Flagship guide · Multi-AI Systems · Updated July 30, 2026

I Was Tired of Being a One-Person Team—So I Built an AI Round Table

I did not need another AI that could write a polished answer. I needed help carrying the five jobs behind every serious project: research, creation, criticism, usability, and distribution. The free prompt below solved the thinking problem. The updated workflow solves what happened next.

By NotebookLM Guide Editorial TeamUpdated July 30, 2026Tested across single-model and multi-tool workflows

Copy the free five-role prompt See what the update fixes

Works in ChatGPT, Claude, Gemini, or another capable model · Multi-AI routing is optional · Human approval stays central

The five-seat system

One human chair. Five specialist seats. Visible disagreement before action.

AI Round Table workflow with five specialist seats coordinated by a human chair — Research, writing, editing, usability, and marketing contribute separate recommendations. The human chair approves the final decision and any consequential action.

30-second answer

An AI Round Table gives several specialist roles separate responsibilities, requires them to expose disagreements, and assigns one chair to make the final recommendation. It can run in one model, across several AI tools, or through limited task execution agents. Human approval remains required for consequential actions.

Start with the free five-role prompt. Use NotebookLM when the Round Table needs cited evidence from a controlled source set. Add other AI tools or agents only when they remove a real coordination bottleneck.

Trend Researcher

Writer

Editor

Usability Engineer

Marketing Director

Human Chairdecides and approves

The AI Round Table keeps the original five-seat metaphor. NotebookLM can supply cited evidence, other AI tools can complete assigned work, and the human chair makes the final decision.

Inside this guide

Story first · practical steps second

The problem was not that AI could not answer me

It was that I still carried the whole project

For a long time, my workflow looked efficient from the outside. I could open one chat, describe a project, and get a respectable answer in seconds. But the answer was only the visible part of the work.

I was still the person who had to decide what to research, find the evidence, explain the audience, turn the research into a draft, edit the draft, check whether it was usable, think about distribution, and remember every unresolved question. When one tool finished, I copied the result into the next tool. When two answers conflicted, I quietly reconciled them myself.

The AI was producing faster. I was still coordinating everything.

That distinction matters because coordination is the work no productivity demo counts. It is the constant switching between roles. It is remembering why one source was trusted, which paragraph still needs evidence, whether the editor removed the important nuance, and whether the final page actually works after the writing is finished.

The problem was not that AI could not produce answers. The problem was that I was still doing all the coordination, judgment, and follow-through.

My first fix: put five specialists in the same conversation

The original Round Table still works

I stopped asking one assistant to be creative, analytical, concise, persuasive, critical, and practical at the same time. Real teams do not work that way. The researcher is not supposed to rewrite every sentence. The editor is not supposed to invent the marketing strategy. Clear boundaries make each perspective sharper.

So I built a five-seat content Round Table and forced the roles to speak separately before the chair made a decision.

Trend ResearcherWriterEditorUsability EngineerMarketing Director

Free AI Round Table Prompt · Updated AI Round Table Prompt

You are the Chair of an AI Round Table. Create five specialists with narrow, non-overlapping responsibilities for the task below.

Default seats:
1. Trend Researcher — evidence, trends, audience questions, and opportunities. Does not write the final deliverable.
2. Writer — story, explanation, structure, and first draft. Does not perform SEO or final QA.
3. Editor — clarity, cuts, logic, and voice. Does not introduce unsupported new claims.
4. Usability Engineer — reader effort, pacing, navigation, comprehension, and friction. Does not rewrite the whole piece.
5. Marketing Director — positioning, title, hook, CTA, and distribution. Does not override factual accuracy.

Process:
Round 1: Each specialist gives an independent recommendation without seeing a manufactured consensus.
Round 2: Each specialist identifies one disagreement, missing assumption, or weakness in another recommendation.
Round 3: The Chair decides. Do not average every view into a vague compromise.

Final answer must include:
- the decision,
- the strongest reason,
- what changed because of the debate,
- evidence still needed,
- unresolved uncertainty,
- and the next action.

