Toolkit for performing structured code reviews. Covers identifying bugs, security issues, performance problems, and style violations. Use when asked to review code changes, pull requests, or individual files.

Use when the user asks to discover, browse, search, suggest, recommend, or install agents and agentic resources. Author, validate, or import OASF records; push, sign, and publish records; discover, search, browse, suggest, or recommend agents like MCP servers, A2A agents, or agent skills; verify signatures, name ownership, or security scans; synchronize data between servers; or install/uninstall agents and agentic resources into coding agents like VS Code Copilot, Claude Code, or Cursor.

Use this skill to interact with an AGNTCY Directory (DIR) instance. This is a bootstrap: it installs the complete agntcy-dir skill (setup, authoring, publishing, discovery, verification, sync, install workflows) from the AGNTCY Directory GitHub repository, then defers to it.

Authors, reviews, installs, and debugs GitHub Agentic Workflows in repositories, including workflow markdown, frontmatter, gh aw compile and run flows, safe outputs, security guardrails, and operational patterns. Use when creating or maintaining GH-AW automation. Don't use for standard deterministic GitHub Actions YAML, generic CI pipelines, or non-GitHub automation systems.

Build, validate, test, and update registries and catalogs that follow the Agentic Resource Discovery (ARD) specification. Use this whenever the user works with an ai-catalog.json, a capability manifest, an ARD/AIR catalog or Agent Registry, urn:air: identifiers, trustManifest/attestations, representativeQueries, or an Agent Finder / discovery service — including authoring a new manifest, scaffolding one, fixing schema or URN errors, running conformance/validation, probing a registry's /search, /explore, or /agents REST endpoints, reviewing trust and federation metadata, or preparing to publish at /.well-known/ai-catalog.json. Trigger it even when the user only says "ARD", "agentic resource discovery", "AI catalog manifest", "agent registry", or "make agents discoverable" without naming the file.

Deploys agent skill collections from any GitHub repository with a /skills folder to one or more distribution surfaces: GitHub releases, Claude Code marketplace, VS Code plugin marketplace, and Copilot CLI plugin marketplace. Handles pre-flight validation, conventional commit analysis, version bumping across surface configs, and surface-specific publishing with dry-run support. Use when releasing, publishing, or deploying a skills collection to any supported marketplace or creating a GitHub release for a skills repository. Don't use for deploying non-skill packages, npm modules, Docker images, or Azure resources.

Installs, configures, audits, and operates Agent Package Manager (APM) in repositories. Use when initializing apm.yml, installing or updating packages, validating manifests, managing lockfiles, compiling agent context, browsing MCP servers, setting up runtimes, or packaging resolved context for CI and team distribution. Don't use for writing a single skill by hand, generic package managers like npm or pip, or non-APM agent configuration systems.

Create or update a project roadmap with major milestones. Generates the configured roadmap artifact (default .ai-factory/ROADMAP.md) — a strategic checklist of high-level goals. Use when user says "roadmap", "project plan", "milestones", or "what to build next".

Run a strict multi-iteration Reflex Loop with phases (PLAN, PRODUCE||PREPARE, EVALUATE, CRITIQUE, REFINE) to improve an artifact until quality gates pass or iteration limits are reached. Use when user asks for iterative refinement, quality-gated generation, or "generate -> critique -> refine" loops.

Reliability gate for answers. Forces evidence-based reasoning, explicit uncertainty, and “insufficient information” instead of guesses. Use when user says “be 100% sure”, “no hallucinations”, “only if verified”, “grounded answer”, or when stakes are high.

Analyze project and generate Docker configuration: Dockerfile (multi-stage dev/prod), compose.yml, compose.override.yml (dev), compose.production.yml (hardened), and .dockerignore. Includes production security audit. Use when user says "dockerize", "add docker", "docker compose", "containerize", or "setup docker".

Archive completed plans and roadmap milestones. Moves finished plans to the archive directory and optionally trims closed milestones from ROADMAP.md. Use when user says "archive plans", "clean up plans", "archive completed", or "trim roadmap".

Custom commit workflow that replaces the built-in aif-commit. Demonstrates the skill replacement feature of the extension system.

A hello world skill provided by the aif-ext-hello extension. Demonstrates extension-provided skills.

Language-independent tokenizer treating text as raw Unicode. Supports BPE and Unigram algorithms. Fast (50k sentences/sec), lightweight (6MB memory), deterministic vocabulary. Used by T5, ALBERT, XLNet, mBART. Train on raw text without pre-tokenization. Use when you need multilingual support, CJK languages, or reproducible tokenization.

Fast tokenizers optimized for research and production. Rust-based implementation tokenizes 1GB in <20 seconds. Supports BPE, WordPiece, and Unigram algorithms. Train custom vocabularies, track alignments, handle padding/truncation. Integrates seamlessly with transformers. Use when you need high-performance tokenization or custom tokenizer training.

Meta's 86M prompt injection and jailbreak detector. Filters malicious prompts and third-party data for LLM apps. 99%+ TPR, <1% FPR. Fast (<2ms GPU). Multilingual (8 languages). Deploy with HuggingFace or batch processing for RAG security.

NVIDIA's runtime safety framework for LLM applications. Features jailbreak detection, input/output validation, fact-checking, hallucination detection, PII filtering, toxicity detection. Uses Colang 2.0 DSL for programmable rails. Production-ready, runs on T4 GPU.

Meta's 7-8B specialized moderation model for LLM input/output filtering. 6 safety categories - violence/hate, sexual content, weapons, substances, self-harm, criminal planning. 94-95% accuracy. Deploy with vLLM, HuggingFace, Sagemaker. Integrates with NeMo Guardrails.

Anthropic's method for training harmless AI through self-improvement. Two-phase approach - supervised learning with self-critique/revision, then RLAIF (RL from AI Feedback). Use for safety alignment, reducing harmful outputs without human labels. Powers Claude's safety system.