$ show nibzard/awesome-agentic-patterns
This repository is a meticulously curated, structured catalog of AI agent design patterns, functioning primarily as a knowledge engineering project with a sophisticated software delivery system. The core data asset is a collection of over 100 Markdown files in the `patterns/` directory, each adhering to a strict YAML frontmatter schema (title, status, authors, category, tags, signals, anti_signals, evidence_grade) and a standardized body (Problem, Solution, How to Use, Trade-offs, References). The primary value is the research and curation effort that synthesizes academic papers (arXiv), industry talks, engineering blogs, and production implementations into a reusable knowledge base. The web application (Astro) is the delivery mechanism, not the primary artifact. The architecture is a clean three-layer data pipeline. The data layer (`patterns/`, `research/`) contains the source-of-truth markdown. The transformation layer (`scripts/build-data.ts`, `scripts/claude-research-loop.sh`, `scripts/arxiv_scanner.py`) processes this data, generating JSON indexes, knowledge graphs, RSS feeds, and research reports. The delivery layer (`apps/web/`) is a well-crafted Astro static site that renders the catalog with full-text search, interactive graph exploration, and a decision guide. The build pipeline is deterministic, enforced by CI, ensuring the generated artifacts are always in sync with the source patterns. The repository demonstrates high craftsmanship in its data engineering. The `scripts/build-data.ts` is a standout example of a deterministic, multi-format static site generator. The `scripts/claude-research-loop.sh` implements a stateful, fault-tolerant LLM orchestration pattern for automated research. The `scripts/pattern_similarity_checker.py` provides a robust deduplication mechanism using Jaccard and Levenshtein distances. These components are genuinely reusable for any project building a structured knowledge base or automated research pipeline. The pattern schema itself is a well-defined standard for documenting design patterns. Caveats: The vast majority of patterns are authored by a single person (@nibzard), creating a bus factor and potential perspective bias. Many patterns are labeled 'emerging' or 'proposed' with low evidence, relying on self-reported or unverified claims. The dual Python/TypeScript toolchain introduces minor friction, and the research loop depends on external APIs (Claude CLI, arXiv) that may change. Despite these caveats, the engineering quality of the pipeline and the value of the curated knowledge base are exceptionally high for a community-driven catalog.
The LLM identified 5 files worth highlighting for reuse:
The central build orchestrator. Reads all pattern markdown files, extracts frontmatter and sections, builds a knowledge graph of related patterns, and generates JSON indexes, RSS feeds, sitemaps, and llms.txt. It is deterministic (CI checks output), well-structured, and handles edge cases. This is the core data engineering asset of the repository and a reusable pattern for any static site generated from structured content.
A stateful shell script that orchestrates an LLM (Claude CLI) to generate research reports for each pattern. It uses a tracker JSON file for recovery, skips existing reports, validates output, and manages API rate limits. This is a reusable pattern for any automated research or content generation task, demonstrating non-obvious craftsmanship in fault-tolerant LLM orchestration.
A Python script that detects near-duplicate patterns using Jaccard similarity on tags and Levenshtein distance on titles and content. It provides a weighted score and flags potential overlaps. Reusable for maintaining any large catalog or knowledge base, solving a common data quality problem.
Implements a clean, two-tier strategy for finding related patterns: first checks for explicit 'related' links in the pattern metadata, then falls back to shared tags. This is a simple, effective, and reusable recommendation algorithm that avoids over-engineering.
A generic, reusable Astro component for sticky section navigation. It uses IntersectionObserver to highlight the current section in the viewport. It is responsive, accepts props for sections and current section, and is decoupled from specific content. Demonstrates high craftsmanship for a UI component.