---
title: "ChatGPT, End to End: A Hands-On Guide From First Chat to Working System"
author: "Ecem Karaman"
source: "https://aiwithecem.com/guides/chatgpt-end-to-end"
published: 2026-09-16
tools: ["ChatGPT"]
topics: ["Setup","Prompts","Workflows"]
---

# ChatGPT, End to End

ChatGPT is now a chat app, a work agent, a coding tool, a website builder and an automation layer in one product. This guide shows you how to use each part, in the order you will need them.

**Last checked against [OpenAI's ChatGPT docs](https://learn.chatgpt.com/docs) on September 16, 2026.** Features change often and depend on your plan, platform, region and rollout. If your screen differs from this guide, your screen is the current version.

> **In this guide**
>
> 1. [How ChatGPT fits together](#how-chatgpt-fits-together)
> 2. [Set up in 10 minutes](#set-up-in-10-minutes)
> 3. [Choose a plan](#choose-a-plan)
> 4. [Choose a model](#choose-a-model)
> 5. [Chat, Work or Codex](#chat-work-or-codex)
> 6. [Prompts that work](#prompts-that-work)
> 7. [Give it the right context](#give-it-the-right-context)
> 8. [Research with sources](#research-with-sources)
> 9. [Create files and analyze data](#create-files-and-analyze-data)
> 10. [Images, visuals and Sites](#images-visuals-and-sites)
> 11. [Voice, browser and your computer](#voice-browser-and-your-computer)
> 12. [Plugins and skills](#plugins-and-skills)
> 13. [Scheduled tasks and long-running work](#scheduled-tasks-and-long-running-work)
> 14. [Privacy and safety](#privacy-and-safety)
> 15. [Five exercises to get hands-on](#five-exercises-to-get-hands-on)
> 16. [Quick reference](#quick-reference)

## How ChatGPT fits together

Every result depends on five layers. When an answer misses, one of these is usually the reason.

| Layer   | What it decides                     | Where you set it                                          |
| ------- | ----------------------------------- | --------------------------------------------------------- |
| Mode    | How much it does on its own         | **Chat** or **Work** above the composer, or **Codex**     |
| Model   | How hard it thinks                  | Model and reasoning control below the composer            |
| Context | What it knows about this task       | Your prompt, files, projects, memory and plugins          |
| Tools   | What it can reach and act on        | Web search, browser, plugins, Computer Use                |
| Reuse   | What it can repeat without you      | Skills and scheduled tasks                                |

**Where you can use it:**

- **Web** ([chatgpt.com](https://chatgpt.com)): Chat and Work. Work runs in the cloud.
- **Desktop app** (macOS, Windows, Linux): Chat, Work and Codex, plus local files and, where supported, the built-in browser, Computer Use, Appshots and Voice.
- **Mobile:** Chat and Work with your plugins. Cloud tasks started elsewhere can continue here.

## Set up in 10 minutes

1. **Start on the web.** Install the [desktop app](https://chatgpt.com/download/) when you need local files, the browser, Computer Use or Voice.
2. **Pick a personality.** Settings > Personalization: Friendly, Pragmatic or None. It changes tone, not capability.
3. **Write custom instructions.** Same page. Use the template below.
4. **Decide on memory.** Settings > Personalization. Memory recalls context from past chats. Put rules that must always apply in custom instructions or project instructions instead.
5. **Set your training preference.** Settings > Data Controls > **Improve the model for everyone**. When it's off, new chats aren't used for training.
6. **Turn on notifications.** Settings > Notifications, so you know when a long task finishes or needs approval.
7. **Coming from Claude or Cursor?** In the desktop app, Settings > Import brings over instructions, skills, plugins, projects and recent chats.

```text
About me: I work as [role] on [main types of work]. I already know [skills or tools], so skip the basics there.

How to respond:
- Lead with the answer, then the reasoning.
- Keep it short unless I ask for depth.
- Use tables for comparisons and numbered steps for processes.
- Tell me when you're unsure or when information may be outdated.
- If my request is ambiguous, ask one question. Otherwise state your assumption and continue.

Avoid: [filler, hype, phrases or formats you dislike].
```

