Naturalize the current or referenced assistant output into context-appropriate Japanese while preserving meaning. Use when the user asks to make the previous response sound natural, remove AI-like Japanese, fix translationese, adjust register, or choose Japanese terminology from audience and domain context.
Build and use repository-local agent memory in .agents/memory.db and .agents/memory.yml. Use when a task needs session memory, work logs, handoff recall, SQLite-backed agent logs, artifact-aware repository memory, or idempotent memory setup across agent sessions.
Research evidence-backed translations and established terms across domain-specific, industry, and general usage. Use when selecting English/Japanese terminology, validating standard translations, resolving product wording collisions, or proposing audience-specific term usage with citations.
Create, update, validate, and review Agent Skills for Codex and Claude Code. Use when the user asks to "create a skill", "make a new skill", "define a skill", "write SKILL.md", "improve a skill description", "set up skill directories", "add Codex openai.yaml metadata", "validate skill structure", apply SkillOpt-style evaluation, or distinguish standard Agent Skills requirements from Codex- or Claude Code-specific practices.
Create atomic Git commits with proper messages and co-author signatures. Use when the user asks to "commit my changes", "make a commit", "save my work", or mentions committing code changes. Focuses on commit message formatting, atomic commits, and maintaining consistent co-authorship attribution.
Reverse-engineers existing implementations, design files, and documentation into a reusable design system while preserving Adobe Spectrum's Principles/Foundation/Content structure. Use when users ask to derive design rules from legacy projects, normalize heterogeneous inputs, and produce evidence-based design system documentation.
Evaluate whether frontend implementation tests deliver meaningful defect-prevention value using the Testing Trophy as the primary decision model. Use when asked to assess frontend test strategy, static vs unit vs integration vs end-to-end layer placement, mock-heavy suites, browser/device realism, accessibility semantics, visual regression value, existing mutation reports, or flaky test risk.
Iteratively refine a GitHub issue body that holds an execution plan by running one or more external review agents (Codex CLI or Claude Code CLI) against it, classifying the feedback by severity, applying deterministic fixes automatically, surfacing ambiguous choices to the user, and updating the issue body via `gh issue edit`. Each batch consumes `concurrency` consecutive round numbers; about 30% of the workers in a batch use a dissenting reviewer for model diversity. Reviewers always run at their maximum reasoning effort. Use when the user asks to "iterate on issue #NNN", "refine the plan in #NNN", "run a review round on #NNN", "run N parallel reviews on #NNN", "loop reviewer feedback on #NNN", or wants to repeat reviewer rounds against a GitHub issue until no blockers remain.