Shipping with Claude Code

In Agentic AI Thoughts I wrote that memory orchestration is the technical moat. Now I've built it with Claude Code — MBTI long-term memory + Deep Companion, and what happens when AI actually remembers what you said.

After years of commercial VPN services, I decided to self-host. One afternoon, one VPS, Claude Code doing all the work. Shadowsocks got blocked within hours, the cert saga had three false starts, but what came out the other side was more complete than I expected.

My GCP devbox died twice in one day. First time: a snap package triggered an infinite crash loop. Applied five fixes, thought it was stable. Hours later it died again — unbounded memory was the real issue. Two rounds of diagnosis, 12 hardening measures, and a machine type swap. A complete field report on VM defense in depth.

Claude Code's Agent Team feature let me break a complex cross-layer plan into ten tasks and assign them to five AI teammates working in parallel. The Lead coordinates dependencies, dispatches resources, handles conflicts — the whole thing feels like managing a remote engineering team.

After building Mio v0.0.2 with five parallel AI agents, I used the same Agent Team pattern for code review — four specialized reviewers found 2 BLOCKs and 2 WARNINGs, three fixers patched everything in parallel with strict file ownership, and two changelog agents handled the release. Building and reviewing are two halves of the same workflow.

A single Claude Code session that combined direct pair programming (three production deploys from user feedback), a background docs agent running concurrently, and an autonomous review team iterating three rounds until clean — all overlapping. The traditional code-test-review-release pipeline collapsed into parallel streams.

The 32GB GCP devbox — upgraded and hardened after Part 3 — died again. Same symptom: SSH timeout, kernel alive, userspace frozen. The watchdog caught it: five node processes at 2-2.7GB each, totaling ~13GB. I assumed Claude Code. It was Cursor's remote server leaking memory for 15 hours straight.

Our Chinese blog articles read like translated English — stiff, academic, lifeless. A reader noticed. We spent a week researching what makes Chinese 公众号 articles work, built a 435-line style guide, and rebuilt the entire writing pipeline. Here's what we learned about the gap between 'correct Chinese' and 'native Chinese.'

From discovering pretext to shipping 7 interactive effects across two sites in one session — explosion physics, bouncing bullets, and an animated crab named Clawd.

Claude Code shipped a virtual pet Easter egg. Two accounts, two penguins — the second one downgraded from Rare to Common. So I used Claude to brute-force a Shiny Legendary Dragon in 0.2 seconds.

After seeing Trae's MTC mode in action, I built a /mtc skill for Claude Code. 6 phases, no skipping, think first, code second. Better results than expected on my game project.

Vanilla Claude Code. Then GSD's multi-agent ceremony. Then superpowers' skill system. Then a custom orchestrator on top. Each rewrite came from a specific session that broke. What I learned: solo dev harnesses and team harnesses are different species, and the right answer isn't a default engine — it's routing by work shape.

For solo dev and small teams, design exploration is usually N=1 — you go with your first instinct because producing a second variant takes a day. Claude Design collapses that: six distinct aesthetic directions in an afternoon. What I kept from the pipeline, what I dropped, and how `big-task`'s UI profile takes over once a direction is locked.

Same pipeline — Claude Design explores, HANDOFF.md locks, `big-task` ships — but it bends around each product. ÉLAN is greenfield and needs wide exploration before converging. Shichuan has a live app and needs incremental category-by-category migration. Panpanmao is a monorepo with a v2 redesign that runs three parallel brand expressions instead of one. What changes and what doesn't.

I run several Claude accounts and kept hitting the 5-hour wall mid-session. Switching meant logging out. So I built a daemon that hot-swaps the active login under a running session — and the whole thing turned on one non-obvious fact about how macOS stores an auth token.

Last time I taught a few Claude accounts to rotate themselves — but I was still chained to my laptop. This time I cut the last cord: the agents live on an always-on box, and my phone connects in to watch and steer them. Close the lid, leave the house, lie in bed — they keep running. Three open-source pieces each own one layer — herdr the sessions, hcom the messaging, Moshi the phone — plus account auto-rotation as the safety net. The mobile vibe-coding setup that currently runs smoothest for me: how to build it, and the traps I hit.

On launch day I replayed 22 real commits from 11 of my own repos, running Opus 5.5 at medium and high effort side by side. Both finished every task; high took about a third longer and 35% more tokens, but blind judges credited it with catching several bugs the tests couldn't see. Cross-checked against the Artificial Analysis Intelligence Index, I made high my default. Recommendations by usage profile at the end.

Every Codex agent I ran was on Astra/high, and usage burned fast. When GPT-6 Sol and Luna landed, I had seven model and effort combinations run through three rounds on real repos: my own monorepo, cross-project tasks in three work repos, and blind replays of historical PRs. More routes wrote correct code than I expected; far fewer also got the real caller path passing in their own integration runner, and several submissions with green self-tests weren't actually fixed. The final routing table and delegation rules are at the end.