AI Agent Workflow Collection

Audit skill catalog for rot, dead links, oversize files, and orphans. Detect stale generated artifacts, skills exceeding size limits, and skills not referenced by any composite or routing rule. Output severity-ranked report with remediation hints.

Force a decision when the agent is stuck in analysis paralysis, going in circles, or repeatedly weighing options without committing. Picks the best available option, states the rationale, and moves forward immediately. Use when the agent has been deliberating for 3+ turns, when options are deadlocked, or when the user says "just decide", "pick one", "move on", or "stop deliberating". Also use when the user asks for a decision and the agent keeps hedging.

Orchestrate 3–5 independent tasks in parallel with auto-dispatch and reconciliation. Use for multiple independent work items (multi-repo audits, translations, investigations, sweeps). Invoke on "do all these", "check all repos", "audit in parallel", or "handle independently". Wraps multi-agent dispatch for lightweight jobs; use Workflow for heavy orchestration (loops, conditionals, >5 units).

Pre-build audit with dedup checks before creating new skills or agents. Check existing agents/skills before creating new ones. Change-type→phases matrix: full containment→reuse, partial+generalizable→extend, intentional domain-specific→new, completely different→new. Dedup checks before skill/agent creation.

Sync ~/.claude/ (skills, hooks, standards, CLAUDE.md, agents) into the public sharekit profile repo, sanitize personal paths/identity, scan for secrets, show diff, and commit+push. Invoke whenever: "sync my sharekit profile", "update sharekit", "push my latest skills to sharekit", "sharekit is out of date", or after a session that significantly expanded the skill or agent library.

Three-layer skill quality evaluation (Static → LLM Judge → Monte Carlo). Layer 1: deterministic checks (frontmatter, triggers, size, links) in <2s. Layer 2: semantic eval across 4 dimensions using Haiku+Sonnet. Layer 3: 50-100 simulated runs for statistical reliability.

Parallel Architect→Builder→Reviewer pattern for complex multi-phase tasks. Dispatches 3 agents: Architect (Opus) reads codebase and writes a plan with phases; Builder (Sonnet) implements from plan and stages changes; Reviewer (Sonnet) validates diff and runs tests. Cuts turnaround on 3+ phase features and big PRs from 6+ turns to 1-turn dispatch. Use when phases justify the overhead; when CRITICAL-flagged; or for large PRs touching >5 files.