惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

博客园 - 司徒正美
aimingoo的专栏
aimingoo的专栏
MongoDB | Blog
MongoDB | Blog
云风的 BLOG
云风的 BLOG
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
酷 壳 – CoolShell
酷 壳 – CoolShell
博客园 - 聂微东
Y
Y Combinator Blog
T
Tailwind CSS Blog
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
S
SegmentFault 最新的问题
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
博客园 - 【当耐特】
IntelliJ IDEA : IntelliJ IDEA – the Leading IDE for Professional Development in Java and Kotlin | The JetBrains Blog
IntelliJ IDEA : IntelliJ IDEA – the Leading IDE for Professional Development in Java and Kotlin | The JetBrains Blog
J
Java Code Geeks
美团技术团队
Google DeepMind News
Google DeepMind News
博客园_首页
Apple Machine Learning Research
Apple Machine Learning Research
T
The Blog of Author Tim Ferriss

DEV Community

Apache Iceberg Metadata Tables: Querying the Internals Hermes, The Self-Improving Agent You Can Actually Run Yourself Unity vs Unreal: 5 Things I Had to Relearn the Hard Way Building Agentic Commerce Infrastructure: Overcoming SQLite Concurrency for Autonomous Procurement Agents Solana Accounts vs Databases HTML Table Borders My first post! I'm kinda excited The Page Root Was the Wrong Unit How to audit what your IDE extension actually sends to the cloud I Migrated 23 Make.com Scenarios to n8n and Cut My Bill by 60% — Complete Migration Guide (2026) Solving a Logistics Problem Using Genetic Algorithms Claude Code Skills Explained: What They Are & When to Use Them (2026) Maintaining Apache Iceberg Tables: Compaction, Expiry, and Cleanup Zero-Idle Local LLMs: Running Llama 3 in AWS Lambda Containers We scanned 8 B2B SaaS companies across 5 categories. ChatGPT named the same 12 brands in every answer. How To "Market" Yourself As A Tech Pro We scanned 500 MCP servers on Smithery. Here is what we found. HTML Basics for Beginners – Markup Language, Elements and Types of CSS DiffWhisperer: How I Turned Cryptic Git Diffs into Architectural Stories with Gemma 4 I built a version manager for llama.cpp using nothing but vibe coding. Unit Testing vs System Testing: Key Differences, Use Cases, and Best Practices for 2026 A game design textbook explains why products with fewer features win How to Build a Raydium Launchpad Bonding Curve in 5 Minutes with forgekit How to turn an AI prototype into a production system How Data Lake Table Storage Degrades Over Time Partition and Sort Keys on DynamoDB: Modeling data for batch-and-stream convergence Auto-Generate Optimized GitHub Actions Workflows For Any Stack With This New CLI Tool Unchaining the African Creator Economy The Treasure Hunt Engine Gotcha - A Lesson in Constrained Performance great_cto v2.17 - no more tambourine dance When Catalogs Are Embedded in Storage SafeMind AI: Instant Health & Safety Intelligence What Is PKCE, How It Works & Flow Examples AI Agent Failure Modes Beyond Hallucination Fastest Way to Understand Stryker Solana Accounts Explained to a Web2 Developer TV Yayın Akışı Sitesi Geliştirirken Öğrendiğim Teknik Dersler $500 Challenge Drop My First Look at Google's Gemma 4: A Quick Introduction How I use an LLM as a translation judge Best Calendar and Scheduling API for Developers — 2026 Comparison Agentic AI in Travel: Why UCP Isn't Travel-Ready Yet — and What We Measured I Finished Machine Learning. And Then Changed The Plan. The Five-Thousand-Line File The AI Whirlwind: Why Your Local Agent Matters More Than Ever I Built an Oracle DBA That Lives in Telegram. It Cut a 500K-Row Scan to 5 - After Asking Permission. The Day 2 Reality of Running a Kubernetes Lab on Your Mac: Stop/Start, CKS Scenarios, and What I Learned Building It. n8n for Airtable Power Users: 5 Automations That Take Your Base to the Next Level Validating Gemma 4 for Industrial IoT: A Governance Pattern VS Code Now Credits Copilot on Every Commit by Default Astro and Islands Architecture: Why Your Portfolio Doesn't Need React for Everything Booting from FAT12: How I added file reading to my x86 kernel Unity’s AI agent went public: the developers of a static analysis tool on what