Post
JA EN

The Expert Who Doesn't Write Prompts — Meta-Prompting and the Evolution to Orchestrator

Posted
By
11 min read
The Expert Who Doesn't Write Prompts — Meta-Prompting and the Evolution to Orchestrator
  • Target audience: Software engineers, IT professionals interested in AI adoption
  • Prerequisites: Basic experience with AI tools such as GitHub Copilot, ChatGPT, or Claude
  • Reading time: 15 minutes

Overview

“People in their 50s have strong verbal abilities, so they can write detailed prompts” — this was a claim made in the AI-Era Career Strategy series.

However, when observing the most productive senior engineers in practice, they aren’t writing detailed prompts themselves. Instead, they’re having AI write the prompts.

Behind this seemingly contradictory behavior lies a technique called meta-prompting and an evolution toward a new role: the orchestrator. This article examines the essence of “not writing prompts” based on the latest research.

The Limits of “Writing Detailed Prompts”

The Conventional Understanding

In the AI-Era Career Strategy series, we argued:

“A concrete scenario where verbal abilities of people in their 50s shine: where younger workers ‘communicate through multiple rounds of trial and error,’ those in their 50s can ‘convey accurately on the first try.’”

This is indeed one valid approach. However, there is another approach.

The Other Approach: Let AI Write It

Some experienced practitioners adopt the following workflow:

1
2
3
4
5
6
7
8
9
10
11
12
13
14

❌ Traditional approach:
"Please implement JWT authentication compliant with OAuth 2.0.
Requirements:
- Access token expiry: 15 minutes, refresh token: 7 days
- Rate limiting: 10 requests per minute for auth endpoints
- Return 429 status with Retry-After header on failure
- Manage token blacklist in Redis (for logout support)
Please consider integration with the existing User model."

✅ Meta-prompting approach:
"I want to implement authentication.
First, please suggest a checklist of requirements
to consider when implementing this kind of feature.
I'll review it before having you write the code."

In the latter approach, AI proposes detailed requirements, and the human evaluates and refines them. In other words, the detailed specification of the prompt is delegated to AI.

Meta-Prompting: The Technique of Having AI Write Prompts

Research Demonstrates Meta-Prompting’s Effectiveness

Zhang, Yuan & Yao (2023) theoretically systematized this technique in their study “Meta Prompting for AI Systems”1:

“Meta Prompting (MP) is a framework that focuses on the formal structure of tasks, emphasizing structure rather than content-specific examples. We further extend it to Recursive Meta Prompting (RMP), realizing an automated process where LLMs generate and improve their own prompts.”

The results are impressive:

  • The Qwen-72B model achieved state-of-the-art results on MATH, GSM8K, and Game of 24
  • Significantly improved token efficiency compared to traditional few-shot methods
  • Enabled a general-purpose approach independent of examples

Advantages of Meta-Prompting

flowchart TB
    subgraph Traditional["Traditional Prompting"]
        direction TB
        T1["Human thinks through details"]
        T2["Human writes details"]
        T3["AI executes"]
        T1 --> T2
        T2 --> T3
    end

    subgraph Meta["Meta-Prompting"]
        direction TB
        M1["Human conveys intent"]
        M2["AI proposes details"]
        M3["Human evaluates & refines"]
        M4["AI executes"]
        M1 --> M2
        M2 --> M3
        M3 --> M4
    end

    Traditional --> Result1["Human cognitive load: High"]
    Meta --> Result2["Human cognitive load: Low<br>Quality: Equal or better"]

The OpenAI Cookbook also recommends meta-prompting2:

“Meta prompting is an advanced technique that uses prompts to directly generate, refine, and interpret other prompts. Rather than answering user questions directly, it enables more dynamic, flexible, and effective AI interactions.”

Practical Examples: How to Use Meta-Prompting

Example 1: Code Review Request

1
2
3
4
5
6
7
8
9

❌ Traditional:
"Please review this PR. Specifically check security,
performance, code style, and error handling."

✅ Meta:
"I'd like to request a review of this PR.
First, please create a checklist of points to verify
when reviewing this type of code change.
Then, let's proceed with the review for each point."

Example 2: Architecture Design

1
2
3
4
5
6
7
8
9
10
11
12

❌ Traditional:
"Design 3 services — user authentication, product management,
and order management — in a microservices architecture.
Use gRPC for communication, PostgreSQL for databases,
Redis for caching..."

✅ Meta:
"I want to design an e-commerce backend.
First, please organize the architectural choices
that need to be made for this type of system,
along with the trade-offs for each option.
After I make my selections, let's proceed to detailed design."

