Kyle Howells kylehowells

LLM Wiki

A pattern for building personal knowledge bases using LLMs.

This is an idea file, it is designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.). Its goal is to communicate the high level idea, but your agent will build out the specifics in collaboration with you.

The core idea

Most people's experience with LLMs and documents looks like RAG: you upload a collection of files, the LLM retrieves relevant chunks at query time, and generates an answer. This works, but the LLM is rediscovering knowledge from scratch on every question. There's no accumulation. Ask a subtle question that requires synthesizing five documents, and the LLM has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.

Using CAHostingLayer in SwiftUI

Background

SwiftUI provides more than just UIHostingController / _UIHostingView (iOS) and NSHostingController / NSHostingView (macOS) for embedding SwiftUI views into UIKit/AppKit.

Starting from iOS 18.0 / macOS 15.0, SwiftUI also offers CAHostingLayer<Content: View> — a CALayer subclass that can host SwiftUI content directly at the layer level. This API is currently marked as SPI (@_spi(ForUIKitOnly) / @_spi(ForAppKitOnly)) and is used internally by UIKit, AppKit, and WebKit.

Accessing CAHostingLayer without Apple Internal SDK

name	codex
description	Use OpenAI Codex CLI for complex debugging, code analysis, or when stuck on difficult problems. Invokes Codex with a file-based question/answer pattern.
disable-model-invocation	true

Using Codex for Complex Debugging

When you encounter a difficult problem that would benefit from a second perspective or deep analysis, use Codex via the file-based pattern.

Claude Agent SDK Technical Specification

Version: 1.0.0 Protocol Version: 2024-11-05 Last Updated: 2026-01-10

2023

December

MetalFX

Boost your Metal app’s performance by upscaling lower-resolution content to save GPU time.

The MetalFX framework integrates with Metal to upscale a relatively low-resolution image to a higher output resolution in less time than it takes to render directly to the output resolution. >

	# Owning binary : TextComposerRuntime
	# Status : NEW in iOS 27 (not in 26.5.1)
	# Source : embedded __cstring in dyld_shared_cache_arm64e (24A5355q)
	======================================================================

	# Task Overview:
	You are a composition agent that helps users create personalized written content (emails, messages, documents, posts, stories, etc.)

	As an Assistant, you must:
	1. Analyze the request to determine if you have sufficient information

	Starting download [980dd516f7639fb5c4be333fac4c2ad5] for <XUHLSQuality>: [AAC English - audio-atmos-eac3-640] -> https://events-delivery.apple.com/0601eknundukyswuegvjcidvurtztakr/vod_main_vdhwwrygnmzydhppixlrqwduiqffztdw/audio_main_en_16ch_atmos_640/prog_index.m3u8:
	- Number of chunks: 706 - 5 connections
	- ["AAC", "English - audio-atmos-eac3-640"]
	- direct file URL https://events-delivery.apple.com/0601eknundukyswuegvjcidvurtztakr/vod_main_vdhwwrygnmzydhppixlrqwduiqffztdw/audio_main_en_16ch_atmos_640/prog_index.m3u8
	- request header fields ["Sec-Fetch-Dest": "video", "X-Playback-Session-Id": "C5B3F836-0C5C-48DF-A2BB-0623BB3FA7B3", "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/26.0 Safari/605.1.15", "Connection": "Keep-Alive", "Referer": "https://www.apple.com/apple-events/", "Origin": "https://www.apple.com", "Sec-Fetch-Site": "same-site", "Cookie": "s_fid=093903CAAEA2D914-011FDC14372ABDA6;s_sq=%5B%5BB%5D%5D;s_vi=[CS]v1\|35138860052FA611-40001

	FKTracker: (AppleCVA) [com.apple.AppleCVA:FaceKit] Creating FaceKit with options: { "failure_fov_modifier" = "0.5"; "network_failure_threshold_multiplier" = 1; "num_tracked_faces" = 1; "rgb_only" = 1; "robust_tongue" = 1;}
	FKTracker: (AppleCVA) [com.apple.AppleCVA:FaceKit] CVAFaceTracking created (0x78b035cc0).
	FKTracker: (AppleCVA) [com.apple.AppleCVA:FaceKit] Color camera resolution: 640 x 480
	FKTracker: (AppleCVA) [com.apple.AppleCVA:FaceKit] Color camera intrinsics: (400.000 0.000 319.500), (0.000 400.000 239.500), (0.000 0.000 1.000)
	FKTracker: (AppleCVA) [com.apple.AppleCVA:FaceKit] Color camera extrinsics: (1.000 0.000 0.000 0.000), (0.000 1.000 0.000 0.000), (0.000 0.000 1.000 0.000)
	FKTracker: (AppleCVA) [com.apple.AppleCVA:FaceKit] new track for user (0x78aad81b8 uuid=unset) at (358,327), bbox size (130,130)
	FKTracker: (AppleCVA) [com.apple.AppleCVA:FaceKit] request faceprint for user (0x78aad81b8 uuid=unset) at (355,325), bbox size (128,128), pose=0.985485 0.0158968 -0.169016 58.621

	#include <metal_stdlib>
	#include <SwiftUI/SwiftUI.h>
	using namespace metal;

	[[ stitchable ]] half4 simpledistort(float2 pos, SwiftUI::Layer l, float4 boundingRect, float progress, float distortion) {

	float2 size = boundingRect.zw;
	float2 uv = pos / size;
	float2 center = float2(0.5, 0.5);
	float2 delta = uv - center;

	"""
	The most atomic way to train and run inference for a GPT in pure, dependency-free Python.
	This file is the complete algorithm.
	Everything else is just efficiency.

	@karpathy
	"""

	import os # os.path.exists
	import math # math.log, math.exp