The hidden part of the new ChatGPT is not one secret GPT-5.6 toggle. It is the growing system around the model: Skills for repeatable instructions, Sites for hosted outputs, Appshots for visual context, plugins for connected tools, Scheduled tasks for recurring work, and Browser or Computer Use for actions beyond chat.
Rob The AI Guy gives a fast tour of these capabilities and demonstrates a content workflow that researches a topic, applies his script process, prepares a thumbnail direction, and finds B-roll. This guide organizes that walkthrough into a safer setup: what each feature does, when it helps, what is still in beta, and which permissions should stay narrow.
Video and walkthrough credit: Rob The AI Guy. The feature and safety details below are checked against OpenAI's official ChatGPT documentation.
Source Note
Rob's video is the source for his interface tour, content-production example, personal recommendations, and timing claims. OpenAI's announcement and current documentation are the factual spine for feature definitions, supported surfaces, permissions, rollout status, and safety boundaries.
Labels matter here. Official means OpenAI documents the feature. Public beta means behavior and availability may change. Creator workflow means Rob demonstrated it on his account, not that every plan or region has the same controls. The prompt-optimizer segment promotes Rob's external product; it is not a built-in ChatGPT capability and is not required for the workflow in this article.
Link Map
| Feature | Official link and status | Use it for | Main caution |
|---|---|---|---|
| ChatGPT Work | Product announcement and quickstart - official | Research, analysis, documents, presentations, spreadsheets, and Sites. | Connected, scheduled, local, and Codex tasks can have different permission boundaries. |
| Skills | Skills and Plugins and Build skills - official | Reusable SOPs, templates, standards, examples, and task-specific instructions. | A vague process becomes a reusable vague process. Test the skill before sharing it. |
| Sites | Sites documentation - public beta | Hosted websites, dashboards, internal tools, web apps, and games. | Every deployment URL is production. Save and review a version before deploying or sharing. |
| Appshots | Appshots documentation - macOS desktop | Share a visible app window plus available text as task context. | It may capture sensitive on-screen information and may miss off-screen document content. |
| Browser | Browser documentation - official | Current research, website interaction, local web QA, screenshots, and visual comments. | Page content is untrusted. Site permission does not make instructions on the page safe. |
| Computer Use | Computer Use documentation - supported regions, macOS and Windows | Scoped workflows that depend on a graphical desktop app or cross-app interaction. | On Windows it takes over the active desktop. Keep sensitive apps closed and approvals narrow. |
| Scheduled tasks | Scheduled tasks - official | Recurring work from Chat or Work, optionally using a tested skill. | Local-file tasks need the desktop app running and the computer available. |
| Plugins and MCP | Plugins documentation - official | Install skills, apps, connectors, MCP tools, hooks, and task templates. | Review the publisher, requested scopes, write actions, hooks, and external-service permissions. |
| Plan and region access | Feature availability - time-sensitive | Confirm what your account and workspace actually include. | Do not assume Rob's interface exactly matches every account, region, or managed workspace. |
What Is Actually New
The model is becoming one component inside a work environment. Chat is for conversation. Work is for research and finished deliverables. Codex is for software work with codebase context and developer tools. Skills define repeatable methods. Plugins connect services and tools. Scheduled tasks decide when a stable method runs. Sites publish a result.
That separation is more useful than calling the whole release a super app. It gives the user a routing decision: choose the lightest surface and narrowest permission set that can complete the job.
Skills: Turn Good Work Into a Process
Rob begins with the strongest feature in the walkthrough. A ChatGPT skill packages instructions and supporting resources for a recurring task. OpenAI says a skill can include its name and description, workflow instructions, templates, examples, brand guidance, schemas, and connected tools.
Rob asks ChatGPT to inspect how he normally works and suggest skills for repeated activities such as YouTube scripts, AI news research, thumbnail ideation, title optimization, channel audits, product launches, and landing-page copy. That is a useful discovery technique, but it should produce a review list, not silently install new behavior from your entire history.
Review only the memories and conversations I have allowed you to use. Identify tasks I repeat often enough to become skills. For each candidate, show the trigger, required inputs, workflow steps, expected output, verification checks, sensitive data involved, and tools it might need. Rank the top five by time saved and consistency gained. Do not create, install, or connect anything yet.
