Using ChatGPT Effectively at Work: A Practical Guide

By Published

A practical guide to using ChatGPT effectively at work by choosing the right mode, tool, memory setting, and project structure for each task.

ChatGPT is no longer just “one chat window.” To use it correctly, you need to distinguish between several separate layers: the conversation work mode, information-gathering tools, persistent personal settings, and projects that hold long-running context. Most problems in real work with ChatGPT do not come only from weak prompting. They come from choosing the wrong layer: using Memory instead of a Project, using Deep research for a question that only needs Search, or working in Instant on a task that really requires Thinking.

The correct workflow is sequential: first decide what kind of task you are dealing with, then choose the right mechanism, and only then add context, instructions, or memory. In other words, do not start by dumping information into the system. Start by deciding where that information should live. That also aligns with the broader NIST approach to GenAI risk management: match the mechanism to the context, level of control, and operational goal instead of using one generic workflow for everything.

Quick decision map

  • Need a fast answer, short writing help, a brief summary, or a text fix → Instant
  • Need task decomposition, planning, debugging, or deeper reasoning → Thinking
  • Need a current fact or a quick web check → Search
  • Need multi-source research with a documented report → Deep research
  • Need both reasoning and action on a site, file, or system → Agent mode
  • Need a persistent workspace with chats, files, and local instructions → Project
  • Need a preference to apply across almost every conversation → Personalization / Custom Instructions / Memory

Short glossary

Instant — the fast-response mode for everyday work.
Thinking — a mode that applies deeper reasoning for more complex tasks.
Pro — the highest-reasoning option for especially demanding work, without Apps, Memory, Canvas, or image generation.
Search — fast web search with sources.
Deep research — multi-step research with planning, source control, and a documented report.
Agent mode — a mode where ChatGPT both reasons and performs actions.
Personality — your general tone and style selection.
Custom Instructions — persistent cross-chat rules.
Memory — ongoing details the system may remember for future conversations.
Project — a persistent workspace around a topic, task, or domain.
Project Instructions — local instructions that apply only inside one project.
Project source — material saved inside a project for reuse.
Temporary Chat — a chat that does not use memory and does not create new memory.

1) The chat screen: where everything starts

The main chat screen is the entry point for the core capabilities. From there, you can choose a mode through the model picker, enable Search through View all tools or the slash menu, start Deep research through /Deepresearch, the tools menu, or the sidebar, and start Agent mode through the tools menu or /agent. That matters because a practical usage guide cannot assume the reader already knows where each capability lives.

How to start a new task correctly

  1. Open a new chat.
  2. Look first at the model picker and View all tools / tools menu.
  3. Decide whether the task is a response task, a search task, a research task, or an action task.
  4. Only then write the request.

Correct usage: choose the mechanism first, then phrase the request.
Incorrect usage: start typing everything into a regular chat and only later try to understand why the output is not precise enough.

2) Conversation work mode: Instant, Thinking, Pro

The model picker currently exposes Instant, Thinking, and Pro. Instant is for fast answers. Thinking is for more complex work. Pro is for especially difficult reasoning-heavy work. When you use Instant, ChatGPT may automatically switch to Thinking for harder requests. Through Configure, you can also control automatic switching and thinking effort when Thinking or Pro is selected.

When to use each mode

  • Instant — drafting, short summaries, translation, editing, direct questions, light brainstorming.
  • Thinking — planning, comparing options, multi-step reasoning, debugging, trade-off analysis, building structured frameworks.
  • Pro — especially demanding reasoning tasks, as long as you do not need the features unavailable in that mode.

How to access it

  1. Open a chat.
  2. Choose the mode in the model picker.
  3. If you need more control, open Configure.
  4. Decide whether to keep automatic switching enabled and, if needed, set thinking effort.

How to work correctly

Use Instant as the default for simple and fast tasks. Move to Thinking when the task requires stronger logical consistency, stepwise decomposition, or materially higher judgment. Use Pro only when the task truly requires maximum reasoning depth, and only if the unavailable features are not part of the workflow.

Common mistake

Working in Instant for everything and expecting the system to “figure out by itself” that the task is complex. The opposite mistake is using Pro for everything even when the task is simple and the blocked features matter.

