Stephen Chuang 莊宗憲

AI-Native Systems Builder

Agentic Systems Engineer13+ years of software engineering

I turn ambiguous problems into concrete systems and production outcomes.My focus is the boundary between AI agents and deterministic software delivery.

See what I buildContact Me
Notes

What I'm thinking about

Short notes on what I'm learning, building and getting wrong. When one needs the full argument, it links to an article.

  3. Stephen Chuang 莊宗憲

    Shipped

    Deckhand・Turn Claude into Your Multi-Agent Mission Control

    (Custom Claude Code Mods)

    Adds a toolbar right above the input box: monitor usage, switch models in a click, run parallel subagents with Sub5, and call Codex, Cursor agent, and agy read-only for second opinions.

    Claude stays in charge as the primary agent; all other AIs are strictly read-only.

    Installs in two commands, MIT licensed.

    ArticleIntroducing Deckhand: Turn Claude Code into Your Multi-Agent Mission Control6 min readRead the full article
All notesAll articles

A delivery gate for AI coding agents

Better Workflows: Evidence first, then it touches your repo

An agent saying it is done is a claim, not a result. Better Workflows sizes up the risk before anything is written: small edits go straight through, anything larger has to produce evidence that matches the current commit. Git changes run in a worktree owned by that task, and the run stops when the state is unknown.

Goal-first
The outcome is agreed before the route is chosen.
Evidence-driven
Every claim binds to the source it was checked against.
Fail-closed
Drift, stale evidence or unknown state halts the run.

It catches stale evidence, false completion claims and external actions with unknown results. It has not yet been statistically shown to reduce drift or rework over long-running work.

betterworkflows.devView @better-workflows/core on npm

V6 Core alpha · Apache-2.0

# gate-demo · illustrative, not a live run

  1. 01goalpassed
  2. 02sourcepassed
  3. 03worktreepassed
  4. 04executepassed
  5. 05verifypassed
  6. 06authoritypassed
  7. 07actpassed
  8. 08reconcileoutcome unknown
  9. 09completenot reached

STOPPED AT 08 / 09

The provider result is unknown: no retry, not marked done. It stops and waits for you.

npm install -g @better-workflows/core@alpha
claude plugin marketplace add stephen-taipei/better-workflows
claude plugin install better-workflows-core@better-workflows --scope local
bw init

6.0.0-alpha.2: Claude Code first; Codex, Gemini CLI and Qwen Code are planned. Needs Node.js 22.14+.

V5 is frozen at RC1 (AGPL-3.0-only) and no longer developed.

See the full walkthrough
More agent tooling
Deckhanda multi-AI toolbar for Claude CodeStephen SkillsAgent Skills shared by Claude Code and Codex

AI Relationship Coach · Client & Contact Follow-Ups

Connectors: AI Personal CRM: Turn Contacts Into Relationships

Forgot what you talked about last time? Connectors remembers what each person cares about and reminds you when it is time to reach out.

AI coach
Suggests who to contact and what to pick up on.
Reminders
Nudges you before a relationship goes quiet.
Pipeline
Shows where each conversation stands.
ctrs.appApp Store
Connectors: AI Personal CRM
Connectors: AI Personal CRM
Connectors: AI Personal CRM

Product & architecture lead · Shipped on the App Store

Live on iOS and the web · Android in development

Stack: React, NestJS, tRPC, SwiftUI, Kotlin Compose (AI-assisted implementation)

Also in developmentAwesome APP 100a suite of native iOS and Android apps

Case study

One Frontend, 100+ Themes

Re-architecting a production Nx/Angular platform. I was the architecture owner and migration lead.

  • core / view / theme leaf layers
  • Typed theme contracts
  • Family completeness tests
  • On-demand theme CSS
2.4 MB→17 KB
CSS per theme request. 136 theme stylesheets went from one bundle to on-demand loading, about 99% less.
Hundreds→23 lines
One representative, complete theme leaf, before and after the refactor. Not every leaf is 23 lines.
50+→90+
Lighthouse Performance on the key measured pages. It came from several changes together, not the CSS alone.

All three are production measurements. The public repo rebuilds the architecture and loading mechanism with original code and three fictional themes; it does not reproduce these numbers. Evidence boundaries and source

Read the full articleLive demo
Stephen Chuang, AI-native systems builder and agentic systems engineer

From a design background to system architecture

I am drawn to learning across fields. I started in design, moved into information management, and settled on full-stack development and system architecture. In recent years my focus has moved to the boundary between AI agents and software delivery: making an agent prove its work before it touches production.

Outside of work, I founded a street dance studio and achieved national champion and world top 8 in dance competitions. These experiences cultivated perseverance, focus, and determination to overcome challenges, which extends to my work attitude.

years of software engineering
13+
frontend engineers I lead directly today
3–4
members in one year, on a customer product I helped deliver
0 → 20,000
Professional Journey · Technical Expertise

Let's Connect

I am glad to talk about agentic systems, AI agent reliability, developer infrastructure and open source collaboration. Questions about anything on this site are welcome too.

Location

Taipei, Taiwan · Remote