Mastering the AI “Golden Squad”: A New Era of Human-Machine Collaboration
Opinion | November 10, 2026
Author: David Tang, Managing Director, Standard Kepler
Originally published in HKET (香港經濟日報)
Executive Summary
AI is accelerating from conversational Q&A to autonomous task execution. Lauren Tan — known online as “Potato Sister” (土豆姐) and widely regarded as “the person who uses Grok Bot best on Earth” — published her pstack AI workflow methodology in September 2026, revealing a staggering metric: she shipped over 2,500 pull requests into production in a single month. That is 100x the output of heavy AI users and several hundred times that of ordinary engineers. This report distills her methodology into actionable principles for building trust with AI agents, separating review from production, and transforming from a “human microphone” into a systems architect.
The Problem: Most People Use AI Like a Better Search Engine
The dominant pattern of AI usage today is question-and-answer: you write a prompt, the AI responds, and if the result is unsatisfactory, you refine and retry. The tools have upgraded, but the workflow has not.
Tan described her early experience with Cursor on a new project: open the software, test, screenshot, copy the error message to AI, wait for the fix, and repeat. She called herself a “human microphone” — because the AI could not see its own output, it relied entirely on her to relay information back and forth.
The critical difference between most users and Tan is that most people simply repeat this cycle. Tan asked a deeper question: Why must the AI ask “me”? Then she systematically taught the AI to eliminate every step that required her intervention.
The Trust Curve: From Supervision to Autonomy
The relationship between humans and AI can be mapped along a trust curve. The more you trust the AI, the more tasks you can delegate, and the more AI agents can work for you simultaneously.
Building the Curve
| Stage | Action | Outcome |
|---|---|---|
| 1. Observation | Watch the AI work, correct errors in real-time | Baseline understanding of capabilities |
| 2. Skill Documentation | Write the corrected workflow into a “skill” | Reusable instructions for future tasks |
| 3. Delegation | When the AI succeeds consistently, step back from the process | Human freed for higher-order decisions |
| 4. Full Autonomy | Multiple AI agents operate in parallel with minimal oversight | Human becomes architect, not operator |
Trust is built through evidence, not hope.
Principle 1: Separate Review from Production
The most important structural principle: the agent that produces output must not be the agent that reviews it. Self-review is the least reliable form of quality control in any system.
Tan’s approach uses multi-agent separation of duties:
- Production Agent — writes code, drafts content, performs the core task
- Review Agent — audits output using a different LLM model
- Human — validates evidence, makes final decisions on ambiguous cases
This is not merely a technical preference. It is a governance principle borrowed from financial services: the person who executes a trade does not confirm it; the person who drafts a contract does not approve it.
Why Different LLMs Matter
Using the same model for production and review risks systematic blind spots. Different models have different training distributions, failure modes, and reasoning biases. Cross-model review catches errors that single-model review misses.
Principle 2: Demand Evidence, Not Explanations
AI excels at generating plausible-sounding but incorrect explanations. The antidote is to require evidence that can be verified at a glance:
| Task Type | Required Evidence |
|---|---|
| Research compilation | Source links with attribution (official vs. secondary) |
| Data analysis | Screenshots, charts, raw data exports |
| Code generation | Test results, execution logs, performance benchmarks |
| Financial analysis | Footnotes distinguishing fact from inference |
If the evidence is not immediately verifiable, the output is not trustworthy.
Principle 3: Eliminate Repetition Through Automation
When the AI performs a highly repetitive task, do not ask it to regenerate the solution each time. Instruct it to write a program instead.
| Approach | Stability | Cost | Scalability |
|---|---|---|---|
| AI regenerates each time | 99% accuracy = still fails 1 in 100 | Higher token cost | Limited by API rate limits |
| Program automates the task | 100% consistency | One-time development cost | Scales to millions of executions |
Use programs for deterministic tasks. Reserve AI for ambiguous, judgment-based work.
Principle 4: First-Principles Decomposition
Before automating any workflow, decompose it to its smallest components and question each step:
Why must it be done this way?
Case Study: AI-Powered Recruitment
A traditional recruitment process: post on platforms, review applications, conduct interviews. Tan’s AI-native approach:
- Data Source: Social network connections become the talent pool (previously unmanageable at scale)
- First Filter (AI Agent Jev): Evaluates each connection against job requirements — “Should we reach out to this person?”
- Second Filter (HR Agent): Writes summaries, conducts deeper screening
- Outreach Agent: Uses a dedicated email/WhatsApp account to contact candidates with personalized messages based on their profiles
- Follow-up Agent: Manages responses, gauges interest, schedules human interviews
Human role: Face-to-face interviews — the only step that genuinely requires human judgment.
This workflow embodies every principle: production/review separation (Jev filters, HR reviews), evidence (summaries for human validation), automation (API-driven messaging for stability), and first-principles thinking (why use platforms when your network is the database?).
When to Start: Why Not Today?
Tan described her first night delegating to AI: she was so anxious she could not sleep. The result? Her monthly output became several hundred times that of an ordinary engineer.
The barrier is not technical. It is psychological.
| Fear | Reality |
|---|---|
| “The AI will make mistakes” | It will. That is why you build review layers and start with low-stakes tasks. |
| “I will lose control” | You gain control over systems instead of being trapped in execution. |
| “It takes too long to set up” | The initial investment pays compound returns. |
The question is not whether to begin. It is why not today.
Conclusion
The AI “golden squad” is not about replacing humans. It is about redefining the human role from operator to architect. The tools are ready. The methodology is proven. The only remaining variable is the willingness to let go of the microphone and pick up the blueprint.
Standard Kepler Research | standardkepler.com