AI as a Student: Teaching It Your Logic Is the Real Advantage

EN #ai-agents#claude-code#community#taipei

Talk by 賴俊吾 (Codotx) at the Taipei WordPress Meetup, at CollaPlay, Wanhua. A mostly non-technical audience, which is exactly why the talk is valuable: it manages to explain AI agents without jargon, using a single sustained metaphor — AI as 學生, a student.

Opening thesis:

“Knowing how to use AI” is already the baseline, not the differentiator. Like “knowing how to use Word” — nobody lists that as an advantage. The real 護城河 (moat) is being able to teach the AI your logic, not just use it.

The student’s bad habits

Three LLM limitations, in metaphor form:

  • 金魚腦 (goldfish brain): you close the window and it forgets everything; next time you’re a stranger
  • 注意力不集中 (lack of focus): on long tasks it only remembers the beginning and the end — the context window
  • 莫名的自信 (inexplicable confidence): it doesn’t warn you when it’s only 60% sure — hallucinations

The three memories: AGENTS.md, MEMORY.md, SKILL.md

The most useful part of the talk is the memory framework, which maps 1:1 to what I use in Claude Code:

AGENTS.md is the persistent principles — the analogy is the reminder list stuck to the fridge. What to put there: role, response language, format, tone. What NOT to put there: pending to-dos, project details that change, passwords.

In Claude Code this is layered: ~/.claude/CLAUDE.md (personal, all projects), /project/CLAUDE.md (shared with the team), /project/CLAUDE.local.md (private). And a rule I found golden: if there are conflicts between layers, the model “丟骰子決定” — decides by rolling a die. Avoiding contradictions between layers isn’t optional.

MEMORY.md is the notebook the AI builds for itself as it works, and consults automatically in similar situations.

SKILL.md is Chun-Li’s combo list in Street Fighter: a predefined sequence you run with one command instead of re-thinking the inputs every time. Their pipeline for WordPress: /plan → /todo --tdd → /review → /verify → /submit-review.

The demo: 23 agents, 22 minutes

The main course was a live Dynamic Workflow. Orchestrator prompt: analyze the last 30 days of LINE conversations, order history, products, and articles from a real e-commerce store, and propose new products prioritizing cheap market validation.

The result: 23 agents in ~22 minutes, organized in phases — 7 in parallel extracting data, 3 analyzing (pain points, revenue, audience), 3 proposing, and an adversarial review phase (“assume it fails, refute it”) from which only 1 of 9 proposals survived. The finding: a market gap in LINE/SMS notifications without WooCommerce.

That adversarial phase is what I’m most interested in replicating: most multi-agent pipelines generate; few systematically tear down what they generated before presenting it.

My takeaways

  • The AGENTS/MEMORY/SKILL framework is equivalent to Claude Code’s system, but with naming that’s more accessible for non-devs — useful for explaining this at work
  • I have things in my global CLAUDE.md that should live in MEMORY or in project context instead
  • The 23-agent Dynamic Workflow is replicable with Claude Code’s Task tool — pending an experiment
  • The 🎲 die rule is real: audit conflicts between config layers

The closing line of the talk: 在 AI 強大的年代，培養自己專屬的工作流 — in the age of powerful AI, cultivate a workflow that’s your own.