Task: [INSERT THE REAL TASK]
Context and sources: [INSERT THEM]
Audience and constraints: [INSERT THEM]
Desired output: [INSERT IT]

Why it helped: the prompt interrupted the “one smooth answer” pattern. It exposed trade-offs, protected specialist viewpoints, and made the final recommendation feel like a decision rather than an average.

The first Round Table solved the thinking problem

And that was a real breakthrough

Before

One overloaded instruction

Every priority competed inside the same answer. The result was often competent, balanced, and strangely flat.

Round Table

Independent perspectives first

Each specialist protected one professional concern before the chair reconciled the conflict.

Human benefit

Trade-offs became visible

I could see why the editor wanted a cut, why the researcher resisted it, and what the marketing angle would cost in nuance.

The original prompt remains the right starting point for brainstorming, critique, planning, and one-time decisions. Most people do not need an autonomous agent team to improve a draft. They need sharper perspectives and a chair that refuses fake consensus.

It sounded like a team. But I was still managing everything.

The missing second half

The five voices could disagree convincingly. Underneath, they were still one model sharing one conversation. When the response ended, so did the “team.”

I still gathered every source

I still moved context between tools

I still assigned every next step

I still checked every factual claim

I still edited every working file

I still ran every final QA check

From a better conversation to a working system

Use the lightest level that removes the bottleneck

The update is not “add more agents.” It is a progression. Stop as soon as the workflow is good enough.

Start here

Run the Round Table in one AI

One capable model plays several tightly clearly limited roles. Use it when the problem is narrow thinking, weak critique, or premature consensus.

Add tools when useful

Give different seats different AI tools

Different tools handle different kinds of work: source-grounded evidence, web discovery, synthesis, production, or independent review.

Only when needed

Let agents complete clearly limited tasks

Agents with limited permissions inspect files, create artifacts, run checks, repeat tasks, or work in parallel. Consequential actions still wait for human approval.

Specialized roles improve thinking. Specialized tools improve division of labor. Agents with limited permissions can improve execution. Human judgment connects all three.

What the updated Round Table actually fixes

Human problem → workflow response

Problems the updated AI Round Table workflow is designed to solve.
Existing problem	Updated response
AI agrees too quickly	Specialists form independent views before reading the others.
Every role blends into the same voice	Each specialist receives a narrow mission, exclusions, inputs, and a required artifact.
Research and writing contaminate each other	Evidence collection, interpretation, drafting, and review become separate stages.
The user copies material between tools	A chair defines compact handoffs and keeps the project state visible.
Nobody verifies the finished result	A separate reviewer or red-team pass checks claims, execution, and acceptance criteria.
Recurring work must be restarted manually	Defined checks can be scheduled when the platform supports recurring execution.
Autonomous changes create risk	Permissions, backups, retry limits, hooks, and approval gates bound the work.
Disagreements disappear into polished prose	The chair records unresolved conflict instead of hiding it inside a smooth answer.

Choose tools by job, not by brand

The method stays the same even when the tools change

ChatGPT, Claude, Gemini, NotebookLM, Perplexity, and execution environments are examples—not mandatory seats. Use only the tools that create a meaningful difference in the work.

Defined evidence

Source-grounded research workspace

Use a tool such as NotebookLM when the Round Table must stay anchored to a controlled source set and show where claims came from.

Open-web discovery

Search-enabled research

Use a search-enabled model or research tool when the project needs current sources, competitor discovery, or questions the existing files cannot answer.

Reasoning and synthesis

General-purpose chair

Use ChatGPT, Claude, Gemini, or another capable model to frame the decision, design briefs, reconcile conflicts, and produce the final recommendation.

Independent challenge

Separate reviewer

Use another model, isolated conversation, or reviewer agent so the checker is not merely defending its own previous answer.

Files and systems

Execution environment

Use a coding, browser, or agent platform only when the project must operate on files, run tests, create artifacts, or repeat defined work.

Accountability

Human chair

The person defines the outcome, chooses acceptable risk, resolves value judgments, and approves consequential actions.

