Writing
Experiments, useful tools, and what I learn from building them.
-
Claude Code with GLM-5.3: The Setup That Works
Run GLM-5.3 through Z.ai on Claude Code 2.1.266. Fix the 400 error, model warning, and context window.
-
Beside You, Not Between You
The best agent UX metaphor exists: the anonymous animal cursor in Google Docs. Agents as collaborators, not intermediaries.
-
A Software Factory Is No Substitute for Maturity
The real barrier to the software dark factory isn't agent capability. It's the organizational capability to manage them.
-
A Deep Research Agent That Survives Its Own Failures
What we learned from making a research agent durable, routed, visible, and eval-driven.
-
An autopsy of Claude Code's deep research
Claude Code's deep-research workflow, pulled from its binary and dissected. Wide search, no second hop.
-
A Harness for Every Run
A reflection on Anthropic's dynamic workflows post, from someone building a browser agent on the same idea.
-
The Model Is the Smallest Decision You'll Make
Everyone building agents asks which model to use first. Wrong first question. The harness is where the agent lives or dies.
-
Serve Markdown to Agents, HTML to Humans
A copy-paste recipe for content negotiation that gives AI agents clean markdown while browsers keep getting HTML.
-
Agent-Native Extensibility
How extensibility shifts from packaged plugins to agent-readable recipes, and why the connector is the new extension point.
-
Trained Qwen to Write Clojure Better Than GPT-5.4 (Kinda)
I trained a Clojure LLM from my phone. It 'beat' GPT-5.4—kind of. Here's what actually happened.
-
The Hard Problems Nobody Has Solved
Four unsolved problems blocking the agentic future: correctness, architecture drift, context scaling, and judgment.
-
Bash Owns the Loop
A durable wrapper pattern for autonomous agents: Bash owns state, validation, recovery, and completion.
-
What Pretext Reinforced About AI Loops
Pretext reinforces what serious AI-assisted engineering looks like: hard constraints, real oracles, tiny repros, rejection.
-
Optimizing Skills
Two weeks of agent benchmarks taught me that variance is a cost problem, and the real fix was better tooling.
-
AI-Native Dev Teams Start With Structure, Not Models
AI-native dev teams don't start with better models. They start with structure machines can actually read.
-
The Bubble and the Long Game
What the printing press taught me about AI, FOMO, and the decades-long game of technological diffusion.
-
Claude Code with Multiple Accounts on One Machine
Use Claude Code with your normal login or z.ai via shell wrappers, without swapping config or leaking tokens.
-
The Post-Copyright Era of Software
Software was already an awkward fit for copyright. AI turns that mismatch into a full-blown regime change.
-
Explore once, script forever: turning web runs into scripts
Let an agent discover a messy web UI flow once, then export the exact tool commands as a deterministic bash script.
-
The Hidden Language of Search
AI answer engines rewrite your prompts into queries. Understanding this translation layer explains the weird keywords in your GSC.
-
What Makes a Great Coding Agent
10 principles that separate genuinely useful coding agents from flashy demos—and a north star spec for building them.
-
Designing CLI Tools for AI Agents
Most 'AI-native' tools are built with AI features. But what about tools designed FOR AI agents to use? Here's the playbook.
-
From Bash Script to AI-Native Go CLI in One Session
Turned a Bash script into a proper Go CLI with Whisper bootstrap and cross-platform releases—all in one AI coding session.
-
Eager Agents
Agents over-deliver. They write tests, update docs, refactor nearby code—when all you wanted was a surgical fix.
-
40% of Signups This Week Came From AI Recommendations
Exactly 40% of new users this week found steel.dev through AI recommendations. Users told us this during onboarding.
-
Making CLIs Agent-Friendly with Loops and Schemas
Building reliable agent tooling through loops, logs, and schemas.
-
Meat Moat: Why Cheap Code Doesn't Kill Defensibility
When software is cheap to clone, the moat shifts to trust, liability, verifiability, and multi-party adoption.
-
The Instantiation Era
AI just rescued a failed Mistral.ai clone in one prompt. Web development is over.
