Teaching Claude Code to review its own PRs
A judge agent, two independent reviewers, and a rule file per repo. What a self-review pipeline catches, what it confidently invents, and the prompt changes that closed the gap.
Writing — the whole archive, failure modes left in.
Notes from actually shipping things: Swift and SwiftUI in production, TypeScript agents on LangGraph, and the agentic-dev harness I keep sharpening around Claude Code. Newest first. Nothing here is a tutorial I haven't run myself.
14 posts
No posts match yet.
2026
9 posts-
-
A worktree-per-agent workflow for parallel development
Give every agent its own git worktree and most coordination problems dissolve. Setup scripts, port allocation, and the cleanup step everyone forgets.
-
LangGraph checkpoints as a debugging superpower
Checkpoints aren't just resumability — they're a time machine. Replaying a misbehaving agent from any node, diffing state between runs, and the checkpointer I ended up writing.
-
SwiftUI's Observation framework, one year in
What @Observable fixed, the re-render traps that survived the migration, and the measurement setup I use before trusting any performance claim.
-
The case for boring test infrastructure
Clever test frameworks rot faster than the code they test. Why I keep choosing flat fixtures, real assertions, and one slow end-to-end suite over anything exciting.
-
Writing an MCP server a model will actually use correctly
Tool descriptions are prompts, and most MCP servers write them like API reference docs. Naming, argument shape, and error strings that steer a model back on task instead of into a retry loop.
-
Evals for agents that edit code
Diffing against a golden patch is the wrong metric — a real edit has many correct shapes. The rubric I settled on, why I run every suite twice, and how much variance to tolerate before calling it a regression.
-
The actor boundary is a design decision, not a compiler chore
Strict concurrency doesn't just want annotations — it wants you to decide what your app's isolation model actually is. Three refactors where the red diagnostics were right and my architecture was wrong.
-
Stop reaching for
as:satisfiesand the shape of your configA type assertion is the point where TypeScript stops helping you. How
satisfiesplus a couple of narrow generics let a config object check itself — and the two places I still cast and say so in a comment.
2025
5 posts-
Snapshot tests I don't regret
Most SwiftUI snapshot suites become a wall of noise nobody reads. Pinning the device matrix, testing states instead of screens, and deleting every test that only ever failed on a system font update.
-
git rerere, or how I stopped re-solving the same conflict
One config flag turns a long-lived branch from a rebase tax into a non-event. What rerere records, the one case where its memory betrays you, and how it changed how long I let a branch live.
-
Structured output beat every prompt trick I tried
Months of prompt tuning to stabilize one agent's output, undone in an afternoon by a Zod schema. Where schema-constrained generation is a free win, and where it quietly costs you reasoning quality.
-
CLAUDE.md is a product surface
Written like documentation, an agent instruction file gets long and ignored. Written like an API — few rules, each testable, each naming the failure it prevents — it starts doing work.
-
Shipping a SwiftUI app with no view-model layer
I cut MVVM out of a real app to find out what would break. Mostly nothing — but the two places it genuinely hurt are worth knowing before you copy the experiment.