that means for code quality Anna's Archive publica un llms.txt para los LLMs que rastrean su catálogo CRDTs for Offline-First Mobile Sync Why I Built Mneme HQ: Preventing AI Agent Architectural Drift Google Antigravity 2.0 Is the I/O 2026 Announcement You Should Actually Care About I Built a Pay-Per-Call Crypto Signal API with x402 — Heres the Architecture JWT Token Refresh Patterns in React 19: Avoiding the Silent Auth Death Spiral 🚀 “From Prompts to Autonomous Agents: What Google I/O 2026 Changed” The Power of Distributed Consensus in Autonomous SOCs Sixteen TUI components, copy-paste, no dependency The Boring Reliability Layer Every Autonomous Agent Needs Nven - Secret manager Building Multi-Tenant Row-Level Security in PostgreSQL: A Production Pattern The Hardest Part of Being a Developer Isn't Coding Building Vylo — Looking for Collaborators, Partners & Early Support I Thought Memory Fades With Time. It Actually Fades With Information. ORA-00064 오류 원인과 해결 방법 완벽 가이드 I registered an AI agent at 1 AM and something cracked open in my head Pitch: Nven - Sync secrets. Ship faster. Why y=mx+b is the heart of AI From Routines to a Crew — Building a System That Plans Its Own Work & executes it 25 React Interview Questions 2026 (With Answers) — Hooks, React 19, Concurrent Mode An open source LLM eval tool with two independent quality signals Using Dashboard Filtering to Get Customer Usage in Seconds from TBs of Data Skills, Java 17, And Theme Accents 4 Hard Lessons on Optimizing AI Coding Agents Arctype: Cross-Platform Database GUI for LLM Artifacts Your robots.txt says GPTBot is welcome. Your server says 403. Organizing How to Use AWS Glue Workflow 5 n8n Automations Every Digital Agency Should Be Running (Bill More, Work Less) Getting Started with TorchGeo — Remote Sensing with PyTorch Designing a Scalable Cross-Platform Appium Framework Google Antigravity 2.0 & Slash Commands Building a Unified Adaptive Learning Intelligence with Gemma 4, Flutter, and Multi-Model Orchestration Looking for beta testers for a £60 server management application The Disk-Pressure Incident That Taught Me to Always Set LimitRanges and Other Lessons from Mirroring EKS Locally. Why AI Should Not Write SQL Against ERP Databases Vibe coding works until it doesn't. The debt is real. Shipping at the Edge: Migrating a Coffee Subscription Platform to Cloudflare Workers Stop Tab-Switching: A Developer's Guide to Color Tools That Actually Fit the Workflow DevOps vs MLOps vs AIOps: What Changes, What Stays, and a Simple Roadmap to Get Started Run Powerful AI Coding Locally on a Normal Laptop 5 n8n Automations Every WooCommerce Store Needs (Save 10+ Hours/Week) What I Learned Building My Own AI Harness Hytale Servers Will Fail Treasure Hunts Until We Fix Our Event Handling Redux in React: Managing Global State Like a Pro Unfreezing Your GitHub Actions: Troubleshooting Stuck Deployments and Protecting Your Git Repo Statistics Unlocking Project Discoverability on GHES: A Key to Software Engineering Productivity
I built a skill that makes AI-generated AWS diagrams actually usable
Alexey Vidan · 2026-05-22 · via DEV Community

Every AWS architecture diagram I generated with AI needed 20–30 minutes of manual cleanup. Colored backgrounds on group boxes, broken icons, inconsistent flow direction, edge labels overlapping services. At that point, I might as well have drawn it from scratch.

Prompt to AI Agent to .drawio diagram

I wanted a draft I could hand to a client the same day. So I built a skill (a markdown file with rules and reference data) that teaches the AI my specific layout and styling rules. It works in both Claude Code and Kiro CLI. No runtime dependencies, no MCP server.

What was wrong with raw generation

Claude Code and Kiro CLI can produce draw.io XML out of the box. The output opens in draw.io. But "opens" and "looks professional" are different things.

Here's what raw generation actually produces:

Colored backgrounds on groups. AWS Cloud boxes had blue fills, VPC boxes had green fills. Real AWS diagrams use transparent groups with just a border.

