A New Home
Moving to an Astro based new home for my personal blog. And a new location.
Archive
All posts, including hidden.
A Tale of Two Harnesses
Adding tool calls to Parley on a flight, watching Claude make a mess, then drawing the architecture boundary for Claude and watching it execute flawlessly within it. The eureka: Claude Code is the execution harness, Parley is the interface for human to inspect and influence — both running on the same shared filesystem state.
WhatsApp GDPR Fine
A close read of the EDPB's €225M WhatsApp fine. The fine is a catch-22 for technical companies, that you need to write your privacy notice concise and easy to understand enough for laymen, yet need to be comprehensive enough to cover all basis. An extremely bad example of regulatory overreach, lacking resemblance of proportionality
Loom workflow system
A parley sketching "Loom": after parley, issues, and vision, a fourth repo-native tool for recurring agent-driven routines — the morning health check, the weekly oncall report. Together they form a kind of operating system for how a startup actually thinks.
Vision alignment tools
A parley on bridging business and tech vision. Mermaid is great for structure but weak for magnitude — Wardley Maps and Graphviz with weighted nodes do better. The deeper move: stop being the human translator and become the maintainer of a shared YAML schema.
what is "Graphviz (DOT language)"
A side-trip into Graphviz's DOT language: a declarative graph format from AT&T Labs where node width, color, and font size let you encode magnitude — exactly the dimension Mermaid lacks. The recommended pipeline: YAML domain model, Python script, DOT, SVG.
AI Workflow Evolved, One Year Jouney
Reflecting on a year of tinkering and adapting to the ever changing world of AI. After a year and 400+ commits on Parley.nvim, I'm convinced the future is humans setting up the environment, with full automation of the coding inside it. Here's my journey so far.
Conversation: further on the ideal of log.md
A follow-up parley: log.md is the least controllable file in an agentic issue structure because it depends on the model voluntarily writing during execution. Drop it. Let git history and plan.md diffs serve as the log instead.
Conversation: agentic state management
A parley conversation on why agentic coding state — issues, plans, logs — wants to live inside the repo, beside the code. Single-threaded topological execution as the sane default; minimal files per issue beat reinventing Jira on disk.
Personalized Software
When agentic coding lets a vim user write a custom mouse-aware floating picker because Telescope couldn't do that. The implication for SaaS: programmability stops being a nice-to-have, it becomes the moat.
My AI Coding Workflow
An iterative spec flow built on a small Makefile and a handful of markdown files — issue, todo, lessons, specs — cuts through the opaque magic of workflow tools. The trick is treating specs as sketches that bound the agent's search space, not exact behavior to confine it. Let AI bringing in world's knowledge!
The Week Everything Changed
The week of February 23, 2026: Dorsey's layoffs, Cloudflare's one-engineer Next.js rewrite, Karpathy's tweet. Picking up Conductor, OpenCode, and OpenSpec over a weekend and shipping 15K lines on Parley — agentic coding is no longer the toy it was ten months ago.
AI as the Next Abstraction Layer
LLMs are fundamentally stochastic, but that's not new — the whole ML field always was, and it worked. The interesting question is how to ground stochastic computation in deterministic domains. Two paths: build a deterministic shell around the model (coding), or apply it directly where the domain itself tolerates noise (recruiting, marketing, design).
To Type or Not to Type
Mainstream type systems are simultaneously too strong (closed-world assumptions hurt during change) and too weak (can't say "positive integer" or "age between 0 and 200"). For startups in unsettled domains, duck typing wins — types are a luxury you afford after product-market fit.
Data vs. Anecdotes: Which to Trust?
Two quotes about data and intuition. "I trust your data, and my intuition" subtly inverts the knowledge hierarchy and subjugates domain expertise to authority. Bezos's version — when anecdotes disagree with data, investigate the measurement — is the better instinct.
On Organic Process
Eight questions to ask before introducing a new process. People overestimate how much process helps, underestimate the cost of maintaining it, and assume it'll work just because it exists. And remember, processes rarely die on their own — the ritual often outlives the purpose.
When Judgment Slips
Judgment slips when decisions become emotional — when you're defending a person, a past call, or a principle rather than solving the problem in front of you. Four signals to watch for in yourself, and four ways to guard against them.
Supreme Court on Trump Tariff
Notes from the IEEPA tariff oral arguments. "Regulate importation" can plausibly mean tariff — but the power to tax is squarely Congress's, the Major Questions and Non-Delegation Doctrines bite, and the one-way ratchet of presidential power makes this near-unanimous against the administration.
How to Parli
Notes from judging a high school parliamentary debate. The opposing side isn't an adversary, it's a collaborator in a joint quest for truth — and the kitchen-sink approach loses on me every time. Logos 70%, ethos 20%, pathos 10%.
Information as Leverage
Andrew Grove treats management as a production system, but the deeper claim is that information workers are managers — their raw material is information, their output is influence. In modern Big Tech, being "a manager" isn't a promotion. It's just different work.
Birthright Citizenship V2
Read narrowly, Wong Kim Ark's actual holding only covers children of permanent legal residents. For 127 years, every branch followed the broader dicta instead — until now. A close reading of holding versus dicta, reliance interest, and what the Court is likely to do.
This is the Way
Healthcare's fragmentation isn't an accident — it's federalism plus misaligned incentives, where EHR vendors profit by locking customers in. NexHealth's Synchronizer attacks the problem at the technical layer, with the law (and a recent 4th Circuit ruling) on its side, eventually.
My Operation Principles
My operating principles: be serious but not too serious, be curious and build a shared mind, be radically transparent, move fast and break things.
How WhatsApp Works
How WhatsApp actually worked: Erlang/BEAM/OTP all the way down, the biggest machines money could buy, location-transparent actors, and code hotloading as deployment. They bucked every Cloud Computing canon and ran the world's largest messaging system with a tiny team.
Designing a Fairer Electoral System
California's non-partisan commission produced a worse minority seat-to-vote ratio than Texas's partisan one in 2024. The deeper culprit isn't gerrymandering — it's winner-takes-all single-seat districts. Multi-seat districts with ranked-choice voting would do real work here.