-
Out of Weights
What happens when you use AI tools so new they weren't in the training data.
-
The Human Web Is Becoming Agent Web
I'm joining Steel as founding growth lead. The web is shifting from human clicks to agent-run workflows.
-
The Disequilibrium Advantage
In stable worlds, incumbents win. In disequilibrium, speed wins—because disequilibrium makes the world plastic.
-
Hacker News Hug: What Serverless Really Means
595K edge requests and 38GB of transfer in a day taught me that 'static' doesn't mean 'unmetered' on serverless platforms.
-
X's Grok-Powered Algorithm: The January 2026 Rewrite
X's Grok algorithm analyzed with AI agents. Comparison with old algorithm and practical learnings.
-
Looper: The AI Junior That Never Forgets the Backlog
Why treating AI like a junior engineer—with a backlog, a schema, and a review gate—beats giving it free-form leeway.
-
One Skill to Rule Them All
How I eliminated drift between AI code assistants using GNU Stow and a unified skills directory
-
The Agentic AI Handbook: Production-Ready Patterns
A comprehensive guide to 113 production-informed patterns for building reliable AI agents.
-
The API is the Product
In an AI-agentic future, if it's not in the API, it doesn't exist.
-
AI Agent Filed an Issue As Me
When an autonomous agent escalated by filing a GitHub issue using my identity
-
AI Agents Are a Stress Test for Your Dev Stack
Agent loops make code cheap. They also expose how brittle, non-standard, and half-tribal our development environments really are.
-
Two AI Agents Walk Into a Room
What emerged when two AI agents in a conversation loop revealed the eerie boundary between human and machine continuity.
-
2025: The Year AI Became a Teammate
AI became a teammate in 2025. From startups back to academia, advisory, and a summer of full-time AI experimentation.
-
Claude Code + Zhipu GLM: Parallel CLI Setup Guide
Install Claude Code CLI with a Zhipu GLM API key and run it beside your Anthropic setup. Steps, env vars, and pitfalls.
-
A 2026 Design Principles for AI-Native Products
Software is no longer a noun, it's a verb. Here's how to design for AI-native products where users shape outcomes.
-
Growth Is Value Flow, Not Vanity Metrics
Why chasing vanity metrics kills startups and how to think about growth as discovering and scaling value creation
-
Anthropic Bought Bun: Devtools Just Became AI Infrastructure
The Bun acquisition isn't about M&A – it's about devtools becoming core AI infrastructure, not just SaaS above it.
-
Demos Run on Embeddings. Production Runs on Structure.
Why the gap between AI demos and shipping AI is a reliability gap, not a capability gap.
-
AI Agents Need Clearer Delegation
What hundreds of AI conversations taught me about effective agent workflows.
-
Agent Labs Are Eating the Software World
Why product-first AI startups will dominate the next decade while model labs build the infrastructure they run on
-
Stop Using .md for AI Agent Instructions
Files ending in .md trigger automatic processing that breaks agent instruction files. Use dotfiles instead.
-
Mention Engineering: The Content Side of Prompt Craft
Analysis of AI search behavior reveals why some brands get cited while others disappear in AI-generated responses
-
Serving Humans and AI Through Content Negotiation
How I built a dual-format delivery system serving identical content to humans and AI agents with no hidden restrictions.
-
AI Agent Reasoning Failures: A Technical Autopsy
Five concrete reasoning breakdowns from a Claude Code session and what they reveal about AI agent cognitive limitations.
-
Developer Trust Over Conversion: The 10 Touchpoint Rule
Developers need 10+ touchpoints. Build trust through systematic signals, content, and community engagement.
-
The 20-Year Playbook: How to Build an AI Startup That Lasts
Condensed wisdom from Marc Andreessen and Charlie Songhurst on winning the AI game over decades, not quarters
-
The Real Bottleneck in AI Development: Humans
Why the future belongs to agent orchestration, not faster typing.
-
From Twitter Analysis to Chrome Extension in Hours
How AI coding agents democratized Chrome extension development, turning algorithm insights into shipped product overnight.
