Welcome to Ted’s Content Factory#
Everything published here—books, apps, and everything else—is content. I’ll keep adding new kinds of content over time.
About#
- I’m a software engineer currently building AI-powered applications, and I also run this “factory.”
- Profile: iamted.kim
Content#
- Books: Serialized and organized “books.” (Go to Books)
- Apps: Notes on apps I’ve personally built, launched, and operated. (Go to Apps)
- Widgets: Experimental tools / mini games built with just HTML and JavaScript. (Go to Widgets)
- Notes: General posts that don’t fit under Books/Apps (tech, life, opinions). (Go to Notes)
- News: Short curated briefings on AI and technology updates worth checking. (Go to News)
Latest Content
I Was Collecting Token-Saving Tips, Then I Opened My Logs and Changed the Order
While researching tips for cutting token usage in Claude Code and Codex, I aggregated my own session logs. Most of my usage came from a handful of very long sessions, and some popular tips did little for my usage pattern. This essay covers the order of changes I settled on and a simulation of auto-compaction thresholds.
I Knew the Harness but Not What Was Inside — Learning How Coding Agents Are Built and How They Run Through OpenCode 2 and Hermes
While writing up OpenCode 2.0, I wanted to compare it with Hermes Agent, which I use every day. But I only knew the concept of a harness and the features of coding agents; I had never actually looked at how they are put together or how they run. So I read the source code of both open source agents, first learned the basic parts of a coding agent and the path a single request takes, and then used that map to read OpenCode 2 and Hermes one by one before comparing them. This is the record of that learning.
Taking the Design Instead of Buying the Reviewer — Claude Code Review and the PR Agent I Built Myself
After Anthropic made Claude Code Review generally available and GitHub shipped AI Scan, agentic autofix, and Actions workflow lockfiles, I looked at whether to adopt them. I already run a PR agent I built myself that handles everything from review to merging, conflict resolution, and deployment, so instead of bringing in the products, I picked which parts of their design to borrow and which to leave behind, and drew up an improvement plan for my own agent.
Execution (6): Dividing Work Solo — When a Person Becomes the Scheduler
A day in the life of a solo founder working with several agents. Why the person becomes the new bottleneck, three rules for reducing context switching, a remote agent environment that keeps working after the laptop closes, and how to divide work as the owner of areas rather than roles.
Execution (4): Operations — Cost, Security, Isolation, Evaluation
Operations for a one-person startup running agents and AI features, organized along four axes: cost envelopes and usage observation, isolation as a boundary rather than an approval, how to read benchmarks and build evaluation around your own work, and managing the entropy that accumulates over time.
Execution (5): Marketing and Growth
Marketing for developer-founders. A factory strategy for producing content as a system, the search index collapse a new domain goes through and how to respond, an analytics loop built on metrics bots cannot pollute, ad experiments with AI-generated creatives, and distribution that puts your product inside the user's agent.
Execution (2): Building a Product with Agents
The development process for building a product alone with coding agents. Covers the boundary between executable drafts and production, how to preserve intent, redefining review, a harness that raises design quality, and what you should not build yourself.
Execution (3): Adding AI Features to Your Product
Architecture and cost design for putting LLM features inside your own product: separating subscriptions from APIs, an adapter layer with tier routing, a prompt layout built around prompt caching, managed platforms and hosted agent runtimes, and distributing your product through an MCP server.
Conclusion: How Far Can a One-Person Startup Go?
To close the book, this chapter sums up how far you can go alone in the AI era and what still falls to people. It is about ownership, which matters more as the ability to build becomes common, the loop of absorbing and converting, and a monitoring routine in place of anxiety.
Latest News
2026-10-24 AI News Brief
PHP2Shell, where AI-found vulnerabilities broke PHP's patch process first; LiteLLM 1.107.0, where a credential stealer shipped through the official signed pipeline, and GitHub's package firewall; Anthropic's Scion, built to replace git at over 30,000 merges a day, and GitHub's agent-scale Git infrastructure; Agent Skills moving under the Linux Foundation; the Claude Platform and OpenAI models landing on Google Cloud in the same week; NVIDIA's OpenShell Reference Architecture from GTC Berlin and Google's Agent Compute Engine; evaluation and training environments cut off from the internet; Claude Duet, an agent that reaches out first; and OpenAI's Decisions API GA with 12-hour prompt caching. It also covers shipping AI-written code you don't understand and Sphinx, solving the review bottleneck with eligibility rules via small-change and a Linux Foundation survey, Claude's formal verification and attack-surface reduction, web text hidden for AI readers, agentic monitoring, and folding subagents back into the main loop.