**Test it:** open a new chat and ask three things you ask often. Rewrite any instruction it ignored.

## Choose a plan

US prices as listed in [OpenAI's pricing docs](https://learn.chatgpt.com/docs/pricing) on September 16, 2026.

| Plan             | Price                                     | Choose it if                                                        |
| ---------------- | ----------------------------------------- | ------------------------------------------------------------------- |
| Free             | $0                                        | You're trying ChatGPT for questions, drafts and light tasks          |
| Go               | $8/month                                  | You want a low-cost paid tier. Compare limits on [the pricing page](https://chatgpt.com/pricing) |
| Plus             | $20/month                                 | You use Work or Codex weekly and want Sites and desktop Voice        |
| Pro              | From $100/month                           | You regularly hit Plus limits. Choose 5x or 20x Plus usage           |
| Business         | $20/user/month annually, $25 monthly      | A team of two or more needs a shared workspace and admin controls    |
| Enterprise / Edu | Contact sales                             | An organization needs SSO, audit logs and retention controls         |

**Know before you choose:**

- **Ads:** Free and Go may show ads. Paid plans don't. See [Ads in ChatGPT](https://help.openai.com/en/articles/20001047-ads-in-chatgpt).
- **Shared usage:** Work and Codex draw from the same limits. Bigger models, higher reasoning, images and long tasks use more.
- **Check your usage** on the [usage dashboard](https://chatgpt.com/codex/settings/usage). Near the limit? Use a smaller model, narrow the task, or buy credits on Plus and Pro.
- **Business data:** Business and Enterprise workspaces aren't used for training by default.

## Choose a model

Open the model and reasoning control below the composer. Start with the default **Power** setting. Slide toward **Smarter** or **Faster** from there, or open **Advanced** to pick a specific model and effort.

| Model         | Use it for                                                                            |
| ------------- | ------------------------------------------------------------------------------------- |
| GPT-6 Astra   | The hardest end-to-end work across research, apps and code. Still rolling out         |
| GPT-5.6 Sol   | Complex, ambiguous or high-stakes work that needs depth and polish                    |
| GPT-5.6 Terra | Everyday work: reports, document analysis, drafting that needs judgment               |
| GPT-5.6 Luna  | Clear, repeatable, high-volume tasks: extraction, classification, reformatting        |

**Reasoning effort:** Light for quick, well-scoped tasks. Medium when it needs a plan. High or Extra High for multi-step work with sources and trade-offs. Use the lowest effort that gets the result.

**Max** gives one task more thinking time. **Ultra** splits a task across subagents working in parallel. Most tasks need neither. If Ultra is missing in the desktop app, turn it on in Settings > Configuration > **Ultra in model picker slider**.

**Heads up:** GPT-5.5 retires from ChatGPT, Work and Codex on October 14, 2026. Move saved settings and scheduled tasks to a GPT-5.6 model before then.

## Chat, Work or Codex

|             | Chat                                 | Work                                            | Codex                                  |
| ----------- | ------------------------------------ | ----------------------------------------------- | -------------------------------------- |
| Use it when | You want an answer, idea or draft    | You want a finished result to review            | The work lives in a code repository    |
| You get     | A reply                              | Files, research, Sites, completed steps         | Diffs, test runs, pull requests        |
| Your role   | You steer every turn                 | You approve at checkpoints                      | You review the implementation          |
| Example     | "Compare these two offers"           | "Build a spreadsheet comparing these offers and flag what's missing" | "Fix the failing checkout test" |

**Three quick tests:**

1. Is one reply enough? **Chat.**
2. Does it need several steps, sources or files? **Work.**
3. Does the result need to be tested and committed? **Codex.**

**Cloud or local:** On the web, Work runs in the cloud. In the desktop app, **Cloud** keeps a task running after you close the app, while **Work locally** can use the files, apps and browser on your computer.