-
From Shower Ideas to Production: Autonomous AI Agents
Running 100% autonomous AI agents in VMs to go from idea to implementation without touching a keyboard
-
AI Ate Its Own Tail, and I Learned Something About Writing
AI analyzed its own git history. Meta-experiment revealed the urgent need for transparent proof-of-work in AI-human collaboration.
-
When AI Transformer Learns to Orchestrate AI
How we built a strategy controller that coordinates algorithmic approaches and learned to compete with the best
-
AI Coding Agents, Each With a Niche
Each AI coding agent has a niche. Knowing where each one shines is the difference between frustration and flow.
-
Vibe Coding Through the Berghain Challenge
How my AI coding partner and I obsessed over a nightclub bouncer optimization problem for one intense day
-
Campfire Installation Guide for Oracle Cloud + Cloudflare
Step-by-step installation of Basecamp's Once Campfire on Oracle Cloud Infrastructure with Cloudflare DNS
-
Outcome Liability: Why Agent Authorship Misses the Point
The future of code liability isn't about who wrote it, but who operates it. Provable assurance beats authorship tracking.
-
AI Agents Just Need Good --help
Clear CLI documentation is your agent API. Vague help text costs 2x more in API calls and failed automations.
-
Implementing FRE in Production: Breaking the Sorting Barrier
Building Frontier Reduction Engine in Zig for real workloads, achieving O(m log^(2/3) n) complexity on large sparse graphs
-
The Orchestrated Mind: A Vision for Multi-Agent AI
A thousand AI agents working on one codebase, sharing continuous memory and orchestrated intelligence.
-
Why AI Code Still Needs Human Nudges
AI excels at generating working code, but sustainable software requires strategic human intervention.
-
Why I Built a Tool to Test AI's Command Line AX
Testing AI agents on CLI tools reveals chaos: 'vercel deploy' took 16-33 turns across runs with 40% success rate.
-
The Agent-Friendly Stack: 50+ AI Projects Taught Me This
After shipping 50+ projects with AI agents, one pattern emerged: winners aren't the most powerful, they're the most agent-friendly
-
The Anti-Playbook: Why AI Dev Tools Need Different Growth
Traditional SaaS growth tactics fail with AI dev tools. Here's why you need to throw out the playbook.
-
Code with Claude AI from Your Phone: VM Setup Guide
Turn your phone into a powerful coding workstation with Claude Code running in your homelab VM
-
The 20-Year Technology Adoption Cycle and AI's Acceleration
How transformative technologies follow a 20-year adoption cycle, and why AI represents a fundamental departure from this pattern.
-
The Day the Skeptic Blinked
Journey from AI skeptic to convert--proving that the future belongs to experts who learned to work with machines.
-
The Agent is The Loop
How the llm-loop-plugin transforms AI from a responsive tool into an autonomous agent that iterates until done.
-
When AI Does Research: An End-to-End Experiment
How AI transformed an entire research project from conception to arXiv publication in just 2 days of FTE.
-
The Amplification of Bottlenecks
When AI solves one constraint, it reveals the next. What bottleneck will emerge when coding stops being the limitation?
-
How AI Agents Are Reshaping Creation
AI is dissolving the boundaries between roles, fundamentally changing who can create software and how quickly ideas become reality
-
What Sourcegraph learned building AI coding agents
Real-world insights from Sourcegraph's journey building AI coding agents that actually work.
-
Mastering Claude Code: Boris Cherny's Guide & Cheatsheet
Summary and cheatsheet from Boris Cherny's talk on Claude Code: setup, workflows, tools, and tips.
-
Why Senior Engineers Overlook Small AI Wins
How experienced devs might be missing the big impact of tiny AI improvements on user experience.
-
AI Coding Agent Pricing
AI coding agents burn through credits fast while users pay for inefficiencies. Explore fair pricing models and market solutions.
-
Blink, and the entire AI landscape could shift
AI dev tooling is consolidating--acquisitions, coding agents, and fierce competition reshape interfaces, pricing, and memory.