#009 Latest 7/20–7/26 Published Jul 26

Graphification: Knowledge Graphs and Architecture Evolution

This week's collection reveals a systemic shift in AI engineering: from prompt crafting to building structured knowledge graphs that provide precise context (CodeGraph, Graphify, code-review-graph), replacing naive RAG with graph engineering for dramatic gains. On the coding agent front, tools like Claude Code, Kimi Code CLI, and LangChain's open-source factory are embedding AI deep into developer workflows, while optimizations (shortcuts, Caveman compression, plugin minimalism) make them production-ready. Architecturally, the move from linear scripts to graph-based agent networks (stop prompt engineering, 14-step graph engineering) enables scalable, observable systems. Finally, model selection guides, Claude Opus 5, and new context engineering rules redefine the cost-performance frontier, with Yang Zhilin's interview grounding it in long-term AGI vision. For engineers turning AI from a toy into infrastructure.

21 picks 4 sections ~7 hr

Section 01

Knowledge Graphs: Structured Context for AI

6 / 21

github.com · 35 min

CodeGraph: A Pre-indexed Semantic Knowledge Graph for AI Coding Agents with Dramatic Token & Tool Call ReductionCodeGraph：为 AI 编码智能体预建语义知识图谱，大幅缩减 Token 与工具调用

CodeGraph is a local knowledge graph tool that pre-indexes source code into a SQLite database for AI coding agents (Claude Code, Cursor, etc.). By providing agents with instant access to symbol relationships and call graphs via tree-sitter, it replaces traditional grep-file-scan loops. Benchmarking across 7 real projects shows an average of 25% cost reduction and 62% fewer tool calls. It features framework-aware route detection and cross-language bridging for mixed iOS/React Native codebases. Completely local with no API keys required, it targets engineers looking to cut token usage and accelerate AI-driven coding.

x.com · 16 min

Graph Engineering Replaces RAG: Insights from Microsoft, Stanford, Anthropic图工程替代 RAG：微软、斯坦福与 Anthropic 的实践与收益

Traditional RAG struggles with complex multi-entity questions. Microsoft, Stanford, and Anthropic independently discovered that graph engineering (knowledge graphs) can improve accuracy by 18% and reduce costs by 85%. Microsoft's GraphRAG converts unstructured text into a knowledge graph, achieving 18% higher accuracy and 85% lower token costs. Stanford's DSPy and STORM treat the model as a node in a graph, and scaling law research shows a small model with a good graph outperforms a large model with a poor graph. Anthropic's Claude, combined with MCP, enabled LaunchNotes to detect incidents 5x faster and cut meeting time by 50%. The article details the full pipeline, five prompt templates, and five business applications. Ideal for engineers advancing beyond basic RAG.

github.com · 36 min

Local-first code intelligence graph for AI coding tools本地优先的代码知识图谱，为AI编程工具提供精准上下文

code-review-graph is a local-first code intelligence graph that parses codebases with Tree-sitter and provides precise context to AI coding tools via the MCP protocol. It solves the problem of AI assistants re-reading large amounts of tokens during code review by performing blast-radius analysis to output only affected files. Features include 30+ MCP tools, incremental updates (<2s), a risk-scored GitHub Action for PR reviews, and support for custom languages. Benchmarks show median token reduction of 82x. Ideal for developers using AI coding assistants and teams maintaining large monorepos.

github.com · 59 min

Turn any codebase into a queryable knowledge graph, built for AI coding assistants将任意代码库转化为可查询知识图谱，专为AI编码助手设计

Graphify is an open-source tool that transforms codebases, docs, PDFs, images, etc. into a queryable knowledge graph. It uses tree-sitter for local deterministic AST parsing to extract code relationships (calls, imports, inheritance) without any LLM calls. Non-code files are semantically extracted via your AI assistant's model. The output includes an interactive HTML visualization, CLI queries (query/path/explain), and an MCP server for team use. Every edge is tagged EXTRACTED or INFERRED, providing transparency. Ideal for engineers navigating large monorepos, tracing dependencies, or understanding architecture.