Deeper comparison: [ChatGPT Work vs Codex](/guides/chatgpt-work-vs-codex).

## Prompts that work

A quick question needs no structure. For anything bigger, cover five parts.

| Part       | Ask yourself                        | Example                                           |
| ---------- | ----------------------------------- | ------------------------------------------------- |
| Goal       | What should exist when it's done?   | A one-page brief                                  |
| Context    | What does it need to read or know?  | The attached survey results                       |
| Output     | Who reads it, in what format?       | My manager. Decisions first, under 300 words      |
| Boundaries | What must not change or happen?     | Keep the budget figures. Don't send anything      |
| Check      | How should it verify its own work?  | Confirm every number matches the source           |

```text
Goal: [the finished thing you need]
Context: [files, links, background, who it's for]
Output: [format, length, structure]
Boundaries: [what to keep, avoid, or ask me about first]
Before you finish: [the check to run on your own work]
```

**Redirect instead of restarting.** Stay in the same chat and say:

- "Shorter. Keep the numbers."
- "Give me three different approaches before you pick one."
- "What assumptions did you make?"
- "Ask me what you need to know, then continue."
- "Which parts are you least confident about?"

**While a task runs** (desktop): **Steer** adds your message to the current run. **Queue** holds it for the next run. Set the default in Settings > General > Follow-up behavior.

**For big tasks** (desktop): type `/plan` to agree on the approach, then `/goal` with a clear finish line. See [long-running work](#scheduled-tasks-and-long-running-work).

**Dictate** in the desktop app with `Ctrl+Shift+D`.

## Give it the right context

Include what changes the answer and leave out the rest. Extra material uses more of your limit and gives it more to sift through.

| Context  | Use it when                                         | How                                                          |
| -------- | --------------------------------------------------- | ------------------------------------------------------------ |
| Files    | The task depends on a doc, spreadsheet, deck or PDF | Attach it and say what to take from each file                |
| Images   | Showing is faster than describing                   | Attach a screenshot and point to the part that matters       |
| Projects | Work continues over time or reuses sources          | Add files to **Sources** and write project instructions      |
| Plugins  | The information lives in another tool               | Type `@` and pick the plugin, such as `@Google Drive`        |
| Appshots | You're working in another desktop app               | Press both Command keys (Mac) or both Alt keys (Windows)     |
| Memory   | Stable preferences you'd otherwise repeat           | Settings > Personalization                                   |

**Project habits:**

- One chat per outcome. Name it after the result: "Q4 budget memo", not "budget stuff".
- Write the brief once in project instructions. They apply to every chat in the project.
- A web project can't see folders on your computer, so upload or connect sources. Desktop projects can attach local folders.
- Pin active projects. Archive finished chats.
- Find old chats on the web with `Cmd/Ctrl+K`.

```text
Help me write instructions for a ChatGPT project about [topic].
Ask me up to five questions about the goal, audience, sources, format and quality bar.
Then draft instructions under 200 words that every chat in this project should follow.
```

## Research with sources

**Quick, current facts:** ask in Chat and say you want current information with sources. Citations appear in the reply.

**Deep research** is for questions that need many sources and a report.

- **Web:** switch to **Work**, open the **+** menu, select **Deep research**.
- **Desktop:** switch to **Work**, open **Plugins**, select **Deep research**.

```text
Research [question] for [who will use it and what they need to decide].
Scope: [time range, regions, source types to include or exclude].
Deliver a short report: the answer first, then a comparison table, then linked sources.
Separate what sources state from your own inference. List anything you couldn't verify.
```

**Pressure-test the report:**

- "Cite the source for each major claim."
- "Which claims rest on a single source?"
- "What would change your conclusion?"
- "What couldn't you access?"

