Agentic Experience: The Agent's Error Log is the Blueprint for Your CLI
An AI agent made over 500 errors using a CLI across 165 sessions. Its error log serves as a reverse blueprint for the interface you should have designed.
frr.dev
Recent content on frr.dev
Latest posts
Cloudflare's "Ask AI" created an API token with read access to my entire account
Cloudflare's in-dashboard AI assistant silently provisioned a permanent, read-everything API token across all my accounts, zones and users.
Sentry Frontend Blocked by Brave: How to Fix it with a Tunnel
Brave Shields blocks Sentry even with shields down. Learn how to fix it with your own Flask tunnel and avoid three undocumented gotchas.
Agentic Experience: 1,324 calls to my CLI, 15.9% error rate
I analyzed 1,324 real calls from Claude Code to my Linear CLI. 15.9% failed. The errors reveal how to design tools that agents can use without breaking.
The Two-Tier Pattern in Apple Silicon: Cheap Deterministic Code First, CoreML for the Heavy Lifting
The same architectural pattern solves sentiment analysis and image cleanup: deterministic first, CoreML only for what remains. PyTorch → ONNX → Apple Neural Engine in three commands.
The Adversarial Collisions that Broke Apple’s CSAM System (and Why They Matter Beyond Apple)
In August 2021, researchers broke NeuralHash in less than two weeks. The attack wasn’t a fluke—it’s a structural property of the perceptual hash family. The EU is legislating on this.
64 Bits Decide What You Can Upload to the Internet: The Algorithm Behind Chat Control
The EU Chat Control and Apple’s failed on-device scanning rely on the same 40-line Python algorithm. Understanding how it works is essential for informed discussion.
Your CLI Has a New User—and It’s Not Human
Agent Experience (AX) principles are changing how CLIs are designed when the primary user is an LLM. Output that’s token-efficient, fuzzy matching, actionable errors, and zero mandatory flags.
Apple's On-Device Model is Terrible for Chat But Surprisingly Good at Structured Output and Tool Calling
I built a macOS app to stress-test every capability of Apple Intelligence on-device. The 3B model is mediocre at free-form text, but with constrained decoding and tool calling, the results are surprisingly good.
Apple's sentiment analysis thinks 'delete the temp file' is a death threat
Apple's NLTagger scores 'delete the temp file' as -0.8 (very negative). It's systematic bias in technical text that nobody has documented. Data, causes, and an open source solution.
Bridging async Swift to sync C: Four Functions to Use Apple's LLM from Any Language
libfoundationmodels exposes Apple's on-device LLM as a 4-function C dylib. The async Swift → sync C bridge using Mutex and DispatchSemaphore, and why this is more useful than a local HTTP server.
I'm paying $15 per million tokens to write 'fix: typo'
Coding agents use $15/M token cloud models for commit messages. foundation-hooks replaces them with Apple's on-device LLM: 300ms, 0 tokens, 0 cost. Open source repo with make install.
TurboQuant, one month later: implementations, controversy, and what actually works
Google's TurboQuant paper got 575 points on HN and crashed memory stocks. One month later: which implementations actually work, the RaBitQ controversy Google won't address, and how to run it on your Mac today.
Your Mac Has a Free LLM and You're Not Using It
Apple Intelligence includes a 3B-parameter model on your Mac. Learn how to use it for classification, naming, triage, and commit messages without paying for APIs. Copy-pasteable Swift recipes.
Three Agents Walk into a Bar: My Experiment to Code More and Spend Less
My plan to distribute work among Claude, Codex, and GLM/Z.AI without automating too early, cutting costs without losing judgment.
In Codex, a Skill Is Not a /Command (but in Claude Code, It Almost Is)
If you're coming from Claude Code, this will save you some serious frustration: in Codex, a skill is not a slash command. They're used for different purposes.
NLTagger and Sentiment Analysis: Why Apple Thinks Your Code is Depressing
Apple's NLTagger scores perfectly neutral technical text as negative. An empirical analysis of bias, why it happens, and what to use instead.
