Skip to content
workflows.diy

Writing

Notes from building software with AI coding agents. Each post covers one build and what broke along the way.

Start here

  1. For leaders

    Shadow AI is already running in your business

    Employees already run AI agents and tools outside IT governance, and the AI-use policies most companies wrote in 2024 don't cover them.

  2. For builders

    Making n8n workflows production-ready

    A production n8n workflow needs header authentication on its webhooks, retries for temporary failures and a central error workflow. In our experience, a single retry resolves 60-70% of temporary API and LLM failures.

  3. The latest build

    MiniFrames turns a web page into a video

    MiniFrames renders an HTML page to MP4 on a virtual clock and makes no AI calls. Its companion tools made the 93-second showcase on my own computers.

2026

  1. MiniFrames turns a web page into a video
  2. Shadow AI is already running in your business

2025

  1. Building an AI video clipper in StreamlitWe turned a Jupyter notebook for AI video clipping into a Streamlit app, and the AI wrote the Python code from the module design we set. The modular app takes YouTube URLs or MP4 uploads and offers three transcription methods.
  2. Local PR Buddy reviews code before you pushLocal PR Buddy is a Python CLI that asks an LLM to review your staged or branch changes on your own machine, before you open a pull request.
  3. Making n8n workflows production-ready
  4. code2prompt and Context7 for AI coding contextcode2prompt packs a codebase into Markdown for an LLM, and Context7 supplies current library docs. Both give Gemini in Google AI Studio targeted context for AI-first development.
  5. Turning the ai-sdlc CLI into an Electron GUIWe re-implemented the ai-sdlc Python CLI as a native SDLC module in our Electron and Vue 3 desktop app. The AI wrote the TypeScript and Vue code in under six hours, and the module still needs end-to-end testing before deeper integrations.
  6. SEO audits with Crawl4AI, Python and an LLMA workflow that uses Crawl4AI to gather a site's pages, Python to clean them, and an LLM to write an SEO audit from an AI-written system instruction. It can flag gaps between a site's claimed expertise and its actual content.
  7. Google Drive RAG chatbot workflow in n8nAn n8n community template indexes Google Drive documents in Qdrant and answers questions with Gemini. You have to set the Drive folder and the connections yourself.
  8. Ten Ollama tips for running local AIAfter a model downloads, Ollama runs it on your own machine with no per-call cost and no internet connection.
  9. Six resources for vibe coding with AIVibe coding moves fast, so MCP directories, GitHub trending and rule directories are worth checking every week.
  10. Running Ollama on NVIDIA GPUs for local AII run LLMs and vision models on an RTX 4090 and an RTX 4070 with Ollama, which keeps data on my machine and avoids API call costs. The same setup runs my Desktop Assistant and local n8n workflows.
  11. Using conversational AI as the new IDEFluid Dev is a workflow where most of the code comes from a chat with Google Gemini, and the IDE mainly runs and debugs it. AI-generated code still needs human testing and an understanding of the principles behind it.
  12. Use AI as a hammer for throwaway scriptsMost small coding problems need one script, not a full project. For a one-off task, AI writes the script and the KISS principle says to skip the framework.
  13. Gemini built an image reorder app in under an hourThe Image Reorder & Renumber app (ImageSortAndSave.com) was built by Google Gemini. It runs in the browser, lets you reorder and renumber images, and downloads the set as a ZIP.
  14. Community tips and pitfalls for n8n workflowsn8n users on Reddit keep hitting the same problems, from PDFs with tables that plain text extraction cannot read to reusing data from earlier nodes.
  15. AI workflows in Google Sheets with Apps ScriptGoogle Sheets and Apps Script can call OpenAI, Gemini, WordPress, and Mailchimp, and AI can write most of the script. You can build one even if you do not code.
  16. The 30+ tools in my AI development stackThe 30+ tools behind my AI applications and workflows run from local Ollama models on NVIDIA GPUs to services such as OpenRouter, Perplexity and n8n.
