Claude Sydney Meetup Recap — Workflows Worth Stealing & A Weekend Build

On Tuesday, April 21, 2026, the 02Ship community gathered at Haymarket HQ (Level 2, 63 Dixon St) for our third Claude Sydney meetup. 50+ builders, engineers, and founders showed up for two longer talks and six community lightning talks — all focused on how people actually use Claude Code at work.

Attendees at the Claude Sydney meetup April 21

Thanks to Haymarket HQ for hosting the space and to everyone who turned up on a rainy Tuesday night.

The event page is here: Claude Sydney Meetup — How People Actually Use Claude Code at Work.

Talk 1: Claude Workflows Worth Stealing

Shaon Diwakar — Founder of Dini Labs, ex-Meta (Instagram/Threads), NewsMaven, Tyro, EY — opened with a rapid-fire tour through 14 Claude workflows that have made it into his daily practice. His framing: these are all "worth stealing" — pick the ones that match your stack.

View Shaon's slide deck →

We'll group the 14 workflows into four clusters so you can scan them fast.

Cluster 1 — Agent infrastructure

InboxAPI — Shaon's own project. Gives every AI agent a personal email address so it can send and receive mail like any other teammate. npm install -g @inboxapi/cli with SMTP exposed via MCP.
Beads — local-first, SQLite-backed task tracker that lets an agent remember its own work. Issues store as JSONL for git collaboration; the agent picks up "ready" work and manages dependencies. Fixes the problem of Claude drifting off plan between sessions.
Tool Gateways (Composio, Merge) — SaaS orchestration with 1,000+ pre-built integrations (Slack, GitHub, Jira, etc.). Managed OAuth, ephemeral sandboxes, and dynamic tool resolution. Skip building MCP servers per integration — buy the platform instead.

Cluster 2 — Agent behavior & knowledge

Agent Skills ecosystem — "Baking engineering culture into agent memory." 20+ production-grade workflows covering TDD, security, and code quality. Shaon quoted the "Beyoncé Rule" — "If you liked the code, you should have put a test on it." Reference: Addy Osmani's Agent Skills on GitHub.
Evals, Evals, Evals — Make the model rate its own work against a rubric before it calls a task done. Structured JSON self-evals documenting what went well, what needed improvement, what's next. Turns a solo dev into a team by enforcing guardrails.
The LLM Wiki Pattern — a compounding knowledge base the agent contributes to. Markdown files in a shared git repo; agents read wiki-first before scanning raw source. Cuts token cost and stops the model from re-deriving the same facts every session. (Nod to Andrej Karpathy.)
Requirements Interrogation — the "Interview me until you're 95% confident" prompt. Flip the dynamic: let Claude ask you the questions before any code is written. Surfaces hidden assumptions stakeholders didn't even know they had. (Addy Osmani again.)

Cluster 3 — Day-to-day dev workflows

Claude Postgres Optimization — point Claude Code at a psql connection (or an anonymized prod replica via MCP). Get slow-query reports, normalization gaps, and schema drift flagged before your users find them.
Headless Browser Ops — Playwright + Chrome DevTools MCP for frontend debugging. Real DOM interaction, network traces, console logs, performance profiling — the agent sees what the browser sees.
Log Detective — read-only access to prod logs + DB schema. Claude correlates a stack trace back to a specific source line, proposes a fix, and verifies against the real failure data. Works with nginx, IIS, Apache. Shaon suggested Opus 4.6 for cost efficiency on long log tails.
Security Auditor — agents review PRs for security and architectural drift. Self-eval against OWASP Top 10 before a commit lands. Web-search enabled so it can flag supply-chain attacks against your dependencies.

Cluster 4 — Meta & ergonomics

Measure AI ROI — build leaderboards comparing AI-assisted output to the human baseline. If you can't show the saving, you can't get organizational buy-in.
Voice Typing — 3x faster than typing. Longer prompts produce better interrogation, which produces better code. Tools: HyperWhisper, Wispr Flow. A decent mic is non-negotiable.

The through-line: give the agent memory, rubrics, and tools — then get out of its way. The 14 workflows are less about clever prompts and more about the scaffolding around them.

Talk 2: How to Build a Feature for Community Engagement

Bob Jiang speaking at the Claude Sydney meetup April 21

Bob Jiang (02Ship) followed with a live walkthrough of a weekend build: the Lightning Talk Voting tool that powers 02ship.com/vote. Five PRs, one prod incident, and a handful of lessons about which AI tool to reach for at each stage.

