📋 Key Takeaways
- Central thesis: Model is Mine + Data is Mine — two constitutional laws delivered tonight, written into the first rule of WukongDojo
- Real vs fake agent watershed: where does intelligence enter — hard-coded if/else is a fake agent, only when Wukong decides at each node is it a real agent
- Universal agent loop: Receive → Recall → Reason → Act → Remember — every agent runs this same circle, your Little Wukong is the same tier as ChatGPT
- 12-model whitelist: from $0.03 floor to $5 frontier, 1000× spread — you pick by task, default to floor, upgrade in one click
- Live Lab: build two agents on WukongDojo with different scripts — same question, see how script design changes the output
- Become Wukong · You Can — Big Wukong awakens · Little Wukong is born · today
🎬 Video Replay
🔧 Materials
✏️ Homework
HW1 (warm-up · no submission): build two agents on WukongDojo with different scripts (system prompts) — same question, see how script design changes the output. HW2 (submit · Golden Hoop 4 · Vision): refine the S3 Little Wukong script into a goal-driven step screenplay. ⚠️ Step 0: declare ONE agent category (A/B/C): · A Goal-driven closer (sell / negotiate / decide / advise — take customer from A to B) · B Knowledge Q&A (RAG / domain expert — answer any in-scope question with citations) · C Companion / coaching (long-term relationship, multi-session — walk WITH the user) TOOL EXECUTOR agents (code/API/system ops) are NOT graded this session — that's a later topic. ① Overall goal four things (FILL PER YOUR DECLARED CATEGORY): A=WHO·WHEN·WHAT·DONE-WHEN; B=WHO·SCOPE·OUT-OF-SCOPE·DONE-WHEN; C=WHO·CADENCE·SESSION-ARC·ESCALATION ② Break into 3-7 units (A=steps; B=question types; C=session moves) ③ Each unit has four elements: 🎯 sub-goal + 🛠 leading move + ✅ observable completion signal + 🗣 dialog (with PUSHBACK/OUT-OF-SCOPE/RESISTANCE + agent's boundary-preserving response) ④ 📚 proprietary knowledge: answers "where does ChatGPT crash on this step". In S5 we feed this together.
⏰ Deadline: 5/16 周六 9:00 AM PT
→ Submit homework (S4-VISION)▶ 🤖 AI Grading Rubric
| Dimension | Weight | Description |
|---|---|---|
| Overall goal specificity | 10% | High when at least 3 of WHO · WHEN · WHAT · DONE-WHEN are specified; DONE-WHEN credits even if not maximally observable, as long as it names what the customer leaves with. A stranger being able to roughly say "agent helps X do Z" is the floor |
| Flow decomposition | 15% | High when broken into 3-7 units AND most units read as advancing customer state (sub-goals); one or two topic-leaning units are fine. Deduct only when fewer than 3 (one-shot fantasy) or more than 7 (workflow software not agent) |
| Sub-goal + leading move | 20% | High when each step names a 🎯 sub-goal plus 🛠 some concrete technique (question phrasing, sequencing, anchoring, evidence move — any of these). Does NOT require "order cannot swap" precision. Only "ask nicely / listen patiently / be professional" is low |
| Observable closing criterion | 20% | High when MOST steps have an ✅ closing criterion that includes at least one observable element (specific answer, data point, confirmation action). "Customer confirms X" is fine without verbatim line. Only deduct heavily when ALL steps rely on unobservable internal states ("felt understood", "vibe was right") |
| Dialog + pushback re-lead | 15% | High when AT LEAST 2-3 steps have real dialog (1 line each from agent and customer) AND AT LEAST 1 step shows customer pushback with agent's boundary-preserving response. NOT required for every step to have a perfect pushback-re-lead |
| Knowledge proprietary | 20% | High when 📚 knowledge list shows DOMAIN SPECIFICITY (names concrete things ChatGPT does not know in this niche). Directly attachable as .md is a bonus, NOT a gate. Only deduct heavily when everything is generic ("needs financial knowledge", "needs communication skills") without a domain name |
90-100 Every step has sub-goal + leading move + closing criterion; agent re-leads back after pushback; knowledge directly feedable in S5 · 60-89 Main steps have sub-goals and closing criteria; leading moves and pushback re-leads occasionally missing · 40-59 Steps written but sub-goals vague, closing criteria missing or unobservable · 20-39 Steps are topics not sub-goals; cannot tell when to advance · 1-19 Nearly empty or entirely ChatGPT one-click output