Build the first skill around a narrow task with an obvious result. A weekly brief, meeting follow-up, content QA checklist, or standard presentation format is easier to verify than a broad "run my marketing" skill.
Sites: Publish Without a Separate Stack
Sites lets ChatGPT create, host, refine, and share websites, web apps, dashboards, and games. It can start from a prompt or a compatible local project. The official docs also describe persistent records, uploaded files, workspace identity, and authentication-enabled Site shapes.
The important hidden detail is that every deployed Sites URL is a production deployment. OpenAI recommends saving a version before deploying when review matters. Keep access set to only you or the smallest intended audience until you have checked content, behavior, data handling, environment variables, uploaded files, and the shared URL itself.
- Describe the audience, purpose, behavior, data, and access requirement.
- Ask ChatGPT to save a reviewable version without deploying.
- Test empty, error, mobile, and permission states.
- Check that no prompt, secret, private file, or customer data appears publicly.
- Deploy the approved version and verify the production URL as an outside visitor.
The Settings I Would Not Enable Blindly
Rob recommends enabling automatic approval review, full access, Browser, Computer Use, and several app integrations. The capabilities are real, but "turn everything on" is the wrong default for a business machine.
OpenAI defines full access as running without normal sandbox restrictions. Its enterprise guidance also says Work permissions, search, apps, plugins, source-system permissions, Codex sandboxing, MCP policy, and approval controls are separate. One toggle does not govern the whole system.
| Setting | Recommended starting point | Expand only when |
|---|---|---|
| Filesystem | One project folder or read-only access. | The task requires edits and you can review the changed files. |
| Browser sites | Allow only the named site for the current task. | You trust the site and understand the signed-in actions available. |
| Desktop apps | Approve one target app for one bounded flow. | The app contains no unrelated sensitive context and the flow is repeatable. |
| Plugins | Read scopes first; avoid write actions initially. | The read-only workflow is stable and the write action has an approval gate. |
| Full access | Leave off. | A specific task cannot work in a sandbox and the machine or VM is prepared for that risk. |
| Schedules | Draft-only output into a review queue. | Several supervised runs have succeeded without hidden assumptions. |
Appshots: Visual Context With Limits
An Appshot sends the frontmost Mac app window to a Work or Codex task. It can include the visible image and text the app exposes, sometimes including text outside the visible scroll area. That makes it useful for showing an error, design, email, settings panel, or current app state without explaining every detail.
Appshots are currently documented for macOS, not Windows. They also do not guarantee full document access. Google Docs, Gmail, Sheets, and Slides may provide only the visible screenshot. If complete structured content matters, use an authorized plugin for that source rather than assuming the Appshot contains everything.
Browser vs Computer Use
Use the built-in Browser first for websites and local web apps. It gives ChatGPT a separate browser profile and can open pages, click, type, inspect rendered state, capture screenshots, and verify results. Use the Chrome extension only when the task genuinely needs an existing Chrome profile or signed-in tab.
Use Computer Use when the task depends on a graphical desktop app or crosses apps that lack structured integrations. OpenAI explicitly recommends preferring a dedicated plugin or MCP server when one exists. Structured tools are easier to constrain and audit than visual clicking.
On Windows, Computer Use operates the active desktop and can move the pointer and type while it works. It cannot run unobtrusively in the background on the same session. For long tasks, use a separate device, a virtual machine, or a carefully scoped remote workflow rather than surrendering the main desktop with every personal app open.
ChatGPT Work: Delegate an Outcome
Rob's most concrete demo asks Work to choose an AI-news topic, propose a title, apply his script skill, find B-roll, and recommend a thumbnail workflow. The value comes from combining a goal, reusable method, current research, creative assets, and a handoff package for an editor.
The creator reports that the result took about two minutes versus roughly an hour manually. Treat that as his observed demo, not a guaranteed productivity benchmark. The real measure is whether the topic is relevant, sources are correct, B-roll rights are clear, the script matches the channel, and the editor can use the package without repair.
I publish practical AI news for [audience]. Research the strongest topic from the last [time window] using primary sources first. Propose three titles under [limit] characters. Apply my [skill name] skill to draft the script. Create a B-roll plan with timestamps, source links, usage-rights notes, and alternatives. Produce a thumbnail brief, but do not publish, download, purchase, or contact anyone. End with a verification table showing every factual claim and source.