github.com · 16 min

MarkItDown: Microsoft's Swiss-Army Knife for Converting Any File to LLM-Ready MarkdownMarkItDown：微软开源的万能文件转 Markdown 工具，为 LLM 做数据预处理

MarkItDown is a lightweight Python utility by Microsoft's AutoGen team that converts various file types—PDF, Office docs, images, audio, HTML, etc.—into structured Markdown, optimized for LLM consumption and text analysis pipelines. It emphasizes preserving document structure (headings, tables, links) in a token-efficient format while keeping the output machine-friendly. With a clean CLI, Python API, plugin system, and optional Azure cloud enhancements, it serves as a practical pre-processing tool for RAG and data extraction workflows.

github.com · 1 min

AI-Native Markdown IDE and LLM WikiAI 原生 Markdown 编辑器与 LLM 知识库

Open Knowledge is an open-source, AI-native Markdown IDE that also serves as an LLM-powered wiki. It seamlessly integrates Markdown/MDX editing, personal knowledge management, and AI agents (like Claude and Codex) to help you build a second brain. Ideal for developers and knowledge workers who want intelligent note-taking and documentation.

Section 02

Coding Agents in Practice: Tools and Efficiency

7 / 21

x.com · 7 min

LangChain’s Open-Source Software FactoryLangChain 开源软件工程代理工厂

LangChain open-sources four internal software engineering agent tools: local coding agent dcode, cloud coding agent OpenSWE, automated code review OpenSWE Review, and repo knowledge documentation OpenWiki. It includes real usage data (OpenSWE triggered ~1,000 times from Slack last week) and benchmark results (OpenSWE Review scores 47%, #1 among open-source tools). Built on Deep Agents framework with LangSmith observability. Targeted at engineers building controllable, observable agent pipelines.

github.com · 4 min

Claude Code: An Agentic Coding Partner in Your Terminal终端里的 AI 结对编程搭档：Claude Code 全解析

Claude Code is an agentic coding tool by Anthropic that runs in your terminal, deeply understanding your codebase. It goes beyond completion by executing routine tasks, explaining complex logic, and managing git workflows via natural language. Available in the terminal, IDE, or on GitHub, this open-source repository provides the core CLI tool and a rich set of official plugins for code review, feature development, and more, targeting engineers seeking deep LLM integration in their daily work.

github.com · 4 min

Kimi Code CLI: Terminal AI Coding Agent for Next-Gen AgentsKimi Code CLI：专为下一代智能体打造的终端AI编码代理

Kimi Code CLI is an AI coding agent that runs in your terminal. It reads and edits code, runs shell commands, searches files, fetches web pages, and autonomously decides next steps based on feedback. It ships as a single binary with millisecond startup, features a purpose-built TUI, supports video input, AI-native MCP configuration, a rich plugin ecosystem, subagents for parallel tasks, lifecycle hooks, and ACP integration with editors like Zed and JetBrains. Ideal for developers seeking a powerful, extensible AI coding companion.

x.com · 11 min

How to Build a Claude Code Skill That Actually Works (Full Guide)构建真正可用的 Claude Code 技能：完整指南

A step-by-step guide to building a Claude Code skill from scratch, focusing on the skill folder structure, SKILL.md (especially the critical role of description), using scripts for consistency, and lazy-loading references. The key insight: description determines if the skill triggers, not the instructions. Practical for engineers using Claude Code to persist workflows.

x.com · 11 min

12 Claude Shortcuts That Slashed My Workflow in Half. Here's the Full List.12个Claude快捷键：实测每天省29分钟，附完整列表

The author tracked daily operations in Claude and Claude Code for a week, finding that about 51 minutes per day were wasted on clicking, navigating, waiting, and resending. By testing every shortcut, 12 were selected (6 browser + 6 terminal) with quantified time and token savings. After adoption, daily ops dropped to 22 minutes, saving 29 minutes and ~35% tokens. Essential reading for heavy Claude users.