Then open the source behind every number, quote or date you plan to use. Web pages are untrusted input, and a confident summary can still be wrong.

## Create files and analyze data

Work creates documents, presentations, spreadsheets and PDFs you can open, revise and download.

```text
Using [attached files or @plugin sources], create a [document, 8-slide deck or spreadsheet] for [audience].
Structure: [the sections, slides, sheets or columns you expect].
Cite each number to its source file and flag anything that needs my review.
Before you finish, check that [the summary matches the data / every action has an owner].
```

**Revise without starting over:**

- Name the exact spot: "slide 4, chart title" or "sheet 2, column D".
- Say what stays: "Keep the layout and colors."
- On desktop, use **annotations**: select part of the preview and describe the change.

```text
Analyze the attached [CSV or spreadsheet]. Don't change the original.
First, describe the columns, row count and any quality issues such as missing values, duplicates or inconsistent formats.
Then answer: [your question]. Show the key numbers in a table and one chart, with each finding explained in one plain sentence.
```

**Before you use a file:** open every page, tab and slide, and spot-check the numbers against the source.

## Images, visuals and Sites

### Images

A strong image prompt is one to three sentences: purpose, subject, composition, style, and what to leave out.

```text
Create a [purpose, such as a blog header] showing [subject and action].
[Setting and composition]. [Style, lighting and colors].
Leave [area] empty. No text, logos or watermarks.
```

- **Edit one thing at a time.** Say what changes and what must stay the same.
- **Select an area** of the image to change only that part.
- **Text in images:** put the exact words in quotes, say where they go, and proofread every letter.
- **Several references:** number them. "Image 1 is the product. Image 2 is the style."
- Image turns use your limits faster than text turns.

### Interactive visuals

Type `@Visualize` to get a chart, diagram, map or calculator you can adjust. It's still rolling out. If the suggestion doesn't appear, your account doesn't have it yet.

```text
@Visualize [concept or data]. Let me adjust [inputs] and show how [outcome] changes. Add a short text summary.
```

A visualization is a snapshot of the data at the time, not a live dashboard.

### Sites

Sites builds and hosts a website, web app or game from a prompt. It's in public beta on Plus, Pro, Business, Enterprise and Edu.

1. Use the word "website" or mention `@Sites`. Describe the audience, purpose, features and any data it should remember.
2. Review the preview. Select **Edit** to describe changes.
3. **Every deployment is live.** Ask it to "save a version without deploying" when you want to review first.
4. A new Site is visible only to you and workspace admins until you widen access in **Share**.
5. Add API keys and secrets in the Site's settings, never in the prompt.

```text
Build a website for [audience] to [main job]. It needs [pages and features].
Save [data it should remember] between visits. Keep access limited to me.
Save a version for my review. Don't deploy yet.
```

Manage your Sites at [chatgpt.com/sites](https://chatgpt.com/sites). Visitor analytics are built in under More actions > Analytics.

## Voice, browser and your computer

These live mainly in the desktop app. Availability depends on plan, region and rollout.

| Feature           | What it does                                              | Start it                                              |
| ----------------- | --------------------------------------------------------- | ----------------------------------------------------- |
| Voice dictation   | Turns speech into prompt text                             | `Ctrl+Shift+D`                                        |
| ChatGPT Voice     | Live conversation that can start, check and steer tasks   | **Start new voice chat** (Plus and up)                |
| Built-in browser  | Browses and acts in its own separate browser profile      | `@Browser` or `Cmd/Ctrl+Shift+B`                      |
| Browser extension | Works in tabs where you're already signed in              | Settings > Computer Use, then `@` your browser        |
| Computer Use      | Sees and operates desktop apps                            | Install the Computer Use plugin, then `@Computer`     |
| Appshots          | Sends the front window to a chat                          | Both Command keys (Mac) or both Alt keys (Windows)    |

- **Voice** uses your plan's usage at $0.05 per minute, plus whatever the task itself uses. On a Mac, turn on **Screen context** in Settings > Voice and say "take a look at this".
- **The extension** works with Chrome, Edge, Brave, Opera and Vivaldi.