Claude in the Morning, Codex in the Afternoon: The Two-Agent Workflow You Didn't Know You Needed
165 sessions with Claude Code and 27 with Codex CLI reveal a clear pattern: each agent excels in different tasks. Learn how to organize your day with two AI copilots.
The Best Thing That's Happened to Python in Decades Is Written in Rust
Python has had a fragmented and slow tooling ecosystem for years. The revolution came from where nobody expected: Rust. uv, Ruff and ty have changed the game. The Cult of Ferris was right.
150 Lines of Apologies Removed
My Linear skill had 150 lines of workarounds: hardcoded UUIDs, curl fallbacks, notes like 'the CLI doesn’t support X.' I didn’t rewrite them — I made them unnecessary by building a tolerant tool.
Madness Driven Design: Don Quixote, Sancho Panza, and Your AI Copilot
MDD: a methodology for designing tools that absorb the hallucinations of your AI copilot. Two layers—error archaeology and adversarial verification—inspired by Don Quixote and Sancho Panza.
Why My CLI Output Isn't XML (And How I Ended Up Reinventing TOON Without Knowing It)
XML repeats tags, JSON repeats keys. When your consumer is an LLM, every token counts. The story of how I ended up with a compact positional format and discovered it already existed: TOON.
Adversarial Programming: When Your AI Copilot Invents APIs
Your AI generates plausible code against non-existent APIs. Schema introspection, real fixtures, and fetch/process separation: the anti-hallucination methodology that works.
Linear Agent Isn’t What You Need. Your Agent Was Already in the Terminal
Linear launched their AI agent. Sounds great, but what we really needed was a functional CLI that didn’t suck. And if we have to rewrite it, let’s do it in Rust.
Transform and Conquer: How Google Compresses LLMs 6x by Changing Coordinates
TurboQuant compresses LLM KV cache 6x without losing precision. The trick: don't invent a better compressor, move data to polar coordinates where compression becomes trivial.
Your AI Writes Code That Compiles but Means Nothing (and a Linter Can Catch It)
LLMs generate code that compiles but abuses strings, optionals, and nil where specific types should exist. How to detect it with a linter and an auditing skill.
Your plan.md Needs a Devil's Advocate (and Codex Volunteers for the Job)
A developer got Claude and Codex to argue about their plan until it was solid. The idea is brilliant: use a second model as an adversarial reviewer. 14 bugs caught, zero human intervention.
Mole: The Mac Cleanup Tool You Didn’t Know You Needed (But Only Two Commands Matter)
Mole promises to clean and optimize your Mac. After testing it all, only two commands are worth your time: mo purge and mo analyze. The rest is just fluff.
Working with an AI agent is like living with the protagonist of Memento
Your AI agent doesn't remember where your projects are. zoxide and CLAUDE.md are the solution: a smart cd and a memory tattoo so they stop getting lost.
A Kalman Filter to Stop Bothering the Server (Or the Guilty Pleasure of Over-Engineering)
How a scalar Kalman filter reduces network calls by 80% in a menu bar app, adapts to battery status, and sleeps on its own when you're inactive.
My Setup: Claude Code + Ghostty + Worktrees on a Mac
How I combine Ghostty, git worktrees, and Claude Code in my daily macOS development workflow. No flickers, massive scrollback, and true parallelism with worktrees.
Codex CLI Lacks Worktrees (Here's How to Set Them Up Yourself)
Codex CLI doesn’t support --worktree or --tmux. But with manual Git worktrees and a few tricks, you can achieve real parallelism. A practical guide with the bugs you’ll encounter.
Five Game-Changing Worktree Tricks in Claude Code to Transform Your Workflow
Claude Code has native support for git worktrees: --worktree, isolated sub-agents, custom agents with isolation, --tmux for parallel sessions. Five tips from Boris Cherny.
