ChatGPT dots: my founder experience and practical guide

I am Akash Singh, co-founder and CTO of CV Infotech. My work moves between development, client discussions, project documents, email, product decisions and the details that connect them. A useful answer often needs more than the message in front of me. It needs the right history.

That is where ChatGPT dots has become useful in my work. I use my dot in ChatGPT to work through long conversations, compare documents, research questions and follow development work. I can ask for a short update, then go deeper into the evidence behind it.

The experience has been genuinely useful. It has also taught me to be specific about access, approval and what counts as finished. I do not want an assistant that sounds confident about everything. I want one that can show me what it checked, explain what is missing and help me move the work forward.

This is my first-person account of those workflows. It is not an official OpenAI guide or a claim that every account has the same features. Connections, permissions, plans and product availability can affect what is possible.

What are ChatGPT dots?

For my work, a personal assistant needs context across an ongoing task. If I ask whether a project document reflects the client's latest feedback, a polished summary of one file is not enough. The assistant needs to compare the document with the relevant discussion, separate decisions from proposals and identify questions that remain open.

OpenAI introduced dots on September 29, 2026. Its official name is “dots”; an individual assistant is “your dot”. OpenAI describes them as agents for ongoing responsibilities across connected tools. (OpenAI announcement)

A dot can coordinate work with ChatGPT Work and Codex. That distinction matters in a coding example: the assistant can organise the work, while a separate coding task carries out the implementation. (OpenAI dots documentation)

I use my dot as a personal assistant in ChatGPT. The practical value comes from combining conversation with the sources and tools I have authorised it to use. It should not have unrestricted access to everything, and a connected account does not automatically mean every action is permitted.

My aim is straightforward: spend less attention reconstructing context and more attention deciding what to do with it. I have not measured a reliable hours-saved figure, so I do not present one here.

Making our dots feel familiar: names, my dog and animation

Across our team, we have used names such as Shaktiman, Jarvis, Veronica, Kimchi and Apple for our dots. I find a familiar name more natural to address than an unnamed assistant.

I have also personalised a dot's appearance with my pet dog's face and animation. It is a small detail, but it makes the experience feel more personal to me. The name and appearance are part of how I recognise the assistant, not evidence that it has more capabilities or access.

These are examples from our use, not a claim that every named dot shares the same memory, tools or responsibilities. Each workflow still needs its own clear instructions. You can make an assistant feel familiar without treating it as a person or assuming that it understands information you have not provided.

Features and how I use them

My strongest hands-on uses are project research, documents, email, development coordination and outreach preparation. Naming and dog-inspired appearance add a personal touch. The following practical guide also explains documented options, including voice and other messaging channels, that I have not all evaluated personally.

How to start using ChatGPT dots

These practical instructions reflect OpenAI's documentation checked on 28 October 2026. They explain documented options; my hands-on experience is identified separately in the examples above and below.

Open dots in the ChatGPT mobile app, desktop app or a desktop browser. Follow the introduction, choose a name and appearance, and connect only the apps needed for your first task. You can add connections later. Mobile setup uses the app, not mobile web. (Getting started)

For a first request, I recommend a small, checkable outcome: “Read this approved document and the latest relevant discussion. Give me the open questions, with sources. Do not send messages or change the document.” Once the result is useful, add a continuing responsibility rather than giving a vague instruction to handle everything.

Names, appearance and animated pets: what to customise

You can personalise a dot through its name and profile controls. Available appearance options may include characters or pets. Use the current setup guide for the controls available in your app; these interfaces can change.

My dog-inspired version is my own experience. A useful creative brief would be: “Use this photo of my dog as a visual reference. Keep the face recognisable and the animation gentle.” This is a suggested brief, not a tested photo-upload-to-avatar tutorial or the exact prompt I used.

ChatGPT's broader Pets feature also offers animated companions. Desktop and web controls are different: on desktop, Settings > Pets lets you choose a pet or begin Create pet; after creation, refresh the picker and select it. The desktop /pet command shows or hides the floating companion. On web, supported accounts use Settings > Personalization > Pet, with an Upload pet option for a compatible sprite sheet. Desktop custom pets do not automatically sync to web. Reduced-motion settings can replace animation with a still frame. (Pets guide)

A pet can be selected for a dot, but dot-profile selection, the desktop floating companion and web Work-chat pet controls are different surfaces. Personalisation makes the experience feel more familiar; it does not change what work the assistant can perform.

Give it a continuing responsibility, not just a prompt

For recurring work, specify the sources, outcome, timing, time zone, notification condition and destination. Ask the dot to confirm the saved schedule, then review Scheduled. Supported event monitoring is another option; connecting an app alone does not create it. (Tasks and memory)

Here is a practical example to adapt: “Each weekday at 10 AM in my time zone, review the approved project sources. Report material changes, blockers and decisions needed. Link the evidence. Do not assign work or contact the team. Confirm the schedule and tell me if a source cannot be read.”