Where NotebookLM fits in the AI Round Table

Why this belongs on NotebookLM Guide

The AI Round Table guides the discussion and the decision. NotebookLM can supply the cited evidence. When the project must stay anchored to a defined collection of reports, transcripts, policies, papers, or internal documents, one seat should be responsible for protecting that source boundary.

That is why this guide belongs here. It does not replace the site’s product-specific tutorials. It shows what happens after NotebookLM has produced grounded findings: the Round Table decides who should interpret them, challenge them, turn them into an artifact, and approve the result.

NotebookLM’s job

NotebookLM protects the source set

It extracts supported findings, keeps citations attached, identifies conflicts between sources, and records what the evidence cannot yet establish.

The Round Table’s job

The Round Table carries the decision forward

It assigns interpretation, writing, implementation, and review to clearly limited roles—then returns the unresolved trade-offs to the human chair.

Curate the approved source set

Load only the materials the Round Table is allowed to treat as evidence. Separate verified sources from inspiration, assumptions, and open-web discovery.

Extract findings with citations

Ask NotebookLM for supported claims, exact source references, contradictions, and missing evidence—not a polished final recommendation.

Create a compact research summary with citations

Return the claims, citations, confidence level, source conflicts, and unresolved questions in a format another role can inspect quickly.

Hand off, do not blend contexts

Pass the research summary with citations to the strategist, writer, or implementation role. Keep the evidence specialist from silently becoming the decision-maker.

Review before action

A separate reviewer checks whether the final artifact still reflects the research summary with citations. The human chair approves publication, submission, or deployment.

Approved sourcesControlled evidence set

NotebookLMFindings + citations

Research summary with citationsClaims + gaps + conflicts

Round Table rolesInterpret + create + test

Human chairApprove or revise

Need a stronger role brief? Use the free Prompt Generator. Need more adversarial review? Open the Council of Agents protocol. Need to route the research summary with citations across several tools? Continue to the Multi-AI workflow. Before designing large research summary with citationss, check the current NotebookLM system limits.

The moment the difference became obvious

A website update is not only a writing task

Imagine maintaining a large educational website. A product changes. Dozens of pages may contain outdated claims. The obvious task is “update the article,” but the real project is much larger.

Someone must find every affected page, identify the claims that changed, locate current evidence, decide what should be rewritten, preserve the site’s voice, edit working files, validate links and structured data, inspect the mobile layout, and prepare a safe deployment.

With a normal chat workflow, AI can help with each step. But the human remains the go-between for all of them.

Before · AI-assisted but human-routed

The person is still the production line

Open one page and ask for a review.
Copy the findings into a research tool.
Move verified facts into a writing model.
Apply edits manually.
Run link, schema, accessibility, and layout checks.
Repeat the process page by page.

Updated Round Table

The human chair reviews visible work

The inventory specialist finds affected pages.
The research specialist verifies changed claims.
The content specialist drafts bounded revisions.
The implementation specialist edits an isolated copy.
The QA specialist tests links, schema, layout, and analytics.
The chair returns a change list for human approval.

Human goalOutcome + limits

ChairPlan + role briefs

SpecialistsResearch + drafts + files

ReviewerClaims + QA + risks

Review summaryChanges + supporting evidence

Human approvalDeploy or revise

The human moves from being the entire production line to being the editor of the system.

Sometimes one agent is still the better answer

Complexity must earn its place

When to use one agent versus a multi-agent Round Table.
Use one AI conversation when…	Use an AI Round Table when…
The task is short, reversible, and easy to inspect.	Several professional priorities genuinely conflict.
All work depends on the same small context.	Specialists need different sources, tools, or separate contexts.
The cost of coordination exceeds the benefit.	Independent work can run in parallel or needs independent review.
You need one coherent draft more than competing views.	The project must create several reviewable artifacts.
The human can verify the result immediately.	The workflow repeats often enough to justify a reusable system.

Rule of thumb: do not create a six-agent system to answer a question that one well-designed prompt can solve.

How to expand the AI Round Table beyond one conversation

Eight stages · one accountable human

1. Define the outcome

Name the final artifact, acceptance criteria, sources, deadline, and actions that cannot happen automatically.