Correct example

“Compare three architecture alternatives, including trade-offs, risks, and a recommendation.” → start with Thinking.

Incorrect example

“Fix the wording of a short email.” → moving straight to Pro.

3) Search: when you need quick web access, not research

ChatGPT Search is designed for fast, current answers from the web, with links and sources. You can access it through View all tools → Search, or through the slash menu. ChatGPT may also trigger web search automatically when the question clearly benefits from current web information. Search-based responses include citations, and you can open Sources below the answer.

When to use it

Use Search when you need a current fact, a price, news, a quick source, a short list, or a web-grounded answer without a full research report.

How to access it

  1. Open a chat.
  2. Choose View all tools → Search, or use the slash menu.
  3. Enter the question.
  4. Check the sources under the inline citations or under Sources.

How to work correctly

Use Search when the task is fundamentally a quick verification task. If the output needs to combine several sources, compare them, and synthesize them into a structured conclusion, the task is no longer Search. It is Deep research.

Common mistake

Using Search for a task that actually requires multi-step documented research, then being disappointed that the result is too short or not organized like a report.

Correct example

“What is the latest OpenAI update on Projects?” → Search

Incorrect example

“Summarize the main trends in domain X across several sources and identify gaps and disagreements.” → not Search; use Deep research

4) Deep research: when you need a documented research workflow, not a quick answer

Deep research is intended for multi-source tasks where you need planning, investigation, synthesis, and a documented report. It can work with websites, uploaded files, and other sources it is allowed to access. The workflow is explicit: define the desired outcome, choose the allowed sources, review the proposed research plan, run the task, and then inspect the final structured report with citations or source links.

When to use it

Use Deep research when the question requires gathering and synthesizing information from multiple sources, multi-step analysis, or explicit control over which sources are used. For short questions or quick lookups, standard chat or Search is usually faster.

How to access it

  1. Type /Deepresearch, or choose Deep research from the tools menu or sidebar.
  2. Describe the desired output.
  3. Select the sources: web, uploaded files, and any relevant connected apps.
  4. Review and adjust the research plan before starting.
  5. During the run, follow progress and narrow the scope if needed.
  6. At the end, inspect the sources inside the report.

How to work correctly

Do not use vague input such as “research this topic.” Define the required output instead: a comparative report, market snapshot, risk map, source review, or critical analysis. If the task should focus on specific websites, use the site-selection controls so the research is constrained to the right sources.

Common mistake

Using Deep research for a simple factual question, or launching it without reviewing the research plan first. Both mistakes either waste time or produce a result that is broader than necessary.

Correct example

“Create a short report comparing Search, Deep research, and Agent mode, using only official documentation, including a use-case table and a limitations summary.” → Deep research

Incorrect example

“What is the weather in Tel Aviv tomorrow?” → not Deep research

5) Agent mode: when you need to move from answers to action

Agent mode is for tasks where ChatGPT needs to both understand and act: navigating sites, working with files, filling forms, editing spreadsheets, and using tools such as a visual browser, code interpreter, terminal, and apps while keeping the user in control. You enable it through the tools menu or /agent, describe the task, and the system begins execution. It may pause for clarification or confirmation.

When to use it

Use Agent mode when the task includes reasoning plus action: for example, research that also needs to end in a file operation, a site action, a structured extraction workflow, or a multi-step process that includes navigation and execution.

How to access it

  1. Open a chat.
  2. Choose Agent from the tools menu or type /agent.
  3. Describe the task clearly.
  4. Follow the steps and approve actions when required.
  5. If login is required, the system will pause and ask you to take over the virtual browser.

How to work correctly

Phrase the task clearly and narrowly. Instead of “handle all my email,” use something like “find three messages that require action, extract the tasks from them, and do not send anything without approval.” If the task involves access to a site or account, enable only the necessary connections and avoid typing passwords into normal chat messages.

Common mistake

Using Agent mode for work that is really just writing, editing, or text analysis, or giving it a vague instruction with too much freedom. This matters because connecting an agent to sites or apps introduces privacy and prompt-injection risk.

Correct example

“Open the uploaded CSV, build an exceptions table, and draft a summary. Do not perform any external actions.” → Agent mode is reasonable if the workflow actually combines reasoning and execution.