For one ongoing investigation, ask for a verified outcome and the next step rather than automatically creating many overlapping schedules. My preference is for meaningful updates, not repeated reassurance.

Review memory, delegated work and progress

Dots can use relevant memory, conversation context and saved notes, and can delegate work. That does not give every delegated task every past conversation. Ask which sources informed an answer and supply the important record when needed. A new task's context is not a complete shared company memory. (Tasks and memory)

In the desktop profile, Activity opens task progress and results. Review the actual output and any request for input. A “completed” run still needs a check against the outcome you asked for. This is how I distinguish a useful result from a task that merely stopped running. (Controls)

Connect apps and choose cloud or local work

Treat app access, messaging and computer access as separate choices. Review the connected account and permitted actions for each plugin. For example, reading Gmail and sending from Gmail are different operations. If a connection expires, reconnect through its supported flow rather than sharing credentials in a chat. (Computers and apps)

The cloud browser has its own sessions; your personal browser login is not automatically copied. Open the dot profile's Computers area to inspect the cloud computer. Use Take over when you need to interact, then Return control. For local work, the desktop profile offers Your computer > Allow access and a confirmation. Keep that computer online with ChatGPT open while it is needed. (Computers and apps)

I choose the environment based on the source or tool the work needs. “Use cloud” is useful direction, but it does not transfer an existing local task or make a private app session appear there.

Talk to your dot and choose where updates arrive

Open the conversation's phone button, or desktop profile's Call option, for supported voice conversations. You can keep typing during a call; ending the call does not itself cancel assigned work. The profile's Add control shows available contact methods, including Slack and Teams. These reach the same dot, but visible chat histories are not all mirrored. (Messaging guide)

On supported iPhones, a Your dot widget and the Open your dot Shortcuts action offer quick access, with messaging, dictation or call modes. Availability depends on the app and account. I have not evaluated every voice, widget or messaging option personally. (Messaging guide)

Messaging options vary by account, region and rollout. Check the current official messaging guide for availability before planning a workflow around texting. I have not evaluated every messaging option personally.

For notifications, tell it what deserves your attention and where to deliver the update. A request for a report every morning is different from an alert only when a decision is needed.

Change permissions, pause work and keep control

Custom rules are available under Settings > Personalization > Custom rules (under Permissions) on desktop/web and Customize > Custom rules on mobile. Add a rule for a specific action and choose the required handling. These rules do not create app access or override built-in safeguards. (Controls)

To stop, distinguish the main activity, delegated tasks and future schedules. Pause affects current main work; inspect delegated tasks in Activity and recurring work in Scheduled separately. Stopping work does not undo actions already completed. I keep those distinctions explicit when changing direction. (Controls)

Turning long conversations into clear project decisions

A project can accumulate decisions in email threads, chat exports, shared documents and task updates. The most recent message may correct an earlier one. A new suggestion may sound definite even though nobody approved it.

I use the assistant to bring those sources together and answer specific questions. What changed? Which version is current? What was agreed? What is only proposed? Who needs to decide the next step?

The result I want is a concise explanation with supporting records, not a summary that removes all uncertainty. When two sources disagree, the useful response is to show the conflict rather than choose the more convenient version.

This approach matters when preparing client-facing scope documents. A requirement can be present somewhere in a long document and still be too vague for the team to implement. I ask for the missing conditions, exceptions, ownership and acceptance checks, and require the revised copy to be checked against the original evidence before it goes back to the client.

Evidence before action

1. 1
   Define the question

   Name the decision and the source that controls it.

2. 2
   Read authorised sources

   Check the current thread and document versions.

3. 3
   Compare the evidence

   Keep requirements, proposals, reports, tests and acceptance distinct.

4. 4
   Review the finding

   Resolve conflicts or ask for a decision.

5. 5
   Take the permitted next step

   Check the outcome and preserve the evidence.

Missing or conflicting evidence? Hold the action, clarify the question, then return to the sources.

Approved requirement

What was authorised for delivery.

Proposed change

A suggestion awaiting a decision.

Reported implementation

A claim about work performed.

Tested result

Observed behaviour within the test scope.

Client acceptance

A separate decision by the client; testing does not establish acceptance.

Using connected email without treating every message as an instruction

Email is one of my practical uses. I ask for relevant threads, the latest replies and unresolved questions, rather than relying only on unread messages.

A useful email review checks whether someone has already answered. It distinguishes a genuine question from a routine acknowledgement, and a request from a decision. It should also identify the correct sending account and recipients before anything is sent.

