mechinterp-next-step-planner
Propose next mechanistic interpretability experiments based on research state, hypotheses, and existing evidence
Discover Testing & security skills, compare source repositories and read the instructions for each workflow.
Propose next mechanistic interpretability experiments based on research state, hypotheses, and existing evidence
Domain expert for Splatoon 3 competitive gear abilities and weapon meta. Use when: (1) Labeling or classifying gear build features from SAE analysis, (2) Understanding why specific ability combinations appear i...
Orchestrate a systematic research program to investigate and meaningfully label SAE features
Manage feature labeling workflow - queue management, label storage, similar features, progress tracking
Quick "first look" overview of SAE features - top tokens, activation stats, weapons, families, sample contexts
Reference catalog of recurring SAE feature patterns - check when labeling new features, add new patterns as discovered
Analyze relationships across multiple SAE features - co-activation patterns, shared drivers, and subsystem identification
Execute mechanistic interpretability experiments from JSON specs - family sweeps, itemsets, interactions, minimal cores, validation
Match SAE features between Ultra (24K) and Full (2K) models based on activation patterns and token overlap
Reference for ability family semantic groupings and gameplay meanings
Use ALWAYS. Core development rules for production-ready code - guard clauses, no shortcuts, no mocking.
Use when creating, reviewing, or merging pull requests. Includes CI checks, review requirements, and merge approval rules.
Use when implementing ANY new feature or fixing bugs. Test-Driven Development - write tests BEFORE implementation code.
Use before claiming ANY work is complete. Evidence before claims, always. Non-negotiable verification gate.
PFLICHT bei JEDER Antwort mit Code-Änderungen! Die 8-Punkt-Checklist MUSS als ERSTES in der Response stehen. Keine Ausnahmen.
Use when deciding whether to delegate work to subagents. Lists available agents with triggers and capabilities. Delegate aggressively.
Use when facing ANY technical issue. Systematic debugging - root cause investigation BEFORE attempting fixes.
Use when starting a session, beginning a new task, or encountering errors. Graphiti is the knowledge graph memory system for storing patterns, fixes, preferences and learnings across all projects.
Use before ANY code changes. Ensures work happens in isolated git worktrees, not on main branch. Critical safety check.
Check Pull Request on current branch, and obtain unresolved review comments. `gh` command is required. Use this when user instructed with `PR review comment` `PR レビューコメント` etc.
Comprehensive Material Design 3 (M3) and M3 Expressive guidance for building modern, accessible, and engaging user interfaces. Use when designing or implementing Material Design interfaces, reviewing component...
下载YouTube视频,支持自定义画质与格式选项。当用户请求下载、保存或获取YouTube视频时,可使用此功能。支持多种画质设置(最佳、1080p、720p、480p、360p)、多种格式(mp4、webm、mkv)以及仅音频下载为MP3。
本书提供专家级指导,帮助读者遵循现代最佳实践编写模块化、可投入生产的Bash脚本。当用户需要执行以下任务时,本书将派上用场:(1)编写新的Bash脚本;(2)改进或重构现有的shell脚本;(3)为脚本添加错误处理和适当的结构...
draw.io图表的创建、编辑和审查。用于.drawio XML编辑、PNG转换、布局调整以及AWS图标的使用。
The default order uses GitHub repository stars from our indexed data. Stars belong to the source repository and may be shared by multiple skills. They indicate repository interest, not skill installs or an independent quality score.