Inconsistent flow direction. Sometimes left-to-right, sometimes top-to-bottom, sometimes random. No two diagrams followed the same convention.

Icon pattern confusion. draw.io has two icon patterns with opposite strokeColor rules. In my generations, the AI mixed them up roughly one in four times, producing empty colored squares. The repo calls this out as the single biggest cause of broken icons in AI-generated diagrams.

Edge labels on top of icons. Orthogonal routing with no explicit exit/entry points meant lines went through other services.

No spacing discipline. Icons crammed together with 50px gaps, or scattered across a huge canvas with no rhythm.

Each one is a 30-second fix on its own. Doing all of them on every diagram adds up to that 20–30 minute tax.

The two-pattern rule

draw.io's AWS library (mxgraph.aws4.*) has two icon types that require opposite styling:

Service-level: strokeColor=#ffffff (white, required)
Resource-level: strokeColor=none (required)

Enter fullscreen mode Exit fullscreen mode

Mix these up and you get empty squares or invisible glyphs. The icon names look interchangeable but they're not. I extracted all 270+ names from draw.io's source code (Sidebar-AWS4.js) and documented which pattern each one uses.

Five rounds of refinement

The first version got icons right but layouts were still mediocre. Each round came from opening the generated diagram in draw.io and noting what I'd manually fix, then encoding that fix as a rule.

Round 1: Icons. Extracted 270+ icon names, documented the two patterns, added a "never guess, always look up" rule.

Round 2: Layout. Increased spacing from 150px to 220px horizontal. Added explicit exit/entry points on edges. Removed edge labels that were redundant with icon labels.

Round 3: Edge routing. Changed from rounded=0 to rounded=1 (sharp corners to smooth curves). Added explicit exitX/exitY/entryX/entryY for vertical connections. This stopped lines from routing through other icons.

Rounds 4 and 5 were about restraint and structure. The AI was labeling every edge with obvious things, "Write" on an AWS Lambda to Amazon DynamoDB connection, so I added a "when NOT to label" rule and a 1–2 word cap. Then a title block, a full-canvas background rectangle for clean PNG export, and an audience-mode toggle (technical vs non-technical) to control detail level.

After five rounds, the skill enforces: left-to-right flow with 220px+ horizontal spacing, no colored backgrounds on any group container, verified icon names only (from 8 category reference files), and explicit edge routing so lines don't cross icons.

Example output

"Create an event-driven order processing architecture with Amazon SQS, AWS Lambda, Amazon DynamoDB, and Amazon EventBridge"

Event-Driven Order Processing

"Create a real-time IoT analytics pipeline with Amazon Kinesis, AWS Lambda, Amazon S3 data lake, and Amazon DynamoDB"

Real-Time IoT Analytics

"Create a 3-tier web application with Amazon CloudFront, Application Load Balancer, Amazon ECS on AWS Fargate, Amazon Aurora, and Amazon ElastiCache"

3-Tier Web Application

Icons render. Flow is left-to-right. No colored backgrounds, no overlapping edges. I can adjust these in under 5 minutes instead of 30.

Install

Claude Code:

/plugin marketplace add vidanov/aws-architecture-diagram-skill
/plugin install aws-architecture-diagram@vidanov-skills

Enter fullscreen mode Exit fullscreen mode

Kiro CLI:

mkdir -p ~/.kiro/skills/aws-architecture-diagram
cp kiro/SKILL.md ~/.kiro/skills/aws-architecture-diagram/SKILL.md
cp -r references ~/.kiro/skills/aws-architecture-diagram/references

Enter fullscreen mode Exit fullscreen mode

Once installed, try this prompt to verify it works:

"Create a serverless API with Amazon API Gateway, AWS Lambda, and Amazon DynamoDB"

You should get a clean left-to-right diagram with correct icons and no colored backgrounds.

What's next

The current output is good. Not perfect. I still adjust things manually. The next step is multiple diagram styles for the same architecture: a technical view for engineers, a simplified view for business stakeholders. Same system, different audience, different drawing.

Try it on your next architecture review. If the generated diagram needs fixes I haven't covered, open an issue. The skill improves from real usage, not theory.

GitHub | Project website


The project was built with Kiro CLI.