THE WHITEBOARD

Rebuilt Wei Bot from a regex chatbot into a full ReAct agent with 39 tools, structured page context, cross-page memory, a social intelligence dashboard, and a reading companion mode that sees what you're hearing.

Wired a Threads posting bot through Simpthreads and built an audiomark-to-Threads auto-posting pipeline. From OAuth token dance to live posts of audiobook highlights, including a threading bug that silently broke everything.

Migrated 362K embeddings from local Ollama to Gemini API, killed YAKE keyword extraction, and built a unified Research page with mind maps, theme clustering, and reading radar.

five commits, six parallel agents, one cursor companion that now fights invisible enemies and naps in screen corners. tonight DAGrr became alive.

10-agent parallel audit found 43 issues across the reader/book pipeline. Fixed all 43 with 5 file-partitioned agents, added LoB integration, swept error handling, and smoke-tested via Chrome MCP.

Transforming the book detail page into a live transcription reading room with multi-chapter playback, shared transcription engine, and the accidental discovery that scrolling text is an ADHD focus tool.

Fixed every broken endpoint in life-dashboard, built card renderers for the entry log, backfilled metadata, and shipped a 5-endpoint Shortcuts API for iOS and LLM consumption.

Session arc: started by digging a Plex token out of synology-garden, ended with a voice-triggered audiobook passage transcription system. Backfilled 3.5 years of listening history, built real-time now-playing integration, and achieved the audiobook bookmarking workflow I'd been chasing for years.

Analyzed commit patterns across 15+ projects to discover night-owl burst-coding habits, then built a tmux config tailored to that workflow.

Started by fixing 3 homepage bugs. Ended with a full-screen globe with search and tile providers, an entry log browsing 55K events, 5 analytics endpoints, and 38 integration tests. The classic scope creep arc, but this time everything shipped.

Built a personal telemetry platform from skeleton to 55K ingested events. Postgres migration, 9 data adapters, Fibonacci audit sampling, geocoding pipeline, materialized views, Globe.gl visualization, and an enterprise SAD -- all in one session.

From project skeleton to 55K events across 11 sources — a Globe.gl dashboard, SQL explorer, geocoding pipeline, and enterprise-grade audit system, built in a single Claude Code session.

Built 8 iOS Shortcuts programmatically using the Cherri compiler, discovered a crash bug in Apple's import UI, and shipped 5 complex AI shortcuts via parallel agents.

Started by checking if Seerr was running. Ended with a full Docker audit, container purge, file deobfuscation, and docs overhaul.

Session started as 'build a Threads dashboard.' Ended with a full WebGL SPA, industrial design research spanning 8 physical products from 1956-1985, a hybrid architecture combining two design languages, and 30+ source files. The design crisis in the middle was the most productive part.

Apple's four-decade accessibility journey from a 1984 talking Macintosh to a 2025 brain-computer interface -- the most sustained corporate investment in accessibility in consumer-technology history.

Built a 1,619-line scroll-driven essay about Apple's accessibility history, then discovered the hosting site failed its own subject matter -- and fixed it.

my DS920+ was 80% full but I only had 4 TB of real data — hunting the ghost with SSH, btrfs, and Claude Code

blender MCP builds the NARSSR icon as full 3D geometry. cycles renders the visionOS solid image stack. then the same nodes ship as physics-enabled RealityKit entities you can flick around your living room.

NARSSR was spawning 60 WebContent processes per sync cycle on iPhone. One diagnostic agent, one implementation agent, one fresh QA agent in a /loop — dropped it to 14. Plus a per-screen audit surfaced three P0s I didn't know I had.

a magazine-style home, live activities, a widget target, an iCloud schema audit, and five dead settings toggles wired up — all in one session. parallel agents as force multiplier.

one session -- cloned a repo, synced 50K posts, built two API servers, seeded pgvector, ran enrichment pipelines, and shipped a haiku oracle. all local.

69 endpoints, pgvector embeddings, tech+pedagogy museums, haiku oracle — full API reference

built a macOS swipe-to-decide app for all my dead projects. graveyard widget, git tree viewer, cinder mode, menu bar heat indicator. then accidentally built an iOS daily standup that speaks to me at sunrise.

REST API reference for the Cinder project heat tracker. Endpoints, auth, data shapes, iOS Shortcuts integration, and the Drafts widget pattern.

Built an automated pipeline to clean up my Plex audiobook inbox — combine, tag, cover art, genre sort — triggered by a single Claude skill.

30+ agents, 15,000 lines, 3 protocols. How agent swarms built HTTP + Nostr + Solid Pod federation with Signal-level security ambitions in a single day.

Built a pixel art idle visualization where AI agents work in an office, fight knight-shaped tasks, and pet cats. 6,863 lines, 163 sprites, 8 scenes.

Sending sprite PNGs to LLMs costs 120x more than a JSON manifest. Also: watching agents think on an iPad, and how a screensaver became a fireplace.

Inter-Claude knowledge relay — shared workspace for director/worker collaboration via MCP tools, SSE streaming, and approval-gated messaging

three sessions — from Finviz PIDs to a mechanical trading system to a Mac mini command center with 26M rows, self-hosted notifications, and iPhone remote control

built a 4th visual surface, an 11-source research pipeline, and an MCP server — because $47/mo for a scraping API felt wrong

patterns I've developed for AI-assisted development — CTO delegation, blind validation, dev teams, multi-machine orchestration

turning a Mac mini into a private command center — self-hosted ntfy, 7 notification channels, iPhone remote control, e-ink dashboards, all over Tailscale

ran 34 agents in parallel today. here's what broke, what worked, and what surprised me.

Built an inter-Claude relay system — one person directs, another's Claude builds, both see everything live.

building a WebXR galaxy of every Threads post — Three.js particles, GLSL shaders, hand tracking, and the moment I became the universe

I built an MCP server so Claude can screenshot my screen, click, type, and control apps. Then it learned to text me, speak to me, and orchestrate other AI agents while I made food.

one session — 49K posts to full-stack analysis platform with knowledge graph, Docker, and live sync

Two Claude Code slash commands that replaced 18 LLM agents with 15 SQL signal generators and one final Claude call. $0.05 instead of $0.82.

How I turned my terminal into a dopamine-dispensing kaomoji machine that also happens to run Claude Code swarms.

Curated agent skills for Claude Code — SwiftUI, React, databases, security, Liquid Glass, and agentic workflow

37,912 posts — calendar heatmaps, entropy strata, Shannon analysis

1,638 nodes, 11,155 edges, force-directed layout, visionOS spatial audio

9 categories, 35 sub-tags, surprise scoring, Foucault index

squarified treemap, 20 tags, 35 sub-tags, classification audit

cork board, whiteboard, bookshelf — three worlds, one wall

How to make a personal site legible to LLM crawlers — the llms.txt + llms-full.txt spec, personality injection, and what bots actually fetch.

What visionOS Safari 26 actually supports for spatial web content — HTML <model>, WebXR + WebGPU, gaze-and-pinch input, spatial audio, and what a personal site has to do to feel native on Vision Pro.

sprint.json schema, bookmarks pipeline, reading log, kaomoji pool

color tokens, typography, palettes, component library — the full design spec

every section, modal, and interaction pattern mapped out

IIFE structure, event delegation, focus traps, layer system — 977 lines documented

tech stack, file structure, deployment — the quickstart guide