Hello OpenAI Product Team,
I would like to propose what I believe could be a major next step for ChatGPT Work: Teach-by-Demonstration workflows.
ChatGPT Work is already evolving beyond a conversational assistant into a system capable of completing real work across files, connected applications, browsers, scheduled tasks, computer use, and parallel subagents.
The remaining gap is not simply more intelligence. It is reducing the amount of work required to teach ChatGPT how a specific user actually works.
Today, users generally have to describe workflows through prompts, instructions, skills, or configuration. For many real-world tasks, however, demonstrating the workflow once would be substantially easier and more accurate than explaining every step in natural language.
I propose a feature tentatively called:
Teach ChatGPT This Workflow
Core concept:
A user performs a task once while ChatGPT observes the authorized interaction.
ChatGPT identifies:
- applications and websites used
- sequence of actions
- inputs and outputs
- decision points
- recurring rules
- exceptions
- required confirmations
- sensitive actions
- dependencies between steps
ChatGPT then converts the demonstration into a reusable workflow that the user can inspect and edit before activation.
For example:
A user demonstrates how they process their inbox every morning.
ChatGPT observes that the user:
1. Reviews unread messages.
2. Prioritizes messages from specific people.
3. Separates meeting requests from general correspondence.
4. Checks the calendar before responding to scheduling requests.
5. Drafts responses using different styles depending on the sender.
6. Adds relevant follow-up items to a task list.
7. Leaves sensitive or ambiguous messages for manual review.
Instead of requiring the user to manually describe all of these rules, ChatGPT would create an editable workflow from the demonstration.
The user could then say:
“Run my morning inbox workflow every weekday.”
The workflow could operate through ChatGPT Work, Scheduled Tasks, connected apps, Computer Use, and subagents.
I believe several design principles would make this particularly valuable.
1. Demonstration should generate an editable workflow, not a black box
Users should be able to inspect:
Trigger
Inputs
Actions
Decision rules
Exceptions
Required approvals
Outputs
Applications used
This would make the automation understandable, auditable, and correctable.
2. Separate observation from authorization
Watching a user perform an action should never automatically grant permission to repeat that action.
ChatGPT should separately request authorization for sensitive capabilities such as:
sending messages
deleting files
publishing content
making purchases
changing account settings
accessing confidential information
3. Include confidence and exception handling
When ChatGPT encounters something outside the demonstrated pattern, it should not blindly continue.
It should be able to:
continue automatically when confidence is high
ask the user when confidence is uncertain
pause when an action is consequential
learn from the correction after the user resolves the exception
4. Allow workflows to improve over time
After repeated executions, ChatGPT could suggest:
“I noticed that you usually handle these messages differently. Would you like me to update this workflow?”
The user would approve any permanent change.
This would create controlled learning rather than uncontrolled behavioral drift.
5. Combine demonstrations with natural-language instructions
The strongest system would allow users to both demonstrate and explain.
For example:
“Watch how I prepare this weekly report. Use the same process every Friday, but always ask me before sending the final version.”
This combination of demonstration and language could dramatically reduce workflow configuration complexity.
6. Allow subagents inside workflows
ChatGPT Work already has the foundation for parallel agent collaboration.
A demonstrated workflow could automatically identify parallelizable components such as:
Research Agent
Data Analysis Agent
Document Agent
Verification Agent
Communication Agent
A coordinating agent could then combine their outputs before completing the workflow.
7. Provide a Workflow Library
Users should be able to see and manage reusable automations in one place.
Each workflow could display:
Last execution
Next execution
Applications accessed
Permissions
Average runtime
Failures
Pending approvals
Execution history
Version history
8. Make workflows portable across devices
A workflow created on desktop should be manageable from mobile and web.
Users should be able to start a long-running workflow from their phone, allow ChatGPT Work to continue remotely where technically possible, and review the outcome later.
Why this matters strategically:
AI assistants are rapidly moving from answering questions to executing workflows.
The long-term competitive advantage will not simply come from which model reasons best. It will increasingly come from which system learns how an individual or organization actually operates with the least configuration effort.
Natural-language prompting reduces the cost of programming.
Teach-by-Demonstration could reduce the cost of automation itself.
ChatGPT already has many of the necessary components:
ChatGPT Work
Scheduled Tasks
Computer Use
Cloud Browser
Connected applications
Skills
Memory
Subagents
The opportunity is to unify these capabilities into a system where users can teach ChatGPT real workflows simply by showing it what they do.
A possible product progression could be:
Phase 1
Record a demonstrated workflow and convert it into editable steps.
Phase 2
Allow the workflow to become a Scheduled Task.
Phase 3
Add exception learning and user-approved workflow updates.
Phase 4
Automatically decompose complex workflows across specialized subagents.
Phase 5
Create a personal workflow operating layer where ChatGPT can manage recurring digital work across authorized applications.
The objective should not be fully autonomous software that acts without oversight.
The objective should be something more useful:
A trustworthy AI teammate that can be taught the way a skilled human colleague is taught — by explanation, demonstration, correction, and repetition.
I believe this capability could significantly strengthen ChatGPT Work as an operating layer for real knowledge work and make advanced automation accessible to users who cannot or do not want to build traditional automation pipelines.
Thank you for considering this proposal.