  17. How n8n ties together workflows, agents and MCPAI workflows follow fixed steps, while agents decide their own next step. Use a workflow when the steps are known and an agent when the number of steps is hard to predict. MCP connects either one to tools and data.
  18. Do million-token context windows replace RAG?Models with million-token context windows can hold a small knowledge base in one prompt, but RAG still matters once the knowledge base outgrows the window. Skip RAG for a few hundred pages, and use more advanced retrieval once the data is fragmented, tabular or spread across documents.
  19. Nail the user story before you buildA clear user story saves rework on n8n workflows and AI-assisted builds, because it fixes the problem before the solution. The format names the beneficiary, the action and the value.
  20. Choosing n8n by codability and integrabilityTwo tests separate AI automation tools: codability, meaning how much control the tool gives through code, and integrability, meaning how well it connects to your business systems.
  21. Using OpenRouter to switch between many LLMsOpenRouter.ai gives you one API for many LLMs, so you can swap models without rewriting integration code. To compare two models, send the same prompt to both through OpenRouter and see which answer fits the task better.
  22. Building PlantDex with Vue 3, Supabase, and PiniaPlantDex is a plant cataloging app where Google Gemini generated nearly every line of code, built with Vue 3, Pinia, and Supabase. Gemini built it in four phases, starting with user accounts and plant data storage.
  23. How Supabase runs the PlantDex backendSupabase gave PlantDex a database, user accounts, file storage and serverless functions without a custom backend. A backend-as-a-service supplies these parts pre-built, so a small team does not build and run its own server for them.
  24. 15 tips for AI-first development in AI StudioA detailed system instruction is the most impactful technique when you build with Gemini 2.5 Pro in Google AI Studio.
  25. AI rebuilt this website in Vue 3Gemini 2.5 Pro rebuilt this website in Vue 3 from vanilla JavaScript. A detailed system prompt and specific task prompts guided it, and I validated the results.
  26. A Tkinter GUI for selecting code2prompt filesSelecting a folder in this Tkinter app selects its contents too, and the app builds a minimal set of code2prompt --exclude arguments from your choices.
  27. Our Jina.ai multipage scraper passes 10,000 n8n runsOur Essential Multipage Website Scraper with Jina.ai template has run more than 10,000 times in the n8n community. It reads a sitemap, pulls each page as clean Markdown with Jina.ai, and saves the results to Google Drive.
  28. My most visited n8n AI workflowsMy multi-platform social media content template leads my n8n templates with 38,157 Total Visitors on n8n.io, ahead of YouTube summarization at 21,447.
  29. 3D undersea world built with Three.js and Anime.jsA child's 2D undersea story grew into a 3D world built entirely in code. Three.js models the squid, fish, jellyfish and shark, and Anime.js animates them.
  30. My first Quasar app was a "Hello World" adventureA first Quasar Framework project built with Vite and TypeScript, starting from a Hello World single-page app.
  31. Crawl4AI RAG in my Electron AI assistantI added Crawl4AI to my Electron and Vue 3 assistant so it can answer from crawled websites. Playwright failed inside the FastAPI server on Windows, so the crawler moved into Docker.
  32. Migrating my vanilla JS site to Vue 3We moved a hand-built vanilla JavaScript site to a Vue 3 app built with Vite. The hardest part was turning static pages into components that Vue Router manages.
  33. My desktop AI stack with Electron and OllamaWe built a desktop AI assistant with Electron, Vue 3, local Ollama models, a FastAPI backend and Docker, and an AI coding assistant wrote all of the code. Our job was to define the what and the why, and the generated code still needed thorough testing.
  34. The tech stack behind our desktop AI assistantOur desktop AI assistant combines Electron, Vue 3, FastAPI, Docker, Ollama and Kokoro TTS, with local models for privacy. Electron's main process handles operating system access, and FastAPI runs the Python AI services.