**Permissions (desktop):** start with **Ask for approval**. ChatGPT works inside the current workspace and pauses before going beyond it. **Approve for me** sends those requests to an automatic reviewer but keeps the same boundary. **Full access** lets it change files, run commands and use the network without asking, so avoid it for unattended work.

**Browser safety:** treat page content as untrusted, since a page can contain instructions aimed at the agent. Review what it plans to submit, buy, send or change. When a site needs a login, enter it in the sign-in flow, never in the chat.

```text
Use @Browser to [task] on [site]. Stop before submitting anything and show me exactly what you're about to enter.
```

## Plugins and skills

**Plugins** connect ChatGPT to tools like Gmail, Google Drive, Slack and GitHub, and can add ready-made workflows. **Skills** are saved instructions for a task you repeat.

### Plugins

1. Open **Plugins** in the sidebar and install the ones you use.
2. Connect the account when asked.
3. Type `@` to call a specific plugin, or describe the task and let ChatGPT choose.

Plugins work in Chat and Work on web, desktop and mobile. Some are marked **Desktop only**.

```text
Use @Slack and @Google Drive to summarize this week's decisions on [project]. List open questions with owners. Don't post anything.
```

### Skills

Once you've pasted the same prompt three times, turn it into a skill.

- **Write one:** type `@skill-creator` and describe the task using the template below.
- **Show one:** on a Mac with Computer Use enabled, go to Plugins > **+** > **Record a skill**, then do the workflow once. It drafts the skill from your recording.
- **Use it:** ChatGPT picks it up when relevant, or call it with `@skill-name` (`$skill-name` in Codex).
- **Test it** on a real request and tighten the step it missed.

Standalone skills run in the desktop app. On web and mobile, a skill has to come bundled in a plugin.

```text
@skill-creator Make a skill for [task].
It starts from: [inputs, such as meeting notes or a link].
A good result looks like: [format, length, sections]. Example: [paste one].
Always: [rules]. Never: [rules].
Before finishing, check: [quality check].
```

**Skill or plugin?** Build a skill for a focused process. Package it as a plugin when you need to bundle it with tool connections or share it with other people.

## Scheduled tasks and long-running work

### Schedule a task

Ask in any chat, for example "Every weekday at 8am...". Manage tasks and their runs in **Scheduled**.

| Type               | What each run does                              | Use it for                                          |
| ------------------ | ----------------------------------------------- | --------------------------------------------------- |
| Standalone         | Starts a fresh chat from the saved prompt       | Independent reports and checks                      |
| Inside a chat      | Returns to the same chat with its context       | Ongoing research, follow-ups, checking until done   |
| Event-triggered    | Runs when a Gmail, Slack or GitHub event occurs | Triaging email, channel summaries, PR feedback      |

- **Event triggers** work on web and mobile on eligible plans, and can't be combined with a time schedule. For Slack, add `@ChatGPT` to each channel it watches.
- **Local desktop tasks** need your computer on and the app open. Cloud tasks don't.

**Before you schedule:**

1. Run the prompt by hand in a normal chat until the output is reliable.
2. Make sure the prompt says what to do, what's worth reporting, and when to stop or ask you.
3. Review the first few runs and adjust.

```text
Every [cadence], check [sources, such as @Slack #channel and a @Google Drive folder] for [what to look for].
If nothing meets [threshold], reply "Nothing new" and stop.
Otherwise, send a brief: [format]. Draft any messages for my review. Don't send or change anything.
```

Not sure a schedule is the right fit? Read [Routine vs Schedule vs Loop](/guides/routine-vs-schedule-vs-loop).