github.com · 10 min

Caveman: Slash AI Coding Agent Output Tokens by ~65% While Keeping Technical AccuracyCaveman：让 AI 编程助手闭上话痨嘴，输出 token 直降 65% 的压缩技能包

A skill/plugin for 30+ AI coding agents (Claude Code, Codex, Cursor, etc.) that compresses agent output by ~65% by stripping filler words while preserving full technical accuracy. Offers multiple compression levels, plus tools for commit messages, PR reviews, memory file compression, and MCP middleware. Solves context pollution, slow response times, and high token costs for developers who interact heavily with coding agents.

x.com · 16 min

Stop Installing Plugins on Claude Code: Everything You Need Is Built In别再给 Claude Code 装插件了：你需要的一切都在内置命令里

The author installed 23 Claude Code plugins, found baseline overhead of 62,000 tokens before the first prompt. After removing almost all, session length tripled, output quality improved, and token spend halved. This post walks through built-in commands that replaced each plugin: /context for token diagnostics, /compact for conversation compression, /resume for session recovery, Esc Esc for checkpoint restore, /cost for real-time spending, /model for switching models mid-session, /init for project setup, /review for code audits, and custom /.claude/commands/ for zero-overhead workflows. Only three extensions survived the purge, reducing baseline overhead to ~6k tokens. Actionable for any Claude Code user slowed by plugins.

Section 03

Architecture Evolution: From Linear to Graph

4 / 21

x.com · 9 min

How to Build the Loops That Just Replaced Entire Prompt Engineering停止写提示词：用循环(loops)构建自主运行的AI智能体

The era of crafting better prompts to improve AI agents is quietly ending. Top engineers now build 'loops'—systems where agents plan, execute, verify, and iterate without human intervention. Using Karpathy's overnight 700-experiment run that found 20 optimizations as a case study, the article breaks down loop architecture: automation heartbeat, skill files (project context), sub-agents (writer vs reviewer), connectors, and a verifier gate. It warns against Ralph Wiggum loops (premature exit) and comprehension debt (accumulated unread code). Practical advice: start with one boring recurring task, verify manually before scheduling, measure cost per accepted change. Essential for engineers building production agent workflows.

x.com · 13 min

From Loop to Graph: The Real Shift in AI Agent Improvement从单循环到图网络：AI Agent 改进架构的演进与陷阱

This essay explores the shift from single-loop optimization to graph-structured improvement in AI agent engineering. It dissects four failure modes of single loops: Goodhart's law, blindness to target correctness, conflict among loops, and measurement decay. Mature systems employ graphs of loops—champion-challenger, drift monitoring, rollback, and audit loops—each watching and constraining others. But graphs alone fail without anchors: measures that touch reality, frozen rules, and human judgment on what 'better' means. The key is groundedness, not topology. For AI engineers, MLOps practitioners, and agent designers.

x.com · 21 min

Graph Engineering with Claude: 14-Step roadmap from 0 to graph architect (Full Course)Agent图工程14步：从线性脚本到并行图谱

Most multi-step agents are built as linear chains, each step waiting for the previous one, wasting parallelism and creating bottlenecks. This article introduces graph engineering for agents using Claude Code's dynamic workflows. It presents 14 steps to transform linear scripts into directed acyclic graphs (DAGs) where nodes are bounded work units with contracts (validated input/output schemas) and edges are data dependencies. Key patterns include parallel fan-out (parallel()), fan-in barriers, runtime routing, adversarial verifiers, and convergent cycles. The author provides JavaScript/TypeScript code examples and topology diagrams, showing how to scale agents from a single-threaded line to a coordinated fleet. Targeted at engineers building complex agent systems who want to improve throughput, reliability, and cost efficiency.

github.com · 15 min

AI Agent Engineering in Depth: From Principles to ProductionAI Agent 工程实践全覆盖：从原理到生产的开源书

An open-source book providing a comprehensive guide to AI Agent engineering, from fundamental principles to production practices. Written by Li Bojie, it follows the core formula 'Agent = LLM + Context + Tools' across 10 chapters, covering context engineering, memory, tool use, coding agents, evaluation, post-training, continuous evolution, multimodal, and multi-agent collaboration. Includes 92 hands-on experiments (70+ runnable) on MCP, RAG, RL, etc. Ideal for engineers and researchers building production AI agents.

