ChatGPT, minus the mystery

ChatGPT is a chat interface to a family of large language models built by OpenAI. You type a request, it writes a response, and the conversation keeps enough context to follow up without you repeating yourself. It is not a search engine, a database, or a person. It is a very well-read text predictor with a good memory for the current thread.

If you arrived by typing chatpt, it is a typo, not a different product.

That definition matters because most disappointment comes from treating it as something else. Used as a search engine, it invents. Used as a first draft, an editor, a tutor that asks questions, or a translator of jargon, it earns its place.

What it is good at

• Turning rough input into a usable shape. Notes into an outline. A transcript into action items. A messy paragraph into three clean sentences.
• Explaining at the level you ask for. “Explain this to a smart 14-year-old” is a real instruction, and it works.
• Generating options. Ten subject lines, five ways to open a difficult email, three schedules that fit constraints.
• Rewriting with a rule. Shorter, warmer, less jargon, same meaning, no new facts.
• Holding a structure. A table, a checklist, a comparison, a lesson plan.

What it is bad at

• Facts you have not checked. It states wrong things with the same tone it uses for right things. Fluent is not the same as correct.
• Anything after its training cutoff unless it can look it up. If the answer needs today’s news, ask it to search or check the source yourself.
• Precision you did not supply. It cannot know your company’s refund policy, your client’s name, or last quarter’s numbers unless you put them in the prompt.
• Taste on its own. Without an audience and an example, it defaults to the average of the internet, which reads like a press release.
• Being a calculator you trust. It can do arithmetic, but for anything financial or legal, verify with a tool built for it.

None of this is a reason to avoid it. It is a reason to give it a job it can do. The full risk register, with the mechanism and the fix for each, is in the downsides guide.

Your first session

1. Go to the official site. The address is chatgpt.com. Bookmark it. There are many lookalike domains; none of them are the product.
2. Create an account or sign in. OpenAI’s own help center covers account creation and sign-in at help.openai.com. If a page asks for your password outside the official domain, close it.
3. Find the composer. The large text box at the bottom is the only control that matters at first. Everything else is optional.
4. Start a fresh thread for a new job. One thread per task keeps the context clean. A long thread about three different things produces mushy answers.
5. Ignore the feature tour. You do not need to configure anything to get value. Learn the composer first; explore the rest when a task demands it.

Exact labels, menus, and available features change. When something in the interface does not match a guide — including this one — the official help center is the source of truth.

Five first tasks worth doing

These are chosen because they are hard to fake and easy to check.

1. Explain something you already understand

Explain how compound interest works to someone who is 16 and hates maths.
Use one concrete example with small numbers.
Then list the two mistakes people make most often.

You already know the answer. This is how you calibrate the tool: you can see where it is clear, where it pads, and where it is wrong.

2. Turn notes into a plan

Here are my raw notes from a project kickoff: [paste notes].

Turn them into a one-page plan with: goal, three milestones, owner for each, and the
first action due this week. Do not invent anything that is not in the notes. If an owner
is missing, write TBD.

The instruction “do not invent” is doing real work. Without it, you get a confident plan with people who were never in the meeting.

3. Edit something you wrote

Edit the text below. Keep my voice. Cut filler and repetition. Do not add facts or
examples I did not write.

Return the edited version, then a list of the three biggest changes and why.

[paste text]