For a tightly defined workflow, I can authorise replies to named team members on a specific subject. That does not authorise new commercial commitments or replies to every person copied on the thread. I keep unsupported scope, pricing and deadline decisions with me.

The benefit is continuity. The assistant can help me understand the question in the context of the project, instead of producing an isolated email that sounds helpful but contradicts an earlier agreement.

Reviewing code and development work more carefully

I also use the assistant in development workflows. It can help inspect source changes, trace a reported problem, prepare a targeted fix and work through tests in an authorised environment.

One lesson is especially important: a change in code is not proof that the user-facing problem is fixed. A passing unit test is not a live deployment. A developer's completion message is not client acceptance.

In one staging workflow, an import began working while the next step, opening the imported item for editing, still failed. Checking the actual screen and tracing the error exposed a separate problem in the editor's data-loading path. The fix needed its own tests and a read-only check of the affected page.

That is the level of detail I want from AI-assisted development. The assistant should name the environment, explain which checks ran and state which actions were deliberately not performed. It should preserve unrelated work and avoid turning a small bug fix into an unapproved production change.

I still need engineering judgement. AI can help investigate and implement, but I do not treat a confident explanation or a green task status as a substitute for verification.

Keeping progress reports useful

When several projects are active, I want short updates that answer four questions: what changed, what has been verified, what is blocked and what happens next.

I use scheduled progress checks for that purpose. The report should be honest when nothing has changed. It should also distinguish an inaccessible source from evidence that no work happened.

This is particularly important with meetings and task systems. A generated recap is not the same as the original conversation. An email notification may describe an update, but it does not prove the current state of the live task board. I prefer a clear coverage gap over an invented status.

Regular reporting is valuable when it reduces the need to reconstruct the same context again. It becomes noise if it repeats vague reassurance or hides a blocker behind the word “ongoing”.

Finding qualified leads and organising outreach without a separate CRM for this workflow

I have also used my dot to help with prospect research and email outreach. For this workflow, I use the dot with Gmail and a shared outreach register rather than a separate CRM. The useful part is not just writing an email. It is keeping the research, contact history and next decision together.

I ask it to look for businesses with a concrete, relevant website or ecommerce need. A generic observation that a site could look better is not enough. The research should explain what was actually found on the current website, why it may matter and whether we have relevant experience to discuss.

When I enable outreach, before a contact becomes eligible, the workflow checks the appropriate public business contact, recipient location and applicable outreach requirements. Publicly finding an address is not the same as having permission to contact it. It also checks previous email, objections, unsubscribes and shared team ownership so two people do not independently approach the same business.

The email should be short and specific: one relevant observation, a verified example of our work where appropriate and an easy question to check whether the issue matters to the recipient. I prefer that to an ambitious pitch that assumes a problem or promises results before we have spoken.

Timing and records matter too. The workflow checks the recipient's local business hours and records a verified sent message in the shared register. An attempted send is not a confirmed send, and a sent record does not prove inbox placement. Human interest, referrals, objections, automatic replies and bounces belong in different categories.

This gives me a practical way to organise this part of outreach without adding a separate CRM to it. It is not a claim that a dot replaces every CRM function. A business may still need a CRM for its wider pipeline, forecasting, account management or reporting. Sales conversations and commercial decisions remain with us, and automatic follow-ups are not part of this workflow.

I only want outreach to run when its checks can be completed. If contact eligibility, connection access, previous activity or the day's sending total cannot be verified, sending should wait. The benefit is a better-prepared, traceable introduction—not a guaranteed reply, sale or daily volume.

Researching and preparing work for another person

The assistant also helps me move from a question to a usable handoff. That may mean researching public information, checking keyword data, preparing a document and writing implementation instructions for the person who will do the next part.

For this content project, for example, the research includes search intent, related questions and keyword estimates. Those estimates guide what to investigate. They do not prove that a product name is official, that a page will rank or that a business result will follow.

A good handoff includes the actual copy, the reason for the page, the evidence behind the claims and a clear implementation checklist. The next person should not need to guess which sentences are public copy and which are internal instructions.

What has not worked perfectly

Connected work depends on the connection. I have encountered integration and MCP startup problems that were not solved simply by asking again or granting broader access. When a tool is unavailable, the assistant needs to say so rather than imply it completed the task.

Document review also needs scrutiny. Missing detail can survive an initial pass. I ask for a second review against the original evidence, especially before sharing an important scope document with a client.

These limitations do not remove the usefulness of the assistant. They shape how I use it. I give it a bounded task, ask for sources, distinguish a draft from an approved decision and require verification for consequential work.

The working rules that make the difference

Start with one workflow. Choose something you can review, such as summarising a project thread or comparing two document versions.

Define the sources. Tell the assistant which documents or accounts are authoritative and which material is only background.