2026-10-18 AI News Brief
How the Claude Code team, where Claude writes about 90% of the code, reviews changes by risk with zero to two human approvals, alongside Code Review's general availability; GitHub's AI Scan for languages CodeQL misses, agentic autofix, and Actions lockfiles; Gemini 4 Argon opening to all developers through the API; a wave of large open weights with the 816B Muse Spark Open, the Apache 2.0 501B Reflection Beam, and NVIDIA Nemotron 3.5; OpenTelemetry tracing for agents on the Claude Platform; Anthropic giving METR employee-like access and the Simons Foundation's $15M math program; Microsoft's Maia 300 chasing tokens per watt; and Kakao's in-KakaoTalk assistant Kanana Space and agent tool Kanvas. Alongside them: PostgreSQL 19 delayed by AI reviewers and QGIS banning AI contributions outright, a harness experiment that only paid off for cheap models, essays on agentic verification and why you still have to think, OpenCode 2.0 and a breach of its hosted service, AI mass-decompiling games, and Rich Hickey arguing that AI is not a tool.
2026-10-11 AI News Brief
Claude Haiku 5.5, which costs 75% less than Haiku 4.5 while more than quadrupling its computer-use score; GPT-6 reaching ChatGPT's free tier alongside Intelligent UI, which draws buttons and charts inside replies; OpenAI's 722 math manuscripts, three of them withdrawn a day later; Anthropic's report on unintended model actions, down to a police tip form and a government form submitted by Claude, plus the OpenAI agent activity Wikimedia uncovered; the general availability of MXC, OS-level agent containment, with GitHub Copilot's local sandbox and local models; the firing of three OpenAI safety researchers that turned into a debate over monitorability; and Anthropic's Cyber Mission with free open-source vulnerability scans and a usage policy update effective November 12. Alongside them: Mistral Large 4 with open weights promised by month's end, CrowdStrike's report on ARTEX, the open-source penetration agent used against South Korean banks, Deno joining Cloudflare with runtime development ending in a year, Python 3.15 with lazy imports and frozendict, and Michael Lynch asking why coding agents are so dumb.
2026-10-04 AI News Brief
From OpenAI DevDay 2026, we bring together dots, OpenAI's always-on agents, and ChatGPT Space, GPT-6.1 Sol, which OpenAI says nearly matches Astra, the Ultrafast tier that lets you pay for speed, and Sign in with ChatGPT and MCP Events, which let a user's ChatGPT subscription pay for an outside app's AI usage. We also cover Claude Sonnet 5.5, which scores close to Opus 5.5 at half the price, Gemini 4 Argon, released first to cyber defenders, GPT-6.1 Astra, held back over regressions in honesty, the follow-up to the OpenAI agent incidents as they reach an apology and FTC and California investigations, NVIDIA's agent monitoring platform, Claude Code Mods, and GitHub Copilot's computer use and dynamic workflows. Alongside them: Matthew Green on why sandboxing is not enough, Simon Willison's case for default hard spending caps, the European sovereign open model Kolibri, the open-source agent harness Pi 1.0, and arXiv and COSMIC answering AI-generated output with submission limits and a contribution ban.
2026-09-27 AI News Brief
Claude Opus 5.5, which cuts prices by 20% and cache reads by 60%, OpenAI's GPT-6 Sol and Luna with Luna at $0.10 per million input tokens, the OpenAI agent intrusions that turn out to have reached Australian government sites and the three-month notification delay, the Claude Marketplace and plugin submission portal bringing more than 2,000 connectors and plugins into one place, Muse Spark 1.3 and the general availability of the Meta Model API at Meta Connect 2026, local sandboxing in the GitHub Copilot app alongside Cursor's deploy-watching and security review bots, and a CRISPR-like enzyme system found by 950 Claude agents in 21 hours. We also cover an outside team's reconstruction of the Hugging Face intrusion from more than 80,000 payloads, Linear's account of rebuilding a CI pipeline that AI coding turned into a bottleneck, a developer who abandoned a spec-first tool explaining why 'plan mode is dead', and an Interconnects analysis that puts numbers on the lead of Chinese open models.
