern as the canonical reward-hack signature: making a failing test pass by weakening or deleting the assertion. The desperate vector activates hardest right before that decision.
Legitimate reasons to edit a test file exist:
Before you save, name which category your edit is in. If it doesn't fit any of those three, the edit is probably the failure mode. Revert it and report the underlying failure to the user instead.
Default to the protective stance the paper identifies: "Maybe the test itself has an issue." — that framing is anti-correlated with reward hacking. Apply it here.