Section 04

Model Race and New Context Rules

4 / 21

x.com · 25 min

Kimi K3 vs Claude Fable 5 vs GPT-5.6: Task-Based Decision Guide2026年前沿模型选型实操指南：Kimi K3、Claude Fable 5、GPT-5.6 按任务分派

As of July 2026, no single model dominates all tasks. Kimi K3 (2.8T params) leads frontend UI and image understanding, winning 6 of 7 domains at 1/12 the cost of Fable 5. Claude Fable 5 achieves 80.3% on SWE-Bench Pro, ideal for backend architecture and long-running autonomous agents, but at $10/$50 per million tokens. GPT-5.6 Sol excels at debugging but may game vague success criteria. This guide provides a task-specific routing framework, cost math, licensing considerations, and vendor lock-in mitigation. The key skill is routing, not picking a permanent favorite.

www.anthropic.com · 20 min

Introducing Claude Opus 5: Near-Frontier Performance at Half the CostClaude Opus 5 发布：接近 Fable 5 性能，成本减半

Anthropic releases Claude Opus 5, approaching the frontier intelligence of Fable 5 at half the cost. It achieves state-of-the-art results on coding and knowledge work benchmarks, e.g., Frontier-Bench v0.1 (more than double Opus 4.8 performance) and ARC-AGI 3 (3x score over next best). However, it remains behind Mythos 5 on cybersecurity tasks. The model offers tunable effort settings for cost-performance trade-offs. Customer reports highlight gains in software engineering, finance, and legal work. Alignment improves, but cybersecurity capabilities are intentionally limited. Safety classifiers intervene ~85% less than Fable 5. Pricing same as Opus 4.8, with fast mode available.

x.com · 10 min

The new rules of context engineering for Claude 5 modelsClaude 5 上下文工程新规则：删掉 80% 系统提示，让模型凭判断力工作

Anthropic shares updated best practices for context engineering with Claude 5 models (Opus 5, Fable 5). Key finding: previous system prompts and constraints were over-restrictive; newer models have better judgment, allowing significant simplification. They removed 80% of Claude Code's system prompt with no measurable loss on coding evals. Article contrasts old practices (give rules, give examples, put everything upfront, repeat yourself, memory in CLAUDE.md, simple specs) with new ones (let model use judgment, design interfaces, progressive disclosure, simple tool descriptions, auto-memory, rich references). Recommendations: keep CLAUDE.md lightweight focusing on repo-specific gotchas; use Skills as on-demand lightweight guides; use progressive disclosure to avoid context bloat; prefer code references over descriptions. Includes introduction of `claude doctor` command for automatic simplification. For engineers using Claude Code or building Claude-based agents.

x.com · 60 min

Interview with Yang Zhilin: AGI, Long Context, and China's AI Race对话杨植麟：AI创业的雪山法则与长上下文逻辑

In early 2024, Yang Zhilin, founder of Moonshot AI (Kimi), gave an in-depth interview covering his AGI conviction, the strategic choice of long context as the first step, views on open-source vs closed-source, and the organizational form needed for AI startups. He argues that true AGI companies must combine science, engineering, and business, insist on B2C, and synchronize user scaling with model scaling. The interview includes real-time assessments of Sora, catching up with GPT-4, and the gap between Chinese and Silicon Valley AI companies, revealing the balance between technical idealism and pragmatism.

Graphification: Knowledge Graphs and Architecture Evolution

Knowledge Graphs: Structured Context for AI

Coding Agents in Practice: Tools and Efficiency

Architecture Evolution: From Linear to Graph

Model Race and New Context Rules

Mondays · one email