ARC Ladies AI Demo Day: Session 2 Recap

Wednesday, July 15, 8:00 PM - 9:00 PM GMT+8

View Luma event

Hans and Cheryl shared practical AI systems for account management, marketing, and automation — built around how they actually work.

HansCherylMellHost
Hans headshot

Hans — "Second Brain" ecosystem updates tool

00:03–24:00
“One-shotting is a myth. Some influencers claim you can build everything in one go, but that’s never been my experience.”
  • Key takeaway: Useful AI systems are built through iteration, grounded in specific context, and designed around the way you already work.
  • Hans is Singapore-based and works in go-to-market at a Web3 infrastructure company. Second Brain is a personal side project, built and run independently, on public data only.
  • She built a locally hosted front end with Claude Code and demonstrated it using a safe, AI-generated mock dataset.
  • Her goal was a more focused, private alternative to bouncing between a social feed, notes, and tasks.

Components shown

FeatureWhat it does
Signals feedRanks posts from a public social feed by relevance.
Feedback loopUpvotes and downvotes teach the system how to adjust its scoring.
Kanban boardCaptures tasks and links them to action items on account pages.
Summary pagesSummarises recent activity for each entity being tracked.
Notes panelKeeps notes in a Notion-inspired side panel.
Drafting helperDrafts a tailored angle from reference material you supply.
Morning digestPushes a morning brief to chat.
Command-K searchJumps anywhere without navigating the sidebar.

Cost discipline

Hans asked Claude not to insert AI calls everywhere. Only tweet scoring and pitch generation spend tokens; rule-based scripts handle the rest, and Haiku handles high-volume scoring.

Key lessons

  • One-shotting is a myth; useful systems take several rounds of feedback and correction.
  • Be precise when you know what you want, and invite the AI to improve things when you do not.
  • Give AI specific reference material instead of an open-ended web search.
  • Borrow interaction ideas from products you already enjoy using.
  • Use rules instead of model calls whenever a rule is enough.
Cheryl headshot

Cheryl — A marketing “team” of skills and agents

24:00–53:00
“Think of skills like a team: if one consistently underperforms, retire it and find a better fit.”
  • Key takeaway: Durable context turns general AI into a structured team of specialists that can produce consistent work.
  • Cheryl previously led marketing at XBTO and is preparing to run marketing end-to-end at a stablecoin payments company.
  • A skill is one specialist doing one task; an agent combines several skills into a project team.
  • Most of her work goes through specialist skills, with agents reserved for recurring multi-step pipelines.

Skill

One specialist, one job, a fixed output format, and a built-in framework.

Agent

Several skills coordinated into a broader, multi-step project team.

System architecture

LayerRoleWhat it does
Layer 0Entry pointA single /marketing router acts as the CMO.
Layer 1SkillsAround 122 focused specialist files.
Layer 2AgentsDepartment heads and project teams with their own context.
Layer 3Brand memoryVoice, positioning, audience, competitors, and assets.
Layer 4ConnectionsResearch, scraping, browser, Gmail, and Calendar tools.
Layer 5SystemThe registry, hooks, and dashboard generation.
Six departments

Brand & Creative, Content, Lead Gen, Product Marketing, Marketing Ops, and Comms & Social.

Examples in action

Bubble tea growth strategy

Research, viral concepts, costing, KPIs, a 90-day plan, and AI product mock-ups.

Front end from markdown

A polished strategy site generated quickly because the brand context already existed.

Key lessons

  • Context files are the highest-leverage step — establish brand, positioning, messaging, and visual rules first.
  • Aim for roughly 80% of the result with 20% of the effort, then expect a human back-and-forth.
  • Treat skills like hires: hire selectively. Too many skills make routing harder and consume more tokens.
  • Retire skills and agents that consistently underperform.
  • Start from a strong skill library, then refine it for your own work.

Glossary

Second Brain

A personal system that collects signals, notes, tasks, and context so important information is easier to retrieve and act on.

Skill

A focused AI specialist for one repeatable task, with a defined method and output format.

Agent

A multi-step AI worker or project team that coordinates several skills toward a broader goal.

Brand Memory

Durable project context covering voice, positioning, audience, competitors, and visual assets.

MCP

Model Context Protocol: a standard for connecting AI systems to external tools and data sources.

Firecrawl

An API that searches, crawls, and scrapes websites, turning web pages into clean Markdown or structured data for AI systems.

Command-K

A keyboard-driven global search pattern used to navigate quickly across an application.

Token

A unit of text processed by an AI model; more model calls and more context generally increase usage and cost.

Self-hosted

Software run in an environment you control instead of entirely inside a third-party hosted product.

Additional Resources

Links mentioned during the session.

Available now

Hans’ X/Twitter API provider

The lower-cost provider used for the signals feed.

Open Xquik ↗

Want to share something you've built with AI?

We want to see it! If you or anyone you know has built an app, agent, or workflow with AI and would love to share, drop Mell (@gmmell) a message and we'll find you a spot!