From Conductor to Orchestrator

Addy Osmani’s “Orchestrator” Concept

Addy Osmani, Engineering Manager at Google Chrome, described the evolution of the engineer’s role as “from conductor to orchestrator” in a November 2025 article3:

“In this agentic era, the role of the software engineer is evolving from implementer to manager — from coder to conductor, and ultimately to orchestrator.”

“Senior engineers may be starting to notice: our job is shifting from ‘How do I code this?’ to ‘How do I get the right code produced?’ — a subtle but profound change.”

The Difference Between Conductor and Orchestrator

flowchart TB
    subgraph Conductor["Conductor"]
        direction LR
        C1["Works with one AI agent"]
        C2["Gives fine-grained instructions in real-time"]
        C3["Makes decisions at each step"]
    end

    subgraph Orchestrator["Orchestrator"]
        direction LR
        O1["Manages multiple AI agents"]
        O2["Sets high-level goals"]
        O3["Reviews final results"]
    end

    Conductor --> Style1["Cognitive load: High<br>Parallelism: Low"]
    Orchestrator --> Style2["Cognitive load: Low-Medium<br>Parallelism: High"]

Osmani explains3:

“For an orchestrator, human effort is front-loaded (writing good task descriptions and specs for agents, setting up appropriate context) and back-loaded (reviewing and testing the final code), but not much is needed in between. This allows a single orchestrator to manage far more work in parallel than working with one AI at a time.”

Why Senior Engineers Are Well-Suited as Orchestrators

RedMonk’s analysis reveals an interesting insight4:

“Multi-agent orchestration remains out of reach for lower-skilled developers. As Gergely Orosz observes, parallel agent work requires skills that experienced tech leads have honed. So far, only senior-and-above engineers have been successful at using parallel agents.”

Reasons why senior engineers excel as orchestrators:

  1. Task decomposition ability: Can break large tasks into appropriately sized pieces
  2. Quality standards: Hold the criteria to evaluate AI output
  3. Risk judgment: Can identify where human judgment is needed
  4. Global optimization perspective: Can avoid the trap of local optimization

Strategic AI Delegation Skills

What Anthropic’s Internal Research Reveals

Anthropic surveyed its own employees’ AI usage and discovered interesting patterns5:

“The fact that AI underperforms in large, complex environments, or where a lot of tacit knowledge or context is needed, maps closely to the kinds of tasks that employees said they don’t delegate to AI.”

In other words, effective AI users accurately judge what should and shouldn’t be delegated to AI. Researchers call this “strategic AI delegation skills.”

An even more important finding:

“People are delegating more autonomy to Claude over time.” “Engineers are delegating increasingly complex work to Claude, and the oversight required for Claude is decreasing.”

Balancing Delegation and Oversight

flowchart TB
    subgraph Delegation["Evolution of Delegation"]
        direction TB
        D1["Early: Delegate simple tasks<br>Oversight: High"]
        D2["Middle: Delegate moderate tasks<br>Oversight: Medium"]
        D3["Mature: Delegate complex tasks<br>Oversight: Low-Medium"]
    end

    D1 --> D2
    D2 --> D3

    D3 --> Skill["Strategic AI Delegation Skills"]
    Skill --> Result["High productivity<br>Appropriate quality assurance"]

This evolution becomes possible as experience accumulates, because:

  • Tacit knowledge functions as a filter: You intuitively sense when something is off
  • Pattern recognition becomes efficient: You quickly spot typical problems
  • Risk assessment becomes accurate: You can judge where to focus attention

Practice: Integrating Meta-Prompting and Orchestration

Workflow Example

flowchart TB
    subgraph Phase1["Phase 1: Structuring"]
        direction TB
        P1A["Convey intent"]
        P1B["AI proposes structure"]
        P1C["Human evaluates & refines"]
        P1A --> P1B
        P1B --> P1C
    end

    subgraph Phase2["Phase 2: Parallel Execution"]
        direction LR
        P2A["Delegate Task A to Agent 1"]
        P2B["Delegate Task B to Agent 2"]
        P2C["Delegate Task C to Agent 3"]
    end

    subgraph Phase3["Phase 3: Integration & Review"]
        direction TB
        P3A["Collect deliverables from each agent"]
        P3B["Quality review"]
        P3C["Integration & final adjustments"]
        P3A --> P3B
        P3B --> P3C
    end

    Phase1 --> Phase2
    Phase2 --> Phase3

Concrete Meta-Prompt Examples

Meta-prompt for feature implementation:

1
2
3
4
5
6
7
8
9

I want to implement [feature name].

