How we index images for RAG
Kapa converts documentation images into text captions at indexing time rather than sending pixels to multimodal models on every query. The approach improves answer quality while adding only 1–6% to per-query cost.
Category
AI
Evaluations, model behavior, and applied automation that holds up in production.
GitHub - Andyyyy64/whichllm: Find the local LLM that actually runs and performs best on your hardware. Ranked by real, recency-aware benchmarks, not parameter count. One command, run it instantly. · GitHub
whichllm is a Python CLI that detects or simulates hardware, then recommends runnable local LLMs using benchmark quality, recency, memory fit, and estimated speed—not parameter count alone. It also supports hardware planning, upgrade comparisons, instant model chat, and scriptable JSON output.
GitHub - addyosmani/agent-skills: Production-grade engineering skills for AI coding agents. · GitHub
Agent Skills is an MIT-licensed, cross-agent pack of 25 structured engineering workflows that pushes AI coding agents toward specs, tests, reviews, security checks, and safe releases. It offers lifecycle slash commands, native adapters for major coding agents, and evaluation-backed skill routing.
What is Ralph Loop? A New Era of Autonomous Coding
Ralph Loop uses persistent AI-agent iteration to autonomously plan, build, and refine software, while Ralph TUI makes those long-running workflows observable and controllable. The article pairs it with Cubic AI Review to address code-quality risks created by faster AI-generated changes.
I used this Claude Skill to turn my vibe-coded projects into coding courses
Codebase to Course is a free Claude Skill that converts a repository into an interactive, visual HTML tutorial. It helped the author understand both an 800-line Java/JavaFX app she wrote and an unfamiliar codebase in minutes.
apfel - Your Mac Already Has AI
apfel exposes Apple’s built-in on-device language model as a zero-configuration CLI, local OpenAI-compatible server, and chat tool. It targets Apple Silicon Macs on macOS Tahoe, emphasizing privacy, scripting, and no token costs.
apfel/demo at main · Arthur-Ficial/apfel · GitHub
A collection of practical shell scripts that use apfel and Apple Intelligence for offline, on-device developer assistance. The demos turn natural-language requests into commands, explanations, project summaries, naming suggestions, and system insights.
microgpt
Andrej Karpathy distills the algorithmic core of GPT training and inference into a dependency-free, 200-line Python script. It trains a 4,192-parameter GPT-2-like model on 32,033 names, making every essential component—from autograd to attention—inspectable.
🍌 Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing
Apple’s Pico-Banana-400K provides roughly 400,000 quality-controlled text–image–edit examples for training text-guided image editors, including supervised fine-tuning, preference-learning, and multi-turn data. It combines Open Images sources, Gemini-generated instructions and evaluation, and Nano-Banana edits.
I Tested Gemini's Nano Banana AI Image Editor—and These 7 Tricks Seriously Impressed Me
Gemini’s Nano Banana makes quick AI image edits—such as object removal, compositing, text generation, and custom filters—surprisingly accessible and often more convincing than Adobe’s generative tools. Its low-resolution output, blur, and lack of precise manual controls still rule it out for professional photo workflows.
The rise of the AI crawler
Vercel and MERJ find that AI crawlers now generate substantial web traffic but generally cannot render JavaScript and waste over one-third of requests on 404s. Sites that need AI visibility should server-render critical content and maintain disciplined URL hygiene.
GitHub - snarktank/ai-dev-tasks: A simple task management system for managing AI dev agents · GitHub
AI Dev Tasks is a lightweight prompt-based workflow for directing AI coding agents through requirements, granular planning, and reviewable implementation steps. It replaces one-shot feature prompts with a PRD-to-task-list process that keeps humans in control.
llm-min.txt: Min.js Style Compression of Tech Docs for LLM Context
llm-min converts technical documentation into compact Structured Knowledge Format manifests for AI coding assistants. Its multi-stage Gemini pipeline claims 90–95% typical token reduction while retaining APIs, relationships, and usage patterns.
Best practices for Claude Code - Claude Code Docs
Anthropic’s guide to operating Claude Code effectively: constrain context, give the agent verifiable goals, and use planning, tooling, and independent review to scale reliable work. It emphasizes that autonomy works best when paired with explicit checks and disciplined session management.
Aqua Voice - Fast and Accurate Voice Dictation
Aqua Voice is an AI dictation tool that converts speech into polished, context-aware text across desktop, mobile, coding, and AI workflows. It claims up to 230 WPM, 49-language support, and 97.3% accuracy on the AISpeak benchmark via its Avalon model.
Ingesting Millions of PDFs and why Gemini 2.0 Changes Everything
Gemini 2.0 Flash makes large-scale PDF-to-Markdown extraction and LLM-driven chunking economically viable, delivering roughly 6,000 pages per dollar with near-perfect OCR in internal testing. Its major remaining limitation is unreliable bounding-box localization for source citations.
How to Train an AI Image Model on Yourself
Cory Zue shows how to fine-tune a Flux image model on 10–15 personal photos using LoRA and Replicate in under an hour. The workflow costs roughly $2.50 per model to train and about $0.03 per generated image.