View the Talk 2 slide deck →

The arc of the build:

Idea → Hermes brainstorm. Before writing code, Bob ran the idea through Hermes (via the gstack office-hours skill). Its job is to kill scope, not help you write. It cut the feature list by ~60% in 10 minutes — dropped ranked-choice voting, Slack integration, and user profiles. The narrowest wedge: "40 people on a phone, 5-minute window, one cookie, three votes."
First PRD with Claude Code + gstack — failed. Asked to "refine the PRD," Claude Code produced 40 pages, 6 services, WebSockets, Redis streams, 3 new tables. Too complex to build or review. Lesson: Claude Code is a coder. If you ask it to architect, it will happily over-architect.
Refine with superpowers + codex review. Switched to the superpowers skill chain — brainstorming → writing-spec → writing-plans — which locks intent before touching implementation. Got a 6-page PRD, 1 KV store, honest edge-case list. Then ran the PRD through codex review (OpenAI's Codex CLI in adversarial mode) for a second pair of eyes. Codex caught three real gaps: "no auth for approvers," "rate limit is per-cookie — trivial to bypass," "what happens when a talk is rejected after a vote?"
Write plan → execute in a worktree. Plan as a checklist of commits, each small enough to review in one sitting. Git worktree keeps the main checkout clean. The verification-before-completion skill requires evidence (test output), not vibes, to close a step.
Code review + E2E — and an E2E trap. Three-lens review: Claude self-review (simplify skill), codex review (adversarial), then a human on the diff. Plus a Playwright E2E covering submit → approve → vote → verify. The trap: the E2E ran against the shared prod KV store. Every CI run left behind "E2E Speaker 1776673201567" records, drowning out real talks. Eventually deleted the test (PR #69). Rule: E2E against shared prod needs teardown — or kill the test.
The deploy cliff. Localhost all green → production returns 500 on every submit. The client showed "Please fill in name, title, intro..." even when every field was filled. Root cause: Vercel KV had been deprecated; the store migrated to Upstash, which provisions different env var names. @vercel/kv reads KV_REST_API_URL; Upstash provides UPSTASH_REDIS_REST_URL. First kv.incr() threw; no try/catch, so a 500 HTML page bubbled to the client, which fell through to its default error text. The fix (PR #66) aliased the env vars at module load and wrapped every voting route in try/catch so real errors reach Vercel logs.

Five lessons from the build

Start with a scope-killer. Hermes / office-hours before any coder.
Match the tool to the stage. Claude Code doesn't write PRDs well; superpowers does.
Always get a second model's eyes. Codex catches what Claude rationalizes.
"Tests pass" ≠ "prod works." Env vars, secrets, and integrations live outside your repo.
Make prod errors observable. Try/catch every KV/API call; log with a tag so you get from "prod broke" to a stack trace in 30 seconds.

Bonus rule: never stack PRs on an unmerged feature branch unless you plan to rebase after the base merges. (PR #67 merged into a stale branch and never reached main — which is why PR #68 exists.)

The PRs, all public on the repo:

PR	Title
#65	feat: community lightning-talk voting tool
#66	fix(voting): Upstash env aliases + route try/catch + clearer errors
#67	feat(vote): read-only results view (merged into stale branch)
#68	feat(vote): read-only results view after voting (reland onto main)
#69	chore(e2e): drop lightning-talk smoke test

Community Lightning Talks — Now Open for Voting

The last block of the night was dedicated to the community. All six submitted lightning talks were presented live — a quick pitch from each speaker. The voting page is live until the next meetup; pick your top three at 02ship.com/vote:

Beyond the Demo — Fayner Brack — "AI can build an MVP in a weekend but can't tell you what to do when a real user shows up on Monday."
pi cloned — Burin Choomnuan — Replicating features of a Claude Code competitor to see how far he can get.
Partner Management System in 2 weeks — Tony Huang — Building a partner management platform on a two-week deadline.
Agentic Workbench — Aboo — An open-source workbench for engineers working with agents.
Next Anthropic — JohnZhou — Investment angle on Anthropic's valuation trajectory and the adjacent opportunities.
Automating the loan serviceability sheet — Linda Liu — Claude-powered tool that extracts financial data and auto-fills mortgage serviceability calculations.

Top-voted talks get the headline slot at the next meetup. Cast your votes →

Join the Next Meetup — Tuesday, May 19

We're back on Tuesday, May 19, 2026 at Haymarket HQ — same venue, same rhythm: 6:00 PM doors, 6:30–7:30 talks, 7:30–8:30 networking.

Still "not a hype session" — real workflows, real lessons, real examples from people shipping with AI. If you've built something with Claude Code and want a 5-minute slot to share what you learned, apply here.

It's free. Spots fill fast.

Register for May 19 on Luma →

See you there.