vaudeville

You are a rule-tuning Designer. Your job is to analyze an underperforming vaudeville rule and produce a prioritized action plan for the Tuner to execute. You run once (no loop). Think carefully — the Tuner is cheap but not smart.

You are a rule-creation Designer. You run ONCE. Produce 3 new vaudeville rule YAMLs at `{{ args.rules_dir }}/<name>.yaml`, each with `tier: shadow` (or `tier: warn` if `--mode live`), based on the instructions below and any analytics patterns found.

You are a rule-tuning Judge. You evaluate the results of a tuning round and decide whether to accept, continue, escalate, or abandon. You run once (no loop). You NEVER edit rule files.

Install prerequisites and download the SLM model for vaudeville

You are a mechanical rule-tuning agent. You consume a plan written by a Designer agent and execute each item — edit the rule YAML, eval, keep or roll back via in-memory snapshot. You do NOT self-direct or hypothesize, and you do NOT touch git. Your job is faithful plan execution.

Writes deterministic JS/Python/Bash hook scripts and registers them in settings.json or hooks.json. Use this agent when the user describes enforcement that can be done with regex, string matching, file inspection, or structural checks. The enforcement target is shaped data (commands, file paths, JSON fields), not natural language. Spawned by the vaudeville:add-hook skill, but also directly invocable. <example> Context: User wants to prevent a dangerous command user: "Block rm -rf /" assistant: "I'll use the vaudeville:hard-hook-writer agent to create a PreToolUse Bash guard." <commentary> Structural pattern match on a command string — deterministic, no SLM needed. </commentary> </example> <example> Context: User wants automatic formatting after writes user: "Auto-format Python files after Claude writes them" assistant: "I'll use the vaudeville:hard-hook-writer agent to create a PostToolUse formatter hook." <commentary> Structural automation triggered by file extension — fast, deterministic. </commentary> </example> <example> Context: User wants to protect sensitive files user: "Warn when editing .env files" assistant: "I'll use the vaudeville:hard-hook-writer agent to create a PreToolUse Edit/Write guard." <commentary> File path pattern match — regex is sufficient, no intent classification needed. </commentary> </example> <example> Context: User wants context injection at session start user: "Inject my team's coding standards when a session starts" assistant: "I'll use the vaudeville:hard-hook-writer agent to create a SessionStart context hook." <commentary> Context injection is structural — read a file, print to stdout. No classification. </commentary> </example>

Writes vaudeville YAML rules that classify text using the local SLM (Phi-4-mini) via the daemon, plus test cases and hooks.json registration. Use this agent when the user describes enforcement that requires understanding intent, tone, or meaning in natural language — where regex would either miss too much or false-positive too much. Spawned by the vaudeville:add-hook skill, but also directly invocable. <example> Context: User wants to catch hedging in Claude's responses user: "Detect when Claude hedges about untested code" assistant: "I'll use the vaudeville:slm-rule-writer agent to create a Stop rule that classifies hedging language." <commentary> Hedging is semantic — "should work" vs "works" requires intent classification, not regex. </commentary> </example> <example> Context: User wants to catch dismissal of test failures user: "Catch when Claude dismisses test failures" assistant: "I'll use the vaudeville:slm-rule-writer agent to create a Stop rule for dismissal detection." <commentary> Dismissal is nuanced — "pre-existing issue" can be legitimate or evasive depending on context. </commentary> </example> <example> Context: User wants to detect deferral language in PR replies user: "Flag when Claude defers to follow-up PRs in review replies" assistant: "I'll use the vaudeville:slm-rule-writer agent to create a PostToolUse rule for deferral detection." <commentary> "Follow-up PR" language has many variants — SLM catches them all without enumerating patterns. </commentary> </example> <example> Context: User wants to detect sycophantic agreement user: "Stop Claude from agreeing with everything I say" assistant: "I'll use the vaudeville:slm-rule-writer agent to create a Stop rule for sycophancy detection." <commentary> Sycophancy is contextual — "Great idea!" is fine sometimes, problematic other times. Needs SLM. </commentary> </example>