Define the action boundary. Reading, drafting, sending, editing and publishing are different permissions. State what is allowed and what needs your approval.

Ask for evidence. A useful result includes relevant links, a clear status and any important uncertainty. It should not expose private material just to make the answer look thorough.

Verify the outcome. For an email, check that it was sent from the intended account. For a document, review the updated version. For code, check the relevant tests and environment. For a live change, confirm the deployed behaviour separately.

My view as a founder and CTO

I see strong value in using a personal AI assistant for the work between tasks: finding context, reconciling details, preparing a decision and carrying a thread forward. Those are real parts of running a development company.

What makes the experience good is not a promise that the assistant can do everything. It is the ability to give it a specific job, connect appropriate sources and inspect the result. Used that way, ChatGPT has become a practical part of how I review information and manage technical work.

If you are exploring this for your own business, start with a real question you already need to answer. Make the result easy to check. Expand the workflow only after you understand where it helps and where a person still needs to decide.

Frequently asked questions

When did OpenAI launch dots?

OpenAI published its dots announcement on September 29, 2026. This article describes my own use and was reviewed on October 28, 2026; it does not claim I began using the product on launch day. (OpenAI announcement)

Are dots available on every ChatGPT plan?

No. OpenAI documents eligibility and rollout conditions, so check the current access page for your account and region. Do not assume that a plan upgrade alone guarantees immediate access. I link to the official page instead of maintaining an unofficial price list here. (OpenAI dots documentation)

Can I use dots on mobile?

OpenAI added mobile creation in its October 5 to 9 update. Check the current instructions rather than relying on launch-day setup articles. (OpenAI October update)

Can ChatGPT work as a personal assistant for a business owner?

It can help with research, summaries, writing and connected workflows when the relevant features and authorised sources are available. My experience is strongest when the task has clear boundaries and a result I can verify. It does not remove responsibility for business decisions.

Can it read all of my emails and documents automatically?

Do not assume that. Access depends on connected services, permissions and the capabilities available in the current session. I ask which sources were actually read and which were unavailable. I also limit the information included in public or third-party outputs.

Can it send emails or make changes for me?

Some connected workflows support actions as well as reading. I keep those permissions specific: the intended account, recipients, purpose and limits. A request to draft a reply should not silently become permission to send it or make a new commitment.

Can I use it to review code?

Yes, I use AI-assisted code review and debugging as part of my workflow. The result still needs engineering review and relevant tests. Source inspection, passing tests, a staging check and a production deployment are separate pieces of evidence.

Will it work while my laptop is off?

That depends on the task. A workflow using a connected cloud service may not need my laptop. A task that depends on a local browser, application or file still needs that environment to be available. I check the dependency instead of assuming all work is either local or cloud-based.

Is a custom MCP connection guaranteed to work?

No. MCP integrations can have authentication, configuration and compatibility problems. I treat a successful real tool call as evidence of a working connection, not the existence of a configuration screen or a broad permission grant.

How much time has this saved you?

I have not established a measured time-savings figure. The practical benefit I can describe is having relevant information organised into decisions, documents and reviewable next steps. A reliable number would require a consistent before-and-after measurement.

Is this an official OpenAI recommendation?

No. This is my personal account as the co-founder and CTO of CV Infotech. Product names belong to their respective owners. This article does not imply an OpenAI partnership, certification or endorsement.

Can I give my dot a name and a personality?

You can choose a name and appearance, and explain your preferred writing style and working habits. Our names include Shaktiman, Jarvis, Veronica, Kimchi and Apple. A name is a personal label, not proof of a separate skill set or shared memory across different people's dots.

Can I use my dog's face and animation?

That is part of my own personalised experience. For your version, distinguish the dot's appearance from the desktop or web Pets feature and follow the current controls for that surface. Use an image you are entitled to use, review the result and avoid assuming a pet created on one surface will appear everywhere automatically.

Can dots find leads and send emails without a CRM?

For my outreach workflow, I use a dot with Gmail and a shared register instead of a separate CRM. It helps organise research, qualification, correspondence checks and records. That is not a replacement claim for every CRM function. Contact eligibility, permission to send, verified records and human sales judgment still matter.

Why does my dot seem to stop or ask for help?

A source may be unavailable, an account may need reconnection or an action may require your decision. Open the task's actual result and request for input. Ask what is missing instead of assuming that more permissions or another “continue” message will fix the underlying issue.

Does this guide cover every possible dot integration?

No fixed article can list every changing plugin or account-specific capability. This guide covers the core experience, practical controls and my main workflows. For a particular integration, check the current supported tools, connected account and actual operation you need. I update the guide when verified changes affect the instructions.

Read how I review project emails and documents.

For a related engineering checklist, see how to review an AI-assisted change before release.