https://patrykreba.com/blog Notes from building AI products in production. https://patrykreba.com/blog/skills-encode-how-you-work https://patrykreba.com/blog/skills-encode-how-you-work Thu, 01 Oct 2026 09:00:00 GMT About 120 agent skills live across my repos. Many were installed, a few dozen were written for the repo, and those are the ones that matter. What a skill is, the kinds I keep, the gotchas they carry, and how I write one now. https://patrykreba.com/blog/headwind-19-days https://patrykreba.com/blog/headwind-19-days Thu, 01 Oct 2026 09:00:00 GMT Headwind is an iOS app for rehearsing hard conversations at work. It went from its first commit on 28 August to the App Store on 16 September 2026, on an engine we already had. https://patrykreba.com/blog/hillclimbing-site-performance https://patrykreba.com/blog/hillclimbing-site-performance Thu, 01 Oct 2026 09:00:00 GMT This site renders a nebula shader, a starfield and animated orbits behind every page. On a mid-range phone it pinned the main thread at 100%. Here is how I measured it, which fixes stuck, and which ones I threw away. https://patrykreba.com/blog/lovestack-first-year https://patrykreba.com/blog/lovestack-first-year Thu, 01 Oct 2026 09:00:00 GMT LoveStack went from an empty project skeleton on 24 September 2025 to an iOS app on the App Store on 16 June 2026. This is the dated build log, and the decisions that shaped it. https://patrykreba.com/blog/sonnet-5-5-and-gpt-6-1-sol https://patrykreba.com/blog/sonnet-5-5-and-gpt-6-1-sol Thu, 01 Oct 2026 09:00:00 GMT Anthropic shipped Claude Sonnet 5.5 on 28 September and OpenAI shipped GPT-6.1 Sol on 29 September, both at $2 in and $10 out per million tokens. What each one is, why the cost per task differs by about ten times, and where I would put each one. https://patrykreba.com/blog/decision-records-for-agents https://patrykreba.com/blog/decision-records-for-agents Thu, 01 Oct 2026 09:00:00 GMT A prompt lasts one session. A decision record lasts until it is superseded, and every agent and person reads it. How I write ADRs so agents stop undoing settled decisions: titles that state the rule, invariants with tests, and pointers from the instruction file. https://patrykreba.com/blog/mcp-server-design https://patrykreba.com/blog/mcp-server-design Thu, 01 Oct 2026 09:00:00 GMT The MCP server behind BBC Marble's ChatGPT desk has 50 tools. The design lessons that made them usable: short lists per person, names that never die, merged reads, descriptions written for the model, and refusals that say what happened. https://patrykreba.com/blog/human-in-the-loop-approvals https://patrykreba.com/blog/human-in-the-loop-approvals Thu, 01 Oct 2026 09:00:00 GMT Drafting is cheap and sending is not. The approval patterns we use when an AI writes a company's email and WhatsApp messages: one word, bound to one version, with free undo, veto windows and no yes from an empty room. https://patrykreba.com/blog/evals-for-product-prompts https://patrykreba.com/blog/evals-for-product-prompts Thu, 01 Oct 2026 09:00:00 GMT Headwind's debrief judges how a person handled a hard conversation, so every number on it has to be real. How we test product prompts: count what can be counted, verify every quote, build small eval sets from failures, and let a verifier try to refute each review finding. https://patrykreba.com/blog/following-ai-since-chatgpt https://patrykreba.com/blog/following-ai-since-chatgpt Thu, 01 Oct 2026 09:00:00 GMT I've followed every major model and tool release since ChatGPT launched, mostly without writing it down. This is the look back: which moments actually changed how I build software, and what I built at each step. https://patrykreba.com/blog/reading-leaderboards-together https://patrykreba.com/blog/reading-leaderboards-together Thu, 01 Oct 2026 09:00:00 GMT Seven leaderboards, seven different questions. What each one measures, who is on top of each today, why they disagree, and the checklist I use to turn them into a model choice. https://patrykreba.com/blog/how-i-use-claude-code https://patrykreba.com/blog/how-i-use-claude-code Thu, 01 Oct 2026 09:00:00 GMT The loop I run most days across three products: write the task down, let the agent build in its own worktree, check the result with commands rather than summaries, and read the diff where a mistake would be expensive. https://patrykreba.com/blog/building-a-voice-twin https://patrykreba.com/blog/building-a-voice-twin Thu, 01 Oct 2026 09:00:00 GMT The Talk to me button on