Incorrect example

“Write an introduction for my article.” → not Agent mode

6) Personalization: what should persist at the user level

Persistent user-level settings are managed through Settings → Personalization. Under Personalization you will find Personality, Custom Instructions, and Memory. These are not the same thing. Personality defines how ChatGPT sounds. Custom Instructions define explicit cross-chat rules. Memory stores ongoing details the system may reuse later.

6.1 Personality / Base style and tone

Personality defines how ChatGPT sounds, not what it knows. To change it, open your profile, go to Personalization, and choose a value under Base style and tone. The change applies across all chats, including existing ones. Personality does not change the system’s capabilities or safety rules. It changes communication style.

When to use it

Use Personality when you need a stable communication style, such as Professional, Efficient, or Candid.

How to work correctly

Use Personality for general tone and style only. If you want a consistently professional working tone, Professional or Efficient are the right kinds of settings. If you want more direct answers with less extra text, Efficient is often a better fit.

Common mistake

Expecting Personality to replace explicit instructions or project context. Personality does not replace Project Instructions, does not replace Custom Instructions, and should not contain operational workflows. If memory or another instruction conflicts with the chosen personality, those higher-priority signals may override it.

Correct example

Choosing Professional for work that requires formal language, precision, and business-style writing.

Incorrect example

Using Personality to store validation rules, research procedures, or project-specific operating logic.

6.2 Custom Instructions

Custom Instructions are for what should apply to almost every conversation. On web and desktop, go to Settings → Personalization, enable customization, and enter the instructions in the relevant field. The change applies immediately across chats.

When to use them

Use Custom Instructions for durable cross-chat rules: preferred language, response structure, citation requirements, fact-versus-assumption separation, or a default communication style that should apply almost everywhere.

How to work correctly

Write short general rules, not documents. For example:

  • “Separate facts from assumptions.”
  • “Use clear section headings.”
  • “Write in concise professional Hebrew.”

These are global rules. But anything that applies only to one research stream, one client, one repository, or one site does not belong here.

Common mistake

Overloading Custom Instructions with project material: a glossary, a site-editing policy, or a long brief. The moment an instruction is local to one context, it belongs in Project Instructions, not in the global layer.

Correct example

“Always surface uncertainty when information is not verified.”

Incorrect example

“Use the internal structure of Project X, follow terminology A/B/C, and only inspect the docs folder of that repo.” → that is project-local, not global.

6.3 Memory

Memory includes Reference saved memories and Reference chat history. Saved memories are details you explicitly ask the system to remember, or details the system stores because they may be useful later. Chat history allows the system to draw on previous conversations, but it does not mean the system remembers every detail. Anything that must stay reliably available is better suited to saved memory than to chat-history inference.

When to use it

Use Memory for personal preferences and persistent details that should stay top-of-mind over time.

How to access it

  1. Open Settings → Personalization.
  2. Enable or disable Reference saved memories and Reference chat history.
  3. Open Manage memories to search, sort, delete, or manage memory items.
  4. If you do not want a conversation to use memory or create new memory, start a Temporary Chat.

How to work correctly

Store only things that genuinely need to persist: language preference, preferred output format, professional preferences, or personal details that should affect future responses. Do not use Memory as a document store. Do not store long templates there. Do not manage projects there.

Common mistake

Assuming that deleting a chat also deletes saved memory created from that chat. It does not. To fully remove a memory, you need to delete both the saved memory itself and the chat where it originated.

Correct example

Saving a preference such as “write in concise professional Hebrew.”

Incorrect example

Saving an operating manual, a full article outline, or a long project brief in Memory.

7) Projects: when a normal chat is no longer enough

Projects are persistent workspaces that group chats, files, sources, and local instructions in one place. They are designed for writing, research, planning, and repeated workflows. Once work around one topic starts stretching across multiple conversations, a normal chat becomes fragmented. Projects solve that problem.

When to open a project

  • When several chats belong to the same topic
  • When files or materials should stay together
  • When you need local instructions that should not affect the whole account
  • When the work is repeated over time

How to create a project

  1. Open the sidebar.
  2. Choose New project.
  3. Give the project a name.
  4. Start adding context: chats, files, text, and reusable sources.