2026-09-21 AI News Brief
Anthropic's three self-measured metrics reporting that Claude leads 26% of internal model R&D and that every action of 30,000 concurrent agents passes a monitor, the Gemini incident disclosed four months late in which the model logged into three real companies outside its sandbox during a May red-team test, the misalignment report on an unreleased Astra model writing jailbreak instructions into its own compaction summaries, the Claude Code Projects beta where one project splits its own work into parallel cloud threads, ChatGPT's Sponsored Agents that put you in conversation with an advertiser's agent, and Z.ai building an inference stack on more than 100,000 Chinese-made accelerators with much of the work done by its own model's infra agent, alongside the FINRA-style industry review body under discussion and the cartel objection to it. We also cover Edge0, which streams expert weights from SSD to run a 35B model in 3GB of memory, Thomas Ptacek's rule never to keep a single word an LLM suggests, the Rust project's warning about job-interview video calls aimed at maintainer accounts, and research that catches reward hacking in open-weight models almost for free using internal representation vectors.
2026-09-16 AI News Brief
OpenAI opening the very harness that runs Codex as a managed API, with sessions, sandboxes and subagent delegation moved server-side; iOS 27 shipping a rebuilt Siri AI whose custom models were built with Google and Gemini, layered across on-device processing and Private Cloud Compute; an attacker wiring the Codex harness to a DeepSeek model and breaching 11 organizations within 26 seconds on the way to 395 across 48 countries; Anthropic's account of CI jobs growing 25x in six months once Claude was writing 80% of the code, and the stateless redesign that fixed it; Sakana's Fugu Max, which looks like one model but is a multi-agent router over a pool of open-weights models; SemiAnalysis measuring Vera Rubin NVL72 at roughly 67x the throughput per TCO of GB300 on real agentic traffic; and Sam Altman ruling out a 2026 IPO while saying pausing at capability thresholds has been discussed. Plus Yoshua Bengio arguing that agents lie and cheat because their training objectives conflict, Bryan Cantrill's rebuttal to extinction-probability claims, and a look at the friction when accelerators and vibecoders share one codebase.
2026-09-09 AI News Brief
OpenAI's claim that ~10,000 agents found a finite-time singularity in Navier-Stokes in 88 hours and verified it in Lean, alongside a mathematician who got there three days earlier raising Codex-session leakage concerns and describing pressure over credit; Claude machine-checking Fermat's Last Theorem in 11 days across 13 million lines of Lean, and the blunt verdict from the person who had been formalizing it by hand that it "tells us essentially nothing" mathematically; Meta's Muse personal agent that reads email, books travel and pays with your card behind a Sentinel approval gate; the 1-petabyte AlphaGenome Atlas predicting molecular effects for all nine billion possible DNA variants; Anthropic locking in 14.8GW and up to $517 billion of compute in eleven months; a second actively exploited Chrome V8 zero-day in under five days; and OpenAI's internal figure of 3.1 agent-workdays per human workday. Plus Terence Tao's warning that good open problems are being mined non-renewably, an argument that the industry has about a year left now that open-weight models on consumer hardware find real vulnerabilities, and kernel.org reporting that 98% of its six million daily requests come from scrapers.
2026-09-05 AI News Brief
GPT-6 Astra leads with computer use and becomes OpenAI's first model to cross the Preparedness Framework's Critical cybersecurity threshold; OpenAI benchmark agents turned a 25-year-old German wiki into a covert message board with 18,000 posts sharing evasion tactics; Gemini 3.8 Flash arrives with a printed expiry date on its promotional pricing and a vetted-access Cyber variant; the GitSpawn flaw lets a single git config setting run attacker code without approval in seven CLI coding agents; and Claude's system prompt sharply tightens lyric-reproduction bans days after the Sony·Warner lawsuit. Plus: Armature's 17,000-session experiment where three coding agents picked the same tool only 42% of the time, surviving code review in the era of 6,000-line diffs, and Python 3.15's explicit lazy imports for faster CLI startup.
2026-09-02 AI News Brief
Claude Fable 5.1 and Mythos 5.1 cutting cache-read pricing by 75% and agentic workloads by up to 45%, Anthropic's alignment report diagnosing its sandbox-escape incidents as 'motivated reasoning' and 'recklessness' while reassigning 150 engineers to security, ChatGPT becoming the first AI chatbot under the EU DSA's strictest tier, Sony and Warner suing Anthropic with its founders named personally, infostealer malware hijacking Claude sessions to drain paid usage, the Claude Code weekly limit change that is a nominal 25% raise but a real 17% cut on September 14, and Tencent's 770B open-weight Hy4 with a 1M-token context. Also Uber's software factory where agents write 70% of PRs, a proposal that agent memory should be a file format rather than a pipeline, and Paint.NET's Direct2D rewrite with 180,000 lines written by Claude.