What practical ChatGPT habits make people better at using it?

Good use starts with a usable brief

ChatGPT cannot reliably infer the background you leave out. A request such as “write an email to my team” omits the audience, relationship, purpose, decision needed, and the facts that must not be changed. The model may fill those gaps with a plausible guess. That can be useful for brainstorming, but it is a poor way to get a dependable work product.

Before you write a long prompt, make five decisions:

1. Outcome: What should exist when the task is done?
2. Audience: Who will read or use it, and what do they already know?
3. Context: Which facts, examples, source text, or prior decisions matter?
4. Constraints: What must the answer avoid or preserve?
5. Quality check: How will you decide whether it is good enough?

You do not need all five for a trivial request. The point is to supply the missing information that would change the answer. A useful first message is often short:

“Draft a 150-word update for customers who already know our product. Explain that the launch moves from 10 October to 24 October because the security review is still open. Keep the tone calm and accountable. Do not promise a date beyond 24 October. Before drafting, list any assumption that could change the message.”

That is not a magic formula. It works because it gives the task a purpose, an audience, boundaries, and a way to surface uncertainty. If the answer starts from the wrong assumption, correct that assumption rather than asking it to “try again.”

Supply representative context, not every file you own

Context is more than background facts. It can include a good past example, a source document, a table, a customer’s actual question, a style guide, or a decision that has already been made. One well-chosen example is often more useful than several abstract instructions.

ChatGPT can work with supported documents, spreadsheets, presentations, and other uploaded files, including for comparison, extraction, and transformation tasks. File availability and limits depend on the account and plan. OpenAI: File Uploads FAQ That makes an upload useful when the file is the source of truth, but it does not make every upload wise or necessary.

Use this decision rule:

The table is about relevance, not quantity. If a fact would change the recommended action, include it. If it is merely interesting, leave it out until it becomes needed.

Example

Hypothetical personal task: you want help choosing between two weekend trips. Instead of asking “Which trip is better?”, provide the dates, budget, transport constraint, preferences, and a small comparison table. Ask ChatGPT to list the trade-offs, say which assumptions it made about weather or availability, and recommend one option only if those assumptions are acceptable.

The action is to give the model the decision criteria, not just the destination names. The takeaway is that you remain the decision-maker, while ChatGPT helps structure the comparison and identify what you still need to check.

Make uncertainty visible before it becomes a mistake

For many tasks, the highest-value follow-up is not “make it better.” It is “tell me what you are assuming, what evidence is missing, and what could make this recommendation wrong.”

Use prompts such as:

• “Separate confirmed facts from assumptions and suggestions.”
• “Give the strongest reasonable objection to this plan.”
• “What information would materially change your answer?”
• “Mark each claim as supported by the supplied material, general knowledge, or an inference.”
• “If the source does not answer this question, say so rather than fill the gap.”

This habit does not make ChatGPT certain. It makes uncertainty inspectable. That is especially helpful when a request contains an ambiguous term, an unstated constraint, or an example that could be unrepresentative.

It also makes collaboration faster. If the model identifies the missing choice as “Is the audience existing customers or new prospects?”, you can answer that question once instead of revising three drafts that were aimed at the wrong audience.

Give factual feedback, not only a thumbs-down

ChatGPT improves within a conversation when you give it usable correction. “Not quite” gives little direction. “The factual content is right, but the opening assumes the reader knows the project. Add one sentence of background, remove the jargon, and keep the final call to action” tells it what to preserve and what to change.

When a draft is partly good, use a three-part correction:

1. Name what to keep.
2. State what is wrong or missing.
3. State the revised acceptance test.

For example: “Keep the three milestones and the concise tone. The second milestone is dated 12 May, not 21 May. Rewrite for a non-technical director and end with the decision we need by Friday.”

This works for writing, planning, coding, analysis, and explanation. It is also a guard against endless regeneration. Each pass should answer a known question: Did it fix the date? Is the decision now explicit? Does it meet the word limit? If you cannot name the next test, stop generating and review the task itself.

Ask for verification when the claim matters

ChatGPT may state an incorrect fact, invent a quotation or citation, or sound more certain than the evidence supports. OpenAI’s own guidance recommends checking important information from reliable sources and visiting cited links directly. OpenAI: Does ChatGPT tell the truth?

Use a different review standard for different kinds of work:

When current facts matter, enable search or use an available research tool and ask for a citation beside each important claim. Then inspect the source itself. ChatGPT search can provide cited answers, but a citation is a path to evidence, not proof that the source says what the response claims. OpenAI: Does ChatGPT tell the truth?

