A practical walkthrough of running Qwen 3.8 Max through ModelScope in Claude Code, from account creation and API setup to a first prompt. Includes a take-home launcher, thinking adapter, and optional web search and browser tools.
Links
Showcase:
Presentations
Presentations from our meetups will go here.
#11 - ABC at Lifelong Learning Institute Fri, 30 Oct 2026 6:00 PM – 9:00 PM
Bennett Clement shares his custom software factory and approach to agentic software development.
Links
#10 - ABC Mini Edition Tue, 29 Sept 2026 6:00 PM – 9:00 PM
Fast is not the same as finished ABC's first MLX demo — running Qwen3.8 Flash-Next agents on a Mac Studio M3 Ultra across oMLX, Unsloth llama-server, and antirez's ds4, and what each engine did with a 30-minute poster job versus a 5-hour OMR build.
Links
Two Brains, One Operator How /remember and /recall route personal and company second brains to the right destination, running on local Qwen 3.8 Flash Next, and where harness-switching primitives are headed next.
Links
#9 - ABC Local LLMs at AmChamSG Thu, 27 Aug 2026 6:00 PM – 9:00 PM
Local Qwen, Real Work How Qwen 3.8 27B runs on a single DGX Spark, how it plugs into a real coding harness, and how local Qwen compares with Qwen Max on real code-review and product-build tasks.
On Speculative Decoding Five rules for when a draft model actually makes your big model faster — and when it doesn't. 199 benchmarks on a DGX Spark.
Links
Slides Lecture Open project Full results Companion write-up
[ℹ]
Every figure in the deck is drawn from 199 benchmark runs on a single DGX Spark; each number links back to the config page that produced it.
#8 - ABC at Three Blind Pigs Mon, 29 Jun 2026 6:00 PM – 9:00 PM
Context Shrinker A quick-start walkthrough for bounded context workflows, with pointers to context-compiler, repo maps, and custom workflow tradeoffs.
Links
Slides context-compiler Aider repo map
[ℹ]
Yi shared this as a quick start for bounded context work. For teams with many custom workflows, he noted that it may be cheaper to fork or build a tailored approach.
[ℹ]
Another reference implementation for mapping repository context.
Hard parts: Building with Gen AI A practical presentation on the hard parts of building with generative AI, from engineering constraints to production workflow decisions.
Introduction to Claw Breeding for Youths Notes from running an OpenClaw bootcamp for nearly a hundred high school students for DSTA's Young Defence Scientists Programme.
Links
Making Software Agents Can Use Notes on designing software surfaces and workflows that agents can use effectively.
Links
#6 - ABC at Lorong AI Thu, 23 Apr 2026 6:00 PM – 9:00 PM
Codex in the Wild How Nexius Labs uses Codex as the main driver for building and teaching SMEs to ship AI-assisted tools.
Links
Infrastructure-based Safety For Your 'Claw A walkthrough of infrastructure-based safety for agentic tools, covering connector threat models, MCP gateways, and practical design patterns.
The Lessons I Learned Coding 10+ Apps with Claude Code Two agentic patterns to fix poor instruction following — Generator + Evaluator and Thin Orchestrator + Specialized Workers — composed inside Claude Code skills.
Workforce OS A file-based multi-agent operating system for a seven-person team. No database, no server — just markdown and git.
Links
Slides Starter prompt
[ℹ]
I was asked by quite a number of you in person and via DM to open source workforce-os. I am not going to, because that is exactly against the philosophy of everything I was sharing (build it yourself, customise it to your needs). However, I recognise that not everybody knows where/how to start building. So instead of open sourcing the software, I asked my agent to generate a "starter prompt" for you to use to start building your own system. You can - if you wish - also reference nanoclaw as a starting git repo (but frankly there's no need because it will find the repos itself in plan + extra high effort mode).
So this is more aligned with my philosophy - learn and build your own thing, for your own needs. This community will all be here to help answer any questions if you get stuck!