The Week When Boring Discipline Beat Magic
PostgreSQL without sharding, a while loop without architecture, cron without cloud platforms. Six articles saying the same thing: boring discipline works better than flashy technology.
Codex CLI Full-Auto Mode: Two Flags to Stop the Approval Prompts
Tired of Codex CLI asking permission for every file? Use --full-auto or -a never -s workspace-write for autonomous agent mode. Step-by-step guide with Claude Code comparison.
Context engineering: the invisible skill that separates good AI agents from mediocre ones
Prompt engineering is writing a good prompt. Context engineering is designing EVERYTHING the model sees: what goes in, in what order, what gets discarded, what gets compressed. And it's what really matters.
I've Declared Email Bankruptcy (Again), but This Time I Have a Plan
Email is an inbox anyone can fill with junk. I've declared bankruptcy multiple times. This time, instead of promising to manage my inbox better, I’ve decided to stop using it as a reminder system.
Five Nonexistent Experts Review Your Startup Before You Build It
I designed a council of five simulated LLM experts to evaluate MVPs before writing code. Paul Graham, Lessig, Godin, Balaji, DHH — and why each one is part of the team.
DIY Codex Automations: Nocturnal Agents with Claude Code and systemd
A practical tutorial to replicate OpenAI's Codex Automations using Claude Code, systemd timers, and Gitea. Agents that work while you sleep, without relying on any desktop app.
OpenAI scales PostgreSQL for 800 million users with a single writer (no sharding)
OpenAI serves 800M ChatGPT users with a single PostgreSQL primary and ~50 replicas. No sharding, no microservices. Deliberate simplicity beats over-engineering.
Your AI coding agent is a while loop with delusions of grandeur
Codex CLI and Claude Code look like magic, but inside they're just a loop with 5 phases, an LLM, and tools. We break down the agentic loop step by step with diagrams.
A 2,500-Layer Neural Network That Turned Out to Be MD5: What This Teaches About Debugging
Jane Street hid MD5 inside a neural network with integer weights. The process to uncover it is a masterclass in debugging that every senior should study.
Your LLM's Cache Charges You Double to Save You Money (And It Makes Sense)
Prompt caching reduces input cost by 90%, but cache writes cost 25% extra. Your bill goes up before it goes down. Here's how it works and how to manage your expenses.
From /simplify to the Jedi Council: How I Built a Code Review with Kent Beck, Martin Fowler, and Mike Acton
/simplify by Claude Code reviews your code with three generic agents. It's okay, but it creates noise. I replaced it with a Jedi Council: Kent Beck, Martin Fowler, and Mike Acton with real decision rules. Fewer findings, zero false positives, and one bug the other tool missed.
/loop in Claude Code: the cron that lives and dies with your terminal
Claude Code 2.1.71 introduces /loop, an inline scheduler for recurring tasks. What it does, how to use it, and why it doesn’t quite replace a proper cron job.
33,000 lines of XML to tell you heavyWork() is slow: how I tamed xctrace for LLMs
xctrace exports 33,000-line XML files that blow up any LLM's context window. ztrace condenses them into 10 actionable lines. How and why.
The Wrong Path Must Be Impossible, Not Forbidden
When an AI agent operates your ETL pipeline, forbidding things doesn't work. Only making the wrong path structurally impossible works.
Anatomy of Claude Code's Remote Control: The Hidden API You Can't Use Yet
Anthropic announced Remote Control for Claude Code. We investigated the bridge API, WebSockets, on-disk traces, and why it still doesn't work for all Max accounts in February 2026.
RustyClaw: I'm rewriting an AI agent in Rust (because the meme demands it)
I’m porting 8,300 lines of Python to Rust using LLMs as copilots. The real goal: testing adversarial development in a hardcore porting process. With raw data and a hallucination counter.
My AI Read a JSON File from Disk 900 Times in a Loop (And Why No Linter Can Save You)
An LLM wrote code that read and parsed a JSON file from disk on every iteration of a 900-file loop. It's a first-year computer science mistake. No linter can catch it.