Sign in to launch Copilot/Codex from the palette.
Home
/
Blog
/blog
Resources
/resources
Projects
/projects
About
/about
Riding Handwired in public
/handwired-in-public
I am learning Cheng Lou's Handwired, where a program is a fixed list of FP32 ops. Five widgets on a toy runtime: the ReLU step trick, both branches always run, the host loop, loops as fixed work, and training a threshold.
The Entropy Wall
/entropy-wall
Models can crack random things we never expected. I tested 25 of them, the 2,300 token functions they wrote, and a Math.random attack in the browser, then asked what Cloudflare's Wall of Entropy is for next.
Terraloop
/terraloop
Two tools that run long tasks on Cloudflare without babysitting them: a heartbeat (loops-yaml) that re-prompts the agent, and terrarium running each job and handing back a receipt. Everything else is one markdown file.
If You're Here, Who's Driving?
/whos-driving
Google and Cloudflare are aiming at 2029 for post-quantum migration. A short note about agents, cryptography, and checking who still has the wheel.
desk and living-artifact: Cloudflare-Backed Hardware with OTA
/built-in-reverse
Two tiny Cloudflare-backed devices. desk installs apps with git push. living-artifact updates its own firmware over Wi-Fi with signed releases and A/B OTA rollback.
Worker Loaders as a Place: Sandboxed Dynamic Workers
/worker-loaders
Notes on using Dynamic Workers as short-lived execution spaces with explicit bindings and egress controls.
Constraint Theory: A Quiz on Where Agents and Humans Differ
/constraint-theory
Twelve statements about agents and humans. The quiz asks where surface behavior looks similar and where the analogy breaks.
Correctness: Software Correctness Tools Compared
/correctness
Reference hub for software correctness tools across runtime verification, model checking, verification languages, SMT solvers, theorem provers, and verified artifacts.
Self: Two Tiny Cloudflare Loop Patterns
/self
Minimal repro for long-running loops on Cloudflare: a Durable Object alarm that re-enters the Worker, plus a Workflow draining the same KV queue.
Prompts Are Wishes: Why Executable Checks Beat Instructions
/prompts-are-wishes
A deja gate passed while allowing broken storage behavior. That moved my attention from instruction wording to executable checks.
Taste: Preference vs. Whether a Design Reaches Its Listener
/taste
A designer note about confusing personal preference with whether an interface, explanation, or prompt reaches its listener.
Liquid Primitives: Building on a Stream
/liquid-primitives
Notes from replacing runners, instructions, and implementations while keeping tests and behavior stable.
Ego-Less: Rebuilding Without Breaking
/ego-less
A personal note about receiving critique, watching code get rewritten, and separating the work from the feeling of being judged.
Compaction: What Survives When the Context Window Resets
/compaction
A 75-call GPT-4.1 experiment comparing five context representations across goals, files, conventions, corrections, and prior reasoning.
Effect-First: Benchmarking Four Approaches to Agent-Assisted Effect Code
/effect-first
Five GPT-5.2 tasks with regex judges comparing curated guidance, human documentation, raw source excerpts, and baseline.
27 prompts for engineering, design, product, strategy, and agent workflows.
Public apps where people sign in with their own Cloudflare account and pay for their own inference. OAuth setup, invariants, pitfalls, and a checklist.
The unit is the experiment, not the repo. Falsifiable claims, a proven/unproven ledger, receipts, and two gates before a repo is earned.
Agents plan forwards. Make them build backwards from a North Star eval, through loop checkpoints, with gates as the floor and proofs where logic allows.
Relentless interviewer mode. AI walks every branch of the design tree, resolving decision dependencies one-by-one until shared understanding is reached.
One line that makes any AI skip the safe answer. What's the single most useful thing you could add right now?
Long-horizon thinking. Map where a project goes in 3 months to 10 years — how it evolves, grows, shrinks, or dies.
Future predictions from current code. What novel ideas have emerged since training ended? What first-mover opportunities exist?
10 attack lenses for any codebase: simplify, destroy, audit architecture, hunt performance, paranoid error handling, test mania, docs, security, refactor, and a final verdict.
Identify opportunities to make code more durable, secure, scalable, and maintainable using Effect. Stream, Cache, Timeout, Retry, Observability.
Modular AI-agent reference for Effect-TS. 9 rules, token-budgeted modules, and a fetch strategy that gives agents exactly the context they need.
Mike Acton's 7 questions for solving performance problems. Before optimizing, ask if you can eliminate, reduce, or approximate the work.
This codebase will outlive you. Fight entropy. Leave it better than you found it.
Before you default to safe patterns — show the version that makes someone stop scrolling. Three variants: constraint-based, reference-anchored, anti-pattern.
10 attack lenses for any design: clarity, UX minimalism, usability, IA, performance, accessibility, responsive, consistency, user testing, and system specs.
Design reference for progress and completion that feels premium. Stripe precision meets PS5 boot sequence.
Make the AI prove it understands what you've built before it touches anything. Summary checkpoint.
Switch the AI into maximum compression mode. Sacrifice grammar for density.
Writing that persuades without pressure. Clarity, honest framing, and reader trust.
Give your AI agent Gmail superpowers. Full read/write/search/send via MCP, npm, or browser connect — no Google security audit required.
Write a prd.ts file for Gateproof. Stories with behavior, evidence, scope, and gate alignment.
20 principles for building products. Speed, team culture, scope management.
Prevent feature creep when building software, apps, and AI-powered products. Decision frameworks, scope management, and saying no.
Brutally honest assessment. Does this already exist? Is this actually a problem? Am I overcomplicating something simple?
Simple demonstration of the Ralph loop concept. Practice the workflow manually before automating.
Set up a self-perpetuating RALPH LOOP with GitHub Actions, state files, and guard constraints.
Command prompts for managing AI agents through verification gates. You steer, not code.
Generate PRD with boundaries, failure modes, and gates. Design backwards from constraints.