I’m an AI builder and the founder of an AI-native startup.

Coding has developed a growing body of public, reusable ways of working with agents. In ambiguous, context-rich work that depends on judgment and collaboration across an organization, many effective practices still live in individual experience. I’m documenting and sharing what holds up in real use so more people can bring these practices into their own work, use agents to think, create, and collaborate, and develop new approaches through use.

This perspective comes from eight years in strategy consulting. As a partner, I also helped colleagues across functions bring AI into their day-to-day work and encouraged them to keep sharing and discussing how they were actually using it. Today, I continue this exploration in the company I’m building, from product and business to the team’s daily work, iterating these practices as agent capabilities change and new real-world problems emerge.

Five Agent Skills are currently public and ready to use:

• Hoist the Elephant — When repeated local corrections pull a long conversation off course, it stops the current momentum, recovers the user’s real objective and accumulated understanding, and hands that mainline back for confirmation. It runs only when explicitly invoked.
• 不是而是 — Edits Chinese final copy only when explicitly invoked, removing recurring model-written filler and writing habits while preserving facts, degree, attribution, and genuine uncertainty.
• Better Decisions — Brings the pre-recommendation work of strategy consulting into an Agent’s decision process: restore context, define the real decision, organize facts, inferences, and unknowns, then form a recommendation that can withstand challenge. The final choice stays with the decision-maker.
• Brief Me — Turns AI’s wall of analysis into a conclusion-first, interactive HTML briefing that helps decision-makers see relationships, compare choices, and carry their own judgment back into the conversation.
• Wish Pool — A local-first Agent Skill and visual board that keeps ideas from different Agent sessions on one continuous card, so they can be found, updated, handed off, and continued later.

Since 2023, I’ve kept bringing ChatGPT deep into real work.

Research / synthesis / hypothesis-building / decision support / writing / communication / management…

From getting up to speed on unfamiliar industries and making sense of scattered material and ambiguous information, to forming and challenging hypotheses, making judgments, communicating with clients, coordinating teams, and managing day-to-day work, AI gradually entered different stages of complex work—from understanding a problem to moving an outcome forward.

To me, a model upgrade stopped being just product news. It moved the boundary of work in practice. Across 2023, 2024, and 2025, each year brought another set of tasks that I or my team could hand to AI: some work disappeared altogether, while some steps could be moved forward by fewer people.

As a partner at a strategy consulting firm, I pushed to put AI accounts in the hands of colleagues across functions—from consultants to branding, administration, HR, and finance—and made “What did you accomplish with AI today?” a recurring subject of internal exchange.

I kept sharing my own aha moments and encouraged the team to share specific examples from their work, because methods can be shared, but whether they actually help someone else often depends on how similar the underlying work is. This kind of cross-functional exchange around real use continued for more than three years.

That experience pushed the question one level deeper: an organization does not truly know how to use AI simply because everyone has an account. The harder part is helping practices scattered across individuals become visible, get picked up, and gradually turn into capabilities a team can keep reusing.

In 2026, I chose to become a founder and bring two threads together inside the company I was building and running.

One is to turn the hard work I had done repeatedly in consulting—understanding unfamiliar problems, processing complex information, forming judgment, and moving outcomes forward—into capabilities carried by agents, workflows, and products. The other is to explore, from the ground up, what AI-native work and organizational collaboration can become.

As a founder, product, business, team, and the organization itself all became questions I had to answer from first principles:

In an AI-native company, what gets faster, and what might actually get slower? When everyone has their own agent—and their own long context that is difficult to hand over—how should a team keep working together?

These are not questions I study from outside the company. They are questions I have to keep answering in the daily work of product, business, and the team.

GitHub is an outward extension of that sharing path. I’m beginning to document practices that recur and take shape in real work so more people can use them, adapt them, and keep developing them in their own work.

By AI-native work beyond coding, I mean exploring how agents can become part of a wider range of real work beyond coding. As agents move from completing isolated tasks to taking part in how a piece of work is thought through, made, and moved forward, the way work begins, judgment takes shape, and handoffs and collaboration happen also starts to change.

Thinking, Creating, and Collaborating are the three threads that keep recurring in this work:

• Thinking: bringing agents into research, synthesis, and challenge, helping people see a problem more clearly and form better-grounded judgments.
• Creating: working with agents to turn ideas that are still taking shape into something that can be used, changed, and developed further.
• Collaborating: carrying goals, context, and progress across time, sessions, agents, and people, so work can be picked up and moved forward together.

In this kind of collaboration, agents can take on more of the searching, organizing, reasoning, and follow-through; people still define goals, make trade-offs, correct course, and remain responsible for the outcome. What I care about is whether this division of work can hold up in day-to-day work and within organizations, while people remain able to understand, challenge, and change what agents are doing.

I’ll document and share recurring problems from real work, together with the useful practices that gradually take shape around them. People in similar work contexts can bring these practices into their own work, adapt them, and keep iterating through new use. I’m starting with a set of Agent Skills.

They come from work I’m actively doing every day. Take Visible Signals / 可见信号, for example: we need to select and maintain information sources worth following, design agent processing pipelines for different types of information, bring industry understanding and human judgment into filtering, interpretation, and output, and keep adjusting the system based on the content the team actually handles and product feedback.

I’ll prioritize practices that have already been used in real work and are beginning to show reusable value. What I share will keep evolving as new business problems, agent capabilities, and feedback from use emerge.