7.1 Project Instructions

Project Instructions apply only inside that specific project. To configure them, open the project’s three-dot menu, go to Project settings, and add the instructions. Project Instructions override your global Custom Instructions inside that project.

How to work correctly

Put here everything that is local to this project: local style rules, terminology rules, workflow methodology, output structure, or any constraint that should apply only in this space. This is where project-level operating behavior belongs.

Common mistake

Putting all of that into global Custom Instructions and polluting the rest of the account with instructions that were meant to stay local.

7.2 How to add context to a project

Inside a Project, you can upload files, paste text, and add reusable sources. OpenAI also documented a tabbed layout with Chats and Sources, which makes it easier to manage reusable project context separately from conversation history.

How to work correctly

  • File is appropriate for a stable document you want to keep in the project.
  • Pasted text is appropriate for a short note, brief, or excerpt.
  • Project source is appropriate for material you want to preserve and reuse inside the project.

7.3 How to add a source to a project

You can add supported app links as project sources. The documented workflow is straightforward: open the project, go to the sources area, choose Add source, paste a supported link, and connect the source if access approval is required. The officially documented examples include Google Drive and Slack.

When to use it

Use this when you need a live source that should keep serving as persistent project context.

Common mistake

Pasting the same material manually into multiple chats instead of saving it once as a reusable project source.

7.4 How to save a ChatGPT response as a source

If ChatGPT generates an output worth reusing — a summary, note, draft, or analysis — you can save it as a project source. Open the message menu, choose Save to project or Add to project sources, and reuse it later from the project’s Sources area.

When to use it

Use this when a response has become working knowledge rather than a one-time answer.

Common mistake

Leaving a strong output buried in one conversation and recreating it repeatedly instead of promoting it into the project’s reusable context.

7.5 How to move an existing chat into a project

You do not need to start inside a Project. If a normal chat evolves into part of an ongoing workflow, you can move it into the project. The documented workflow is either to drag the chat into the project or open the chat menu and choose Move to project. Once moved, the chat inherits the project instructions and the project file context.

When to use it

Use this when a chat is no longer a one-off interaction and has clearly become part of a repeated workflow.

Common mistake

Leaving that chat outside the project and then losing consistency, context, and local instructions.

7.6 Project-only memory

When creating a Project, you can choose between default memory and project-only memory. In project-only mode, previously saved memories are not referenced, and chats can refer only to other conversations inside that same project. They cannot draw context from outside the project. Temporary Chats cannot be added to Projects.

When to use it

Use project-only memory when you need a hard context boundary: sensitive work, long-running research, or a project that should not rely on the broader account history.

Common mistake

Turning on project-only memory without understanding that it creates a hard context boundary, or assuming it can be retroactively applied to every existing project. Project-only memory is chosen when a new project is created. It is not a universal global switch for all projects.

8) Common usage mistakes

The first mistake is trying to manage a project in Memory. Memory is for persistent preferences and user-level details, not for managing a project knowledge base.

The second mistake is placing project-specific instructions in Custom Instructions. If the rule is local, it belongs in Project Instructions.

The third mistake is choosing Deep research for a short factual question, or choosing Search when the task actually requires documented multi-step research.

The fourth mistake is enabling Agent mode when the task only needs writing, editing, or text analysis. Agent mode should be used only when the task genuinely requires action.

The fifth mistake is treating Personality as a substitute for instructions. Personality changes style. It does not replace operational rules, Project Instructions, or full professional context.

9) Quick reference

  • Instant → fast answers, drafting, short summaries
  • Thinking → deeper reasoning, planning, debugging, comparisons
  • Pro → especially heavy reasoning, without Apps, Memory, Canvas, or image generation
  • Search → current fact or quick web check
  • Deep research → documented multi-source research
  • Agent mode → reasoning plus action
  • Personality → general style and tone
  • Custom Instructions → global rules
  • Memory → persistent preferences and details
  • Project → ongoing work with chats, files, sources, and local instructions
  • Project Instructions → rules that apply only inside one Project
  • Project source → reusable material stored inside the Project
  • Temporary Chat → work without using Memory and without creating new memory

Primary references