First, please organize the following:
1. Requirements to consider (functional and non-functional)
2. Implementation options and trade-offs for each
3. Typical failure patterns and mitigation strategies
4. Scenarios that should be tested

After I review and select from these, let's proceed with implementation.

Meta-prompt for code review:

1
2
3
4
5
6
7
8
9
10
11
12

I want to review this PR/code change.

First, please create a checklist for reviewing this type
of code change ([type of change]):
- Security perspective
- Performance perspective
- Maintainability perspective
- Testing perspective
- Other commonly overlooked points

After I confirm the checklist, let's proceed with
the specific review for each item.

Meta-prompt for design decisions:

1
2
3
4
5
6
7
8
9
10
11
12
13
14

I'm considering the design of [system/feature].

First, please organize the key choices to be made
in this type of design:
1. Architecture pattern options
2. Technology stack options
3. Data modeling options

For each option, please organize:
- Pros/Cons
- Suitable use cases
- Cases to avoid

After I make my selections, let's proceed to detailed design.

The Relationship Between Meta-Prompting and “Verbal Ability in Your 50s”

Not a Contradiction, but an Evolution

“The ability to write detailed prompts” and “meta-prompting” are not contradictory. Rather, strong verbal ability makes meta-prompting more effective.

  1. Clear communication of intent: Can precisely convey what you want to achieve
  2. Evaluation of AI proposals: Can judge the quality of AI-generated structures
  3. Accurate correction instructions: Can articulate directions for improving AI proposals
  4. Setting quality standards: Can define what constitutes “good” output
flowchart LR
    subgraph Skills["Where Verbal Ability Shines"]
        direction TB
        S1["Communicating intent"]
        S2["Evaluating proposals"]
        S3["Giving corrections"]
        S4["Quality standards"]
    end

    subgraph Application["Application in Meta-Prompting"]
        direction TB
        A1["High-level goal setting"]
        A2["Verifying AI output"]
        A3["Course-correction feedback"]
        A4["Final quality judgment"]
    end

    S1 --> A1
    S2 --> A2
    S3 --> A3
    S4 --> A4

How the Use of Experience Is Changing

Conventional UnderstandingEvolved Understanding
Write detailed promptsEvaluate detailed prompts
Verbalize requirementsVerify AI’s proposed requirements
Dictate solutionsSet the direction for solutions
Ensure quality yourselfJudge and correct quality

Summary

  1. Meta-prompting: The technique of having AI write prompts has been validated by research. It improves token efficiency and enables a more general-purpose approach

  2. From conductor to orchestrator: The engineer’s role is evolving from a “conductor” who gives fine-grained instructions to a single AI, to an “orchestrator” who manages multiple AI agents

  3. Senior advantage: Parallel agent work and strategic AI delegation require the task decomposition ability, quality standards, and risk judgment that senior engineers possess

  4. Leveraging verbal ability: Verbal abilities developed over decades are valuable not just for “writing detailed prompts” but also for “evaluating AI proposals and providing course corrections”

  5. Strategic AI delegation skills: Effective AI users accurately judge what should and shouldn’t be delegated. This skill develops with experience

“Not writing prompts” doesn’t mean “not thinking.” It means thinking at a higher level and leveraging AI strategically.

Explore other articles related to this topic:

References

References corresponding to citation numbers in the text are listed in order.

Additional References (not cited by number in the text)

On the accuracy of citations: The research cited in this article has been verified through:

  • Confirmation via academic databases (arXiv, Google Scholar)
  • Verification on official websites
  • Cross-referencing against multiple independent sources

  1. Meta Prompting for AI Systems - Zhang, Y., Yuan, Y., & Yao, A. C. (2023). arXiv preprint. [Reliability: High] ↩︎

  2. Enhance your prompts with meta prompting - OpenAI Cookbook. [Reliability: High] ↩︎

  3. Conductors to Orchestrators: The Future of Agentic Coding - Osmani, A. (2025). [Reliability: Medium-High] ↩︎ ↩︎2

  4. 10 Things Developers Want from their Agentic IDEs in 2025 - RedMonk (2025). [Reliability: Medium-High] ↩︎

  5. How AI Is Transforming Work at Anthropic - Anthropic Research (2025). [Reliability: High] ↩︎

AI・Technology
AI Meta-Prompting Expertise Orchestration Agentic-Coding Delegation
This post is licensed under CC BY 4.0 by the author.