Great thought to collab with multiple people so that provides a comprehensive view and different options available from different folks!
QQ: I use the appropriate expert: TypeScript-type system problems use a TypeScript-focused expert, React architecture uses a React-focused expert
> what does the author mean by appropriate expert? Is it setup as a skill? Or MCP. Can you share an example prompt of everything described as generation vs implement and this expert setup?
Hi, I'm the author, and happy to respond: "appropiate expert" usually means role/routing, not a built-in feature by default. It can be implemented as a prompting pattern ("act as a TS type-system expert"). Or it can also be implemented via skills (reusable instruction packs). What I mean is that I use the right specialist for the problem: Type-level bugs/inference/conditional types -> TypeScript expert behavior. For component boundaries/state/data-flow/perf -> React architecture expert behavior. You can combine skills with MCP, where skill defines expert workflow, and MCP provides capabilities.
As example prompts:
*1) Router prompt (expert setup)*:
```
You are a task router + implementer.
Routing rules:
- If request is about TypeScript types (inference, generics, conditional types, mapped types, overloads, utility types), use TypeScript Type-System Expert behavior.
- If request is about React architecture (component boundaries, state strategy, hooks design, rendering performance, folder/module structure), use React Architecture Expert behavior.
Execution mode:
- If user asks for ideas/spec/options only, run Generation mode (no code edits).
- If user asks to build/fix/refactor, run Implement mode (edit files, run checks/tests).
Always state:
1) Chosen expert
2) Chosen mode (Generation or Implement)
3) Output accordingly
```
*2) Generation mode prompt*:
```
Generation mode only.
Do not edit files or run destructive actions.
Task: Propose 3 approaches for [problem].
For each approach include:
- architecture sketch
- pros/cons
- risk
- recommended choice with rationale
Keep it concise and implementation-ready.
```
*3) Implement mode prompt*:
```
Implement mode.
Use [TypeScript Type-System Expert | React Architecture Expert] based on task.
Task: Implement [feature/fix] in this repo.
Requirements:
- make code changes
- preserve existing behavior unless requested
- run lint/tests relevant to changed files
- report changed files + verification results
- if ambiguous, choose the safest convention-aligned default and proceed
Really interesting to see how different team sizes approach AI-assisted engineering. I manage 10+ autonomous AI agents daily as part of my PM workflow and the pattern I keep seeing is that the teams who treat AI tools as team members rather than utilities get dramatically better results. The orchestration layer matters more than any individual model choice.
We should focus on architecture of the project and not only on design, Agents make silent mistakes which are not apparant upfront, and can be project killer. I use Cloud coding agents heavily but i want the entire control on the architecture flow, so that if something goes wrong, i know where it went wrong. its about being responsible for the code you ship.
Great thought to collab with multiple people so that provides a comprehensive view and different options available from different folks!
QQ: I use the appropriate expert: TypeScript-type system problems use a TypeScript-focused expert, React architecture uses a React-focused expert
> what does the author mean by appropriate expert? Is it setup as a skill? Or MCP. Can you share an example prompt of everything described as generation vs implement and this expert setup?
Hi, I'm the author, and happy to respond: "appropiate expert" usually means role/routing, not a built-in feature by default. It can be implemented as a prompting pattern ("act as a TS type-system expert"). Or it can also be implemented via skills (reusable instruction packs). What I mean is that I use the right specialist for the problem: Type-level bugs/inference/conditional types -> TypeScript expert behavior. For component boundaries/state/data-flow/perf -> React architecture expert behavior. You can combine skills with MCP, where skill defines expert workflow, and MCP provides capabilities.
As example prompts:
*1) Router prompt (expert setup)*:
```
You are a task router + implementer.
Routing rules:
- If request is about TypeScript types (inference, generics, conditional types, mapped types, overloads, utility types), use TypeScript Type-System Expert behavior.
- If request is about React architecture (component boundaries, state strategy, hooks design, rendering performance, folder/module structure), use React Architecture Expert behavior.
Execution mode:
- If user asks for ideas/spec/options only, run Generation mode (no code edits).
- If user asks to build/fix/refactor, run Implement mode (edit files, run checks/tests).
Always state:
1) Chosen expert
2) Chosen mode (Generation or Implement)
3) Output accordingly
```
*2) Generation mode prompt*:
```
Generation mode only.
Do not edit files or run destructive actions.
Task: Propose 3 approaches for [problem].
For each approach include:
- architecture sketch
- pros/cons
- risk
- recommended choice with rationale
Keep it concise and implementation-ready.
```
*3) Implement mode prompt*:
```
Implement mode.
Use [TypeScript Type-System Expert | React Architecture Expert] based on task.
Task: Implement [feature/fix] in this repo.
Requirements:
- make code changes
- preserve existing behavior unless requested
- run lint/tests relevant to changed files
- report changed files + verification results
- if ambiguous, choose the safest convention-aligned default and proceed
```
For my personal assistant Screech, described in https://www.decodingai.com/p/scaling-120-ai-agents-two-tier-orchestration, the answer is like this:
"Appropriate expert” = a Screech subagent, not primarily a Cursor skill or MCP feature.
So I will pick typescript-pro for type-system work and react-specialist for React architecture.
If task spans multiple domains, I use agent-organizer to route/orchestrate.
Really interesting to see how different team sizes approach AI-assisted engineering. I manage 10+ autonomous AI agents daily as part of my PM workflow and the pattern I keep seeing is that the teams who treat AI tools as team members rather than utilities get dramatically better results. The orchestration layer matters more than any individual model choice.
We should focus on architecture of the project and not only on design, Agents make silent mistakes which are not apparant upfront, and can be project killer. I use Cloud coding agents heavily but i want the entire control on the architecture flow, so that if something goes wrong, i know where it went wrong. its about being responsible for the code you ship.