2. Choose the seats

Choose the fewest specialist seats that create a real division of labor. Give every seat explicit exclusions.

3. Separate context and tools

Give each specialist only the files, sources, tools, permissions, and background it actually needs.

4. Decide what can run independently

Parallelize research, extraction, alternatives, or tests that do not depend on one another. Keep true dependencies sequential.

5. Require visible disagreement

Make specialists identify conflicting assumptions, missing evidence, and the cost of choosing one recommendation over another.

6. Run an independent review

Use a separate reviewer to check claims, execution, omissions, edge cases, and the stated acceptance criteria.

7. Chair the decision

The human chair selects the recommendation, records unresolved uncertainty, and writes a concise decision summary.

8. Require human approval

Publication, deletion, purchase, submission, deployment, and external communication remain blocked until a person approves them.

Free Prompt 2: design the workflow before choosing the tools

Copy this prompt to plan the workflow

This prompt prevents the most common mistake: choosing a platform first and inventing work for it afterward.

AI Round Table Workflow Prompt

Design the lightest controlled AI advisory-Round Table workflow that can complete the project below.

First decide whether this project needs:
A. one model with several roles,
B. several AI tools with structured handoffs,
C. or an agent tool that can complete clearly limited tasks.
Do not recommend a more complex level unless it removes a specific bottleneck.

Return:
1. The human problem the workflow is solving.
2. The smallest useful set of roles.
3. A role card for each specialist: mission, exclusions, inputs, allowed tools, prohibited actions, required artifact, verification method, stop condition, and escalation rule.
4. Which work is independent, which is sequential, and why.
5. The compact artifact passed between stages.
6. The chair's responsibilities.
7. The independent review gate.
8. Maximum retries and failure handling.
9. Actions requiring human approval.
10. A final acceptance checklist.

Project: [DESCRIBE THE PROJECT]
Available tools: [LIST THEM — OR SAY NONE]
Sources and files: [LIST THEM]
Sensitive or irreversible actions: [LIST THEM]
Final deliverable: [DEFINE IT]

One example: Gemini leads the discussion, Antigravity handles assigned tasks

An example stack · not a requirement

This combination is useful when a project needs tool-assisted research plus parallel or scheduled execution. It is one possible setup, not the definition of the Round Table.

Gemini leads the discussion

Frame, ground, and synthesize

Use Gemini to translate the goal into role briefs, work with multimodal or web-connected inputs where appropriate, and reconcile the returned artifacts.

Explore the Gemini Orchestration Hub →

Antigravity handles assigned tasks

Delegate, run, collect, and validate

Use projects, subagents, scheduled work, artifacts, and hooks when specialists must work on files or systems and return results for review.

Official Antigravity overview →

NotebookLM or another evidence workspace

Keep claims tied to the source set

Use a source-grounded workspace when the central risk is unsupported interpretation rather than execution.

Use NotebookLM as a research workspace →

Other tool combinations can work too. A ChatGPT or Claude project can serve as the chair; another search tool can handle discovery; a coding or browser agent can handle implementation; and an isolated model can perform review.

Three Round Tables you can copy

Change the seats to match the task

Content Round Table

From source to publishable asset

Researcher → Story Architect → Writer → Ruthless Editor → Audience Advocate → Chair.

Research Round Table

From evidence to defensible conclusion

Search Strategist → Source Analyst → Methods Auditor → Contrarian Reviewer → Synthesis Chair.

Product Decision Round Table

From idea to controlled recommendation

Customer Voice → Product Strategist → Technical Feasibility → Financial/Risk Reviewer → Devil's Advocate → Chair.

Do not let the Round Table vote. The chair must make a decision, explain the trade-off, and preserve any uncertainty the human still needs to resolve.

Agentic does not mean absent

The human chair remains responsible

The goal is not maximum autonomy. It is less message carrying, less repetitive coordination, and more time for the decisions only a person should make.

Least access required

Backups or version control

Explicit source requirements

Maximum retry budget

Reviewer differs from implementer

Logs and artifacts preserved

Low-confidence escalation

Human approval for consequential action

The human chair decides

Outcome, values, risk, and permission

The system can collect evidence and perform limited assigned work. It cannot decide what your audience deserves, what trade-off is acceptable, or whether a consequential action should happen.

