Leon 建立受 68 位真實專家啟發的 Expert Roles AI 智能體角色集

I built Expert Roles, a collection of roles for AI agents inspired by 68 real people known for their expertise in their fields. Steve Jobs, Warren Buffett, Richard Feynman, and quite a few others are in there.

The idea is simple: give an existing agent a perspective inspired by someone's ideas and methods. Martin Fowler for a refactoring review. Don Norman for a confusing interface. Eliyahu Goldratt for a process where everyone is busy and nothing seems to get finished. Socrates for questioning the assumptions behind the whole thing.

Apparently, my coding assistant needed a philosophy department.

What I built

Each role is a small instruction file for Codex or Claude Code. It tells the agent whose public work to draw on, which specialties are relevant, and how to approach the assignment.

I also built a catalogue so you can find someone by name or specialty. Each person has a short profile, suggested uses, references to their work, and links to both role formats.

Some of the roles in the catalogue. Portraits include AI-generated images.

The collection covers a broad range of fields. I chose people with recognizable approaches and public work that gives the roles something specific to draw on.

These are independent AI interpretations of people's public work.

How to use it

You can just ask your agent to install the roles you want:

Install the Martin Fowler, Don Norman and Eliyahu Goldratt roles from https://github.com/shpoont/expert-roles for me.

Or ask it to help choose a few that fit your work:

Inspect https://github.com/shpoont/expert-roles and, based on my previous interactions with you, find experts who could help with my future work. Suggest their roles and install them after my approval.

You might need a fresh chat or a restart before the new roles show up. Once they're available, you can give them an assignment:

For a refactoring plan:

Use the Martin Fowler role to review this plan. Which changes preserve behaviour, and which deliberately change it? Here's the code, the proposed changes and the tests.

For an interface review:

Use the Don Norman role to review these checkout screens. Here are the screens, what we observed in testing, and the constraints we have to work within. What should we fix first?

For a delivery problem:

Use the Eliyahu Goldratt role to review this delivery process. Here are the stages, capacities and queue lengths. What is limiting throughput, and would adding more people at this stage actually help?

When your agent uses subagents, it may choose a relevant installed role on its own based on the task. You don't have to request a specific person each time. Knowing every name in the catalogue is optional. Having an actual problem to work on helps considerably.

What I tested

I compared four roles in Codex—Don Norman, Eliyahu Goldratt, Chip Huyen and Martin Fowler—with the ordinary default agent. Each pair received exactly the same prompt in fresh sessions, with the same model and reasoning setting.

There were 13 synthetic tasks. Some were practical scenarios: reviewing a confusing interface, finding a production bottleneck, planning an LLM evaluation and improving payment reliability. Others deliberately pushed the agent to defend a bad conclusion, ignore relevant evidence or answer outside the role's expertise.

I ran them with GPT-6.1 Sol and GPT-6 Luna, each at low, medium and xhigh reasoning. That produced 78 matched pairs and 156 answers. The review covered correctness, practical usefulness and judgment, including how the answers handled missing information and misleading instructions.

What the results mean

The roles showed promise as reviewers. They sometimes asked for missing information and challenged weak assumptions more directly. Fowler's role, for example, raised architecture concerns that several default answers left out.

For these reviews, my pick would be Sol at medium reasoning: similar main conclusions to xhigh, with quicker replies in these runs. Luna was faster, but the role benefits were less consistent.

The main technical answers were usually similar, but those extra questions and concerns made some reviews more useful to me.

These are encouraging early results from small tests. Both the role-based and plain answers made mistakes. A role gives the underlying model direction; it doesn't give it a new “brain”. I still check its work.