For a factual task, a good final question is: “What would you want me to verify before I act on this?” It keeps the model in the role of a useful assistant instead of an unaccountable authority.

Put reusable guidance in the right place

Repeatedly explaining the same durable preference wastes time. Putting every instruction into every chat creates noise. ChatGPT gives you two different places for reusable guidance, and they serve different purposes.

Custom instructions are for broad, stable preferences that should affect most chats, such as your preferred level of detail, writing style, units, language, or a request to flag uncertainty. They are applied to all chats and can be edited or deleted for future conversations. OpenAI: ChatGPT Custom Instructions

Projects are for a bounded, continuing effort such as a job search, a course, a product launch, or a recurring research brief. They keep related chats, files, and project-specific instructions together. Project instructions apply only inside that project and override global custom instructions. OpenAI: Projects in ChatGPT

Use a regular chat when the work has no future value. Use a project when the answer depends on a shared source set or repeated decisions. Use custom instructions when the preference genuinely applies across topics. This separation prevents a useful instruction for one project from becoming a bad default everywhere else.

Example

Hypothetical professional task: a project manager prepares a weekly internal update for a product launch. The team has an approved ChatGPT workspace and a project containing the current roadmap, risk log, and update template. The project instruction says to distinguish facts, open risks, and decisions needed, and to ask a clarifying question if the source dates conflict.

Each Friday, the project manager adds only that week’s changes and asks for a 250-word update for senior stakeholders. They then check dates and ownership against the roadmap before sending. The takeaway is not “store every work document in a project.” It is to keep the small, approved source set and repeated instructions with the recurring task, while retaining human ownership of the published update.

Use ChatGPT as a challenger, not only a writer

Drafting is the obvious use, but questioning your plan can be more valuable. Once you have a proposal, ask the model to examine it from a defined perspective:

• “Act as a skeptical customer. Which claim is unsupported or confusing?”
• “List the strongest arguments against this recommendation, then say what evidence would resolve each.”
• “Find internal contradictions between these two requirements.”
• “Turn this plan into a pre-mortem. What are three plausible failure paths?”
• “Ask me the three questions that would most improve this brief.”

The important constraint is to specify the perspective and evidence. “Critique this” often produces generic objections. “Critique this release plan as a support lead who must handle a surge in tickets, using only the supplied launch notes” produces a review you can test.

Do not mistake an adversarial-looking answer for independent validation. It is still generated reasoning. Use it to surface questions, then resolve them with data, documents, experiments, or people who have relevant expertise.

Know when a chat should not be a ChatGPT task

Some tasks are poor fits because the cost of a plausible mistake is too high or because the source material should not be supplied.

Do not use a general ChatGPT conversation as the final authority for:

• A diagnosis, treatment choice, or urgent medical decision.
• Legal advice, contractual commitment, regulatory interpretation, or an individual rights decision.
• A financial trade, tax filing, credit decision, payment instruction, or other material financial action.
• Hiring, firing, discipline, admissions, insurance, benefit, or safety decisions about a person.
• Work involving confidential client material, security credentials, personal identifiers, protected health information, child data, or trade secrets unless the approved environment, policy, and terms specifically permit it.

In these areas, ChatGPT can still help prepare a plain-language explanation, a checklist of questions, or a draft for qualified review. It must not replace the applicable professional, policy, or authoritative source.

Privacy is also a habit, not a settings task you do once. ChatGPT Data Controls let signed-in users turn off “Improve the model for everyone,” which prevents new conversations from being used to train ChatGPT while they remain in chat history. OpenAI: Data Controls FAQ Temporary Chats do not appear in history, do not create memories, and are not used to improve models while they remain temporary, but OpenAI may retain them for up to 30 days for safety purposes. OpenAI: Temporary Chat FAQ Neither setting overrides an employer’s policy, client agreement, data-protection law, or the terms of a connected third-party app.

A small weekly practice that builds judgment

Pick one real but low-stakes task each week and keep a short record:

1. Write the working brief in two or three sentences.
2. Ask for assumptions and an initial answer.
3. Give one factual correction or a concrete quality constraint.
4. Verify one claim, number, or source yourself.
5. Note which missing input would have improved the first response.

After four weeks, look for the pattern. You may find that you regularly omit audience, leave decisions implicit, provide contradictory examples, or accept claims without checking the source. The insight is not about which prompt to memorize. It is about which part of your own task definition needs more care.

Common habits that look helpful but are not

The practical alternative to both prompt tricks and avoidance is deliberate use: define the task, give relevant evidence, ask for uncertainty, revise with facts, and verify before acting.