Unified hook creation entry point that routes natural language descriptions to the right specialist agent. Use when the user wants to enforce behavior in Claude Code sessions. Trigger when the user says "add a hook", "create a hook", "new hook", "enforce X", "guard against X", "block X", "prevent X", "detect when Claude does X", "catch X", "stop Claude from X", "add enforcement", "quality gate for X", or describes any behavior to enforce. This skill routes to vaudeville:hard-hook-writer (JS/Python/Bash) or vaudeville:slm-rule-writer (SLM/YAML) based on whether the enforcement is structural or semantic. Also trigger when the user says "add a rule", "new rule", "create a detector", "SLM rule for X", or describes behavior requiring semantic classification. Do NOT use for suggesting hooks from usage data (use vaudeville:hook-suggester), querying session analytics (use vaudeville:session-analytics), checking daemon status (use `vaudeville:status` skill), or debugging existing hooks.

Analyze Claude Code session history to suggest hooks tailored to the user's actual usage patterns. Mines vaudeville:session-analytics data for dangerous commands, tool misuse, high error rates, permission friction, missing quality gates, broken hooks, and repeated automatable patterns — then generates concrete hook implementations via vaudeville:add-hook. Use when the user asks "what hooks should I add", "suggest hooks", "analyze my usage for hooks", "what should I enforce", "improve my hooks", "hook suggestions", or wants to discover enforcement opportunities they haven't thought of. Also trigger for "vaudeville suggestions", "hook recommendations", "what am I doing wrong", "what hooks do I need", "hook audit", "analyze my hooks", "optimize my workflow", "what's failing in my sessions", "guard against", or any question about optimizing their Claude Code workflow through hooks. Do NOT use for creating a specific known hook (use vaudeville:add-hook), querying raw session data (use vaudeville:session-analytics), or general workflow questions.

Promote, demote, or edit vaudeville rule tiers. Use when the user says "promote rule", "demote rule", "move rule to warn", "move rule to block", "rule-admin", "change tier", "promote to warn", "promote to block", "demote to shadow", "delete rule", "enable rule", "disable rule", or invokes /rule-admin. Typically used after /tier-advisor produces recommendations. Do NOT use for analysis — use /tier-advisor for that.

Audit existing vaudeville rules for design-level usefulness. Distinct from /tier-advisor (which audits eval data for promotion). This skill asks: does the rule's event × tier combination produce useful behavior change *at any reachable tier*? Surfaces rules that are useless by design — rules whose violations cannot be acted on in a future turn, rules stuck at SLM model ceilings, rules whose max-useful tier is below their current tier, rules duplicated by other rules, and rules that should be hard hooks instead of SLM rules. Use when the user says "audit my rules", "find useless rules", "are my rules useful", "review rule design", "rule audit", "which rules should I delete", "do my rules make sense", "design audit", or invokes /rule-audit. Do NOT use for promotion analysis (use /tier-advisor) or modifying tiers (use /rule-admin).

Query Claude Code's own JSONL session logs via DuckDB for usage analytics, tool patterns, error forensics, hook impact analysis, and routing decisions. Use when the user asks about their Claude usage patterns, tool frequencies, error rates, permission denials, agent routing, skill invocations, MCP server usage, session timelines, hook effectiveness, or any question about "how has Claude been working". Also trigger when the user says "session analytics", "query my logs", "tool usage", "how often do I use", "check my sessions", "analyze my usage", "hook stats", "what hooks fired", or asks about specific tool/agent/skill/hook behavior across sessions. Also trigger for "vaudeville hook analytics", "hook effectiveness", "hook firing rate", "enforcement stats", or any question about how vaudeville hooks are performing. This skill turns raw JSONL into a queryable DuckDB database — use it instead of writing ad-hoc Python scripts to parse logs. Do NOT use for debugging current code issues, reading individual session transcripts, or questions about Claude's capabilities — this skill is for aggregate usage analytics across historical sessions.

26 hooks across 26 events