### Hand off a long task

- **Web:** use Work and put the outcome, constraints and review criteria in the first message. Add context in the same chat as it runs.
- **Desktop:** type `/goal` and your definition of done. Pause, edit or clear it from the progress row. If you can't define done yet, run `/plan` first.
- **Parallel work:** run independent tasks in separate chats, but never let two tasks write to the same file or source.

```text
/goal [Outcome]. Constraints: [tools, sources, what not to touch]. Done when: [checks it can verify, such as every claim linking to a source file].
```

## Privacy and safety

| Control         | Where                                                      | What it does                                           |
| --------------- | ---------------------------------------------------------- | ------------------------------------------------------ |
| Model training  | Settings > Data Controls > Improve the model for everyone  | When off, new chats aren't used for training           |
| Temporary Chat  | Switch it on when starting a chat                          | Not saved to history, creates no memories, not used for training |
| Memory          | Settings > Personalization                                 | Review, delete or turn off what it remembers           |
| Shared links    | Settings > Data Controls > Shared links                    | See and revoke chats you've shared                     |
| Permissions     | Control below the composer (desktop)                       | Decide what it can do without asking                   |

Temporary Chats are still kept for up to 30 days for safety. See the [Temporary Chat FAQ](https://help.openai.com/en/articles/8914046-temporary-chat-faq).

**Rules to work by:**

- Don't paste passwords, API keys or data you aren't allowed to share. On a work account, follow your company's policy.
- Keep a human check before anything is sent, published, bought or deleted. Write "draft only" into the prompt.
- Treat websites, emails and documents it reads as untrusted. They can carry instructions meant for the agent.
- Review a chat or Site before you share its link.
- For legal, financial, medical and security decisions, use ChatGPT to prepare, then get expert review.

## Five exercises to get hands-on

Each exercise builds on the one before. Together they take you from a first prompt to a workflow that runs on its own.

1. **Calibrate.** Set your custom instructions, ask three questions you'd normally ask, and fix the instruction that didn't hold.
2. **Research.** Run deep research on a decision you actually face. Pressure-test it with the four questions in [Research with sources](#research-with-sources) and open three of the sources.
3. **Produce.** In a project, have Work turn real files into a document or spreadsheet you'll use this week. Revise it twice without starting over.
4. **Package.** In the desktop app, turn the prompt from exercise 3 into a skill with `@skill-creator` and run it on new inputs.
5. **Automate.** Schedule that skill. Review the first three runs and tighten the prompt.

## Quick reference

| Action                           | Where   | Shortcut or command                                  |
| -------------------------------- | ------- | ---------------------------------------------------- |
| Call a plugin or skill           | Any     | `@`                                                  |
| Call a skill in Codex            | Desktop | `$skill-name`                                        |
| Search past chats                | Web     | `Cmd/Ctrl+K`                                         |
| Dictate a prompt                 | Desktop | `Ctrl+Shift+D`                                       |
| Open the built-in browser        | Desktop | `Cmd/Ctrl+Shift+B`                                   |
| Take an appshot                  | Desktop | Both Command keys (Mac), both Alt keys (Windows)     |
| Quick chat from floating controls | Desktop | `Option+Space` (Mac), `Windows+Alt+P` (Windows)     |
| Open Activity view               | Desktop | `Cmd+Option+U` (Mac), `Ctrl+Alt+U` (Windows)         |
| Agree on an approach first       | Desktop | `/plan`                                              |
| Run a task to a finish line      | Desktop | `/goal`                                              |
| Control memory for this chat     | Desktop | `/memories`                                          |
| Check usage                      | Web     | [Usage dashboard](https://chatgpt.com/codex/settings/usage) |

Floating controls need a pet or **Mini** selected in Settings > Pets.

**Go deeper:** [OpenAI's ChatGPT docs](https://learn.chatgpt.com/docs) · [How to Give AI Taste](/guides/how-to-give-ai-taste) · [ChatGPT Work vs Codex](/guides/chatgpt-work-vs-codex) · [5 Codex Features](/guides/5-codex-features)