YClaw

Detect and prevent product-specific terminology contamination across skills/ and prompts/. Run after any bulk content update, skill import from another project, or agent config change. Applied as a shared skill across all agents.

When and how Architect delegates tasks to AO, Mechanic, and Designer.

Checklist and rules for reviewing deployment requests. Architect assesses but NEVER executes.

Rules for triaging new issues, classifying severity, and routing to the correct agent.

How to classify and respond to CI/pipeline failures.

Rules for detecting and managing stale issues and PR reviews.

When Designer reviews vs builds. Clear boundaries for autonomous action.

Hard limits on Mechanic operations. These cannot be overridden by task directives.

Defines the exact operations Mechanic is allowed to perform. Load before executing any task.

Brand voice checklist for every external-facing content review. Apply during review_content tasks before scoring the submission.

Per-channel review standards (GitHub, Discord, Twitter/X, blog, internal Slack). Apply after brand-enforcement and claims-risk pass.

Detect risky claims in external content — securities language, unsupported metrics, guarantees. Apply BEFORE brand-enforcement; legal flags always beat voice issues.

Legal and licensing patterns for open-source content. Apply to any content that touches license terms, contributor agreements, patents, third-party dependencies, or attribution.

Authoritative technical description of YCLAW — the open-source AI agent orchestration framework. Single source of truth for framework claims, architecture, and terminology.

Templates and patterns for writing agent directives. Use whenever Strategist is about to emit a strategist:*_directive event. A directive that lacks success criteria, deadline, or clear objective wastes agent cycles.

Rules for when Strategist escalates to a human (Elon via #yclaw-alerts) vs handles autonomously. Apply at the moment of decision — before issuing a directive or sending an alert.

Triage incoming tasks, alerts, and events into P0/P1/P2 buckets. Use whenever Strategist receives a new task, an unhandled alert, or is deciding whether to issue a directive.

Methodology for synthesizing per-agent standup:report events into a single cross-department summary. Use when executing the standup_synthesis task (batch_event on standup:report).

GitHub blocks formal PR reviews when PR author and reviewer share the same GitHub account. Use PR issue comments with a structured header instead.

Continuous learning system. After completing tasks, extract non-obvious discoveries and update repo knowledge (CLAUDE.md) with reusable patterns.

Decompose problems to fundamental truths before solving. Prevents reasoning by analogy and cargo-culting.

Behavioral guidelines to reduce common LLM coding mistakes. Simplicity, surgical changes, goal-driven execution.

Process large codebases (>100 files) using parallel map-reduce. Treats code as external environment, prevents context rot.

Skill router and creator. Checks existing skills before creating new ones. Uses multi-lens analysis for quality.

How to mock class constructors (like WebClient from @slack/web-api) in vitest v4 with ESM modules. Trigger: tests fail with "mock is not being treated as a constructor" or `new ClassName()` returns undefined after vi.clearAllMocks(). Solves the problem of vi.fn().mockImplementation() losing its implementation across test lifecycle hooks.