Agentic means the human spends less time carrying information and more time exercising judgment.

What to explore next

Related systems

Adversarial role design

Council of Agents

Go deeper on debate, self-critique, chair design, and red-team protocols.

Explore Council of Agents →

Cross-tool routing

Multi-AI Workflow

Route evidence, reasoning, production, and review to the tool that fits each job.

Explore Multi-AI Systems →

Source-grounded work

NotebookLM Research Workspace

Use NotebookLM to ground the Round Table in cited sources before adding more tools.

Explore the research workspace →

Sources and testing notes

Primary documentation · reviewed July 30, 2026

This guide separates documented product capabilities from editorial workflow recommendations. Product-specific claims link to official documentation; the role design, handoff formats, review gates, and human-approval model are the publication’s recommended operating pattern.

NotebookLM

Google’s official NotebookLM help explains source-grounded research and how sources are used in responses.

Official help →

ChatGPT

OpenAI documents Projects as workspaces for chats, files, instructions, and longer-running efforts.

Official help →

Claude Code

Anthropic documents subagents with separate context, prompts, tool access, and permissions.

Official docs →

Antigravity

Google documents subagents, scheduled tasks, artifacts, and hooks for limited task execution workflows.

Official overview →

Execution hooks

Google’s hooks reference explains how checks and custom scripts can run at defined execution points.

Official docs →

Testing note: The free role prompt and workflow structure were reviewed for clarity, portability, and human-control requirements. Tool interfaces, limits, and product names can change; verify current capabilities before automating consequential work.

Frequently asked questions

AI Round Tables and multi-AI workflows

What is an AI Round Table prompt?

It is a prompt that assigns several specialists narrow roles, requires independent recommendations and visible disagreement, and gives one chair responsibility for the final decision.

Does the free prompt work in ChatGPT, Claude, and Gemini?

Yes. The prompt works in any capable model that can follow separate role instructions and a multi-round discussion.

Is a role prompt the same as a real AI agent?

No. A role prompt simulates several perspectives inside one conversation. An execution agent can have separate context, tools, permissions, tasks, artifacts, and execution state.

Do I need more than one AI tool?

No. Add another tool only when it contributes a meaningful capability, such as grounded evidence, current web discovery, independent review, file execution, or recurring checks.

What is the difference between the older and updated Round Table?

The older version solved one-dimensional feedback by creating five specialist perspectives. The updated version keeps that prompt and adds structured handoffs, tool routing, independent verification, optional execution, and human approval.

How many seats should the AI Round Table have?

Use the fewest seats that produce a real division of labor. Four to six active seats is a useful default, but a smaller Round Table is often better for a narrow task.

When should specialists run in parallel?

Parallelize work that does not depend on another specialist's result, such as separate source reviews, alternative analyses, extraction from different files, or independent tests.

When is one agent better than a multi-agent workflow?

Use one agent when the task is short, reversible, easy to inspect, and based on one shared context. An AI Round Table adds value only when specialization, isolation, parallel work, or independent review outweighs the coordination cost.

Where do Gemini and Antigravity fit?

They are one possible setup. Gemini can lead the discussion and combine the findings, while Antigravity can carry out assigned tasks through projects, subagents, artifacts, scheduled work, and hooks. Other tool combinations can follow the same Round Table process.

How do I use NotebookLM inside an AI Round Table?

Load the approved source set into NotebookLM, extract supported findings with citations, record conflicts and evidence gaps, and pass a concise research summary to the strategy, writing, or review seats. The Round Table coordinates the next steps while the human retains approval.

How do I prevent unsafe autonomous changes?

Use least-privilege access, isolated working copies, backups or version control, source requirements, retry limits, independent review, preserved artifacts, and human approval before publication, deletion, purchase, submission, deployment, or external communication.

Keep the workflow. Choose how far to build it.

Start with the free research kit, or use the Multi-AI Systems templates when you need reusable handoffs, review gates, and approval checklists.

Get the free Starter Kit Explore Multi-AI Systems