this site starts a voice call with an AI version of me. It runs on GPT-Live over WebRTC, and a second, smaller model listens for requests to see something and puts it on your screen. https://patrykreba.com/blog/open-weights-september-2026 https://patrykreba.com/blog/open-weights-september-2026 Thu, 01 Oct 2026 09:00:00 GMT Where open weights stand at the end of September: sizes, licenses, scores and prices for Qwen3.8, GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, plus the Xiaomi model that now tops them all, and why the license file matters more than the label. https://patrykreba.com/blog/what-a-task-costs-opus-5-5 https://patrykreba.com/blog/what-a-task-costs-opus-5-5 Thu, 01 Oct 2026 09:00:00 GMT Claude Opus 5.5 lists at twice the per-token price of Sonnet 5.5 and GPT-6.1 Sol. Reading Anthropic's cost post next to Artificial Analysis's per-effort data shows where it is cheap per finished task, and where it is not. https://patrykreba.com/blog/subagents-and-worktrees https://patrykreba.com/blog/subagents-and-worktrees Thu, 01 Oct 2026 09:00:00 GMT Subagents let one session think in parallel; git worktrees let several sessions write in parallel. How I use both, how I split and merge the work, the trap that kept my instructions out of most worktrees, and when one agent is the better answer. https://patrykreba.com/blog/realtime-voice-gpt-live https://patrykreba.com/blog/realtime-voice-gpt-live Thu, 01 Oct 2026 09:00:00 GMT How LoveStack moved voice practice to OpenAI's GPT-Live: a WebRTC call owned by the app, a Convex action that only swaps SDP, a judge model instead of in-band tags, and Gemini kept for the builds already on people's phones. https://patrykreba.com/blog/business-from-chatgpt-mcp https://patrykreba.com/blog/business-from-chatgpt-mcp Thu, 01 Oct 2026 09:00:00 GMT How the operators of a natural stone export company in Dubai run their outreach from ChatGPT, through an MCP server that holds the record, the rules and the one door a message leaves by. https://patrykreba.com/blog/shipping-ios-with-agents https://patrykreba.com/blog/shipping-ios-with-agents Thu, 01 Oct 2026 09:00:00 GMT LoveStack and Headwind are React apps in a native iOS shell, and most of their commits are written with coding agents. The rules, skills and guards that let agents ship TestFlight builds without stranding the builds already on people's phones. https://patrykreba.com/blog/speech-to-speech-2026 https://patrykreba.com/blog/speech-to-speech-2026 Thu, 01 Oct 2026 09:00:00 GMT OpenAI made GPT-Live generally available on 10 September and Google followed with Gemini 3.8 Live on 15 September. How the two differ, what the independent numbers say, and the five things that changed for anyone building voice. https://patrykreba.com/blog/intelligence-index-october-2026 https://patrykreba.com/blog/intelligence-index-october-2026 Thu, 01 Oct 2026 09:00:00 GMT Artificial Analysis now scores 688 models on ten evaluations. Here is the top 12 as of today, what each evaluation measures, where the labs cluster, and what a single number leaves out. https://patrykreba.com/blog/what-voice-ai-costs https://patrykreba.com/blog/what-voice-ai-costs Thu, 01 Oct 2026 09:00:00 GMT Realtime voice bills by the minute, and a call nobody ends keeps billing. The rates we pay across LoveStack, Headwind and this site, the five layers of caps between a call and an open-ended bill, and the gaps we still have. https://patrykreba.com/blog/spending-effort-on-agent-work https://patrykreba.com/blog/spending-effort-on-agent-work Thu, 01 Oct 2026 09:00:00 GMT When an agent writes most of the code, my effort moves to the brief, the checks and the review, and the model's effort becomes a setting I choose per role. Effort levels, model choice, reading diffs, evals and cost. https://patrykreba.com/blog/choosing-a-coding-agent https://patrykreba.com/blog/choosing-a-coding-agent Thu, 01 Oct 2026 09:00:00 GMT I have nine coding agents installed. Two of them do nearly all the work. What each one is for in my setup, which ones I have real evidence for, and how I would choose if I were starting today. https://patrykreba.com/blog/claude-md-that-agents-follow https://patrykreba.com/blog/claude-md-that-agents-follow Thu, 01 Oct 2026 09:00:00 GMT Patterns from the instruction files in my three repos: a map before rules, traps with their reasons, hard stops, a definition of done, links instead of copies, and the step most people skip, which is tracking the file so every worktree gets it.