Have fun building!
— Gaurav
#5 - ABC at Google Fri, 27 Mar 2026 5:30 PM – 9:00 PM
Tips, Threats, and What Actually Works Logan Goh's ABC at Google talk on why viral prompting tricks sometimes appear to work, and why expert framing and multi-perspective prompting are more reliable tools.
Links
Slides Page Warmthink/coldthink prompts
[ℹ]
Original Coldthink and Warmthink prompt documents shared by Logan Goh in the ABC chat on Feb 18, 2026.
Projects
Want to add something? Follow the contribution guide
A local-first desktop AI team that turns goals into reports, slide decks, videos and web pages on macOS and Windows.
An early hosted, proprietary network for agents to publish artifacts and exchange reviews.
Every MRT and LRT line in Singapore as a 3D world in your browser. Drive the train, or ride along, walk through the cars and change platforms, with the hour's real crowds and frequencies. Rebuilt in 10 days with Claude Code from a passenger-flow map made a year earlier.
A weekly shortlist of what's worth doing in Singapore. Every night 23 scrapers feed a Claude Opus 5 pipeline that filters, writes, researches on the web, and ranks 5 to 10 picks, served on the web, in a Thursday email, in an iOS app, and over an MCP server. Built solo with Claude Code.
Credited motion references and recipes for humans and their authorised agents, with a shared profile and private saved library.
A shared work ledger for people and AI agents — every piece of work carries who claimed it, what was done, on what basis, and the evidence, kept outside the agent that did it.
ContextSpectre
A CLI and TUI over Claude Code's local session files — see what is filling the context window and what the session has cost, then cut the parts that no longer earn their place.
Pastewatch
Secret redaction for AI coding agents — file reads and outbound API requests are rewritten with typed placeholders before they leave the machine, and shell commands carrying credentials are blocked before they run.
Bulwark
An OS-level read gate for AI coding agents — a protected file open is denied, allowed, or held for consent before the bytes reach the agent.
A free persistent multiplayer game where builders can try agent tool use, session recovery, and exploration alongside human players.
An open-source suite of 291 Singapore-focused MCP tools, agent skills, and Claude plugins for evidence-based AI workflows using public data.
24/7 live entertainment platform for AI agents
102 original songs across 11 albums, generated from Lenny Rachitsky's 349 newsletters and 289 podcast episodes. it turns his product wisdom into 2-3 minute tracks you can stream and shuffle. made in singapore.
A pay-as-you-go AI model aggregator that gives people one place to use leading chat and image models without juggling multiple subscriptions.
A native macOS editor for Markdown and JSON, built for quickly writing, previewing, validating, and moving between structured and prose-heavy text.
pg_ffmpeg
A PostgreSQL extension that brings FFmpeg media processing into SQL, with safe filter graphs and optional hardware acceleration.
A retro-styled data sequencer that turns Singapore public datasets into generative music, with cassette-like datasets, sector pads, and live controls for shaping rhythm, density, and effects.
Articles
Want to add something? Follow the contribution guide
a code riff conversation with Roger Olofsson, CEO and Co-Founder of Olofsson Consulting, on running an ai-native executive search firm. includes a live walkthrough of the matching platform his team built across 395,000 candidates, why the whole firm now works in claude code after building their own software, the new roles ai created including an agentic product manager, and the promise he made his team that nobody would lose their job.
a code riff conversation with Gaurav Keerthi, former Deputy Chief Executive of Singapore's Cyber Security Agency and now CEO of StrongKeep, on using AI without leaking your data or API keys, building your own lightweight harnesses, and why good cybersecurity shouldn't be a luxury for small teams.
A rebuilt HTML archive of Code with Claude Tokyo workshop and founder-stage sessions, reconstructed from attendee photos with the original source images appended beneath each section.
Introduction to Hermes Agent, an open-source AI agent from Nous Research.