Scheduled Tasks: Automate Stable Work
Scheduled tasks can run recurring work in the background and can invoke skills. OpenAI's clearest rule is excellent: skills define the method; Scheduled tasks define the schedule. If the workflow still needs frequent steering, improve the skill before automating it.
Web schedules can use uploaded context and connected tools but cannot work directly in a local computer folder. Desktop schedules can use local projects, but the computer must be on and the app running when local files are required.
Good first schedules produce reviewable information: a Monday research brief, overdue follow-up list, source monitor, draft release notes, or weekly analytics summary. Avoid automatic publishing, customer messaging, payments, deletion, or account changes until a human-approved version has proved reliable.
Plugins and MCP: Connect Tools Carefully
Plugins can include reusable skills, apps, MCP servers, browser capabilities, hooks, and scheduled-task templates. Rob demonstrates searching for services such as Stripe and using a vidIQ plugin to find outlier videos. Those examples show the upside of direct tool access, but each plugin creates a new trust relationship.
Before installing a plugin, record its publisher, authentication method, read and write scopes, external data destination, hooks, scheduled behavior, and revocation path. For financial, messaging, CRM, or publishing tools, begin with read-only analysis and require confirmation for every external action.
Four Prompts to Copy
1. Find the first skill worth building
Review the tasks I repeat in our allowed context.
Rank five skill candidates by frequency, time saved, and ease of verification.
For each one, list inputs, steps, output, checks, tools, and sensitive data.
Recommend one narrow first skill. Do not create or install it yet.2. Build and test the skill
Build a skill for [focused task].
Inputs: [files, links, fields].
Process: [ordered steps].
Output: [format and destination].
Always check: [acceptance criteria].
Never do: [prohibited actions].
Ask before: [external or irreversible actions].
Include one good example and three test cases.3. Turn the stable skill into a schedule
Use [skill name] every [cadence].
Use only these sources: [approved sources].
Create a draft in [review destination].
Include citations, changes since the last run, uncertainty, and failed sources.
Do not publish, send, purchase, delete, or change account settings.
Notify me when the draft is ready for review.4. Audit permissions before enabling more access
Review this workflow before it runs.
List every file, app, site, plugin, credential, and external action it may use.
For each one, recommend the narrowest permission and explain why it is needed.
Flag prompt-injection, privacy, payment, publishing, deletion, and account risks.
Produce an approval checklist. Do not change any setting.JQ AI SYSTEMS Setup Checklist
- Update and verify access. Check your plan, region, workspace policy, and desktop operating system.
- Choose one workflow. Avoid activating every capability just because it appears in Settings.
- Create a focused skill. Add examples, prohibited actions, and acceptance checks.
- Test manually three times. Use realistic inputs, including one failure case.
- Add one tool at a time. Prefer structured plugins over Computer Use where possible.
- Keep permissions narrow. Use project folders, named sites, named apps, and read-only scopes.
- Review the output. Check sources, edits, logs, and external actions before accepting the result.
- Schedule only proven work. Send outputs to a review queue before allowing delivery.
- Revoke unused access. Remove old app approvals, plugin connections, browser permissions, and public Sites.
Bottom Line
Rob is right about the larger shift: using ChatGPT only as a chat box leaves much of the new platform untouched. Skills, Work, Sites, Browser, Computer Use, plugins, Appshots, and Scheduled tasks can turn a conversation into a repeatable operating workflow.
The best setup is not the one with every toggle enabled. It is the smallest system that reliably completes one valuable task, exposes its sources and actions, asks before anything consequential, and can be reviewed by the person responsible for the result.
Sources
- Rob The AI Guy: New Hidden ChatGPT 5.6 Features You Did Not Realize Existed
- Rob The AI Guy on YouTube
- OpenAI: ChatGPT for your most ambitious work
- OpenAI: ChatGPT quickstart
- OpenAI: Skills and Plugins
- OpenAI: Build skills
- OpenAI: Sites
- OpenAI: Appshots
- OpenAI: Browser
- OpenAI: Computer Use
- OpenAI: Scheduled tasks
- OpenAI: Plugins
- OpenAI: ChatGPT Work Admin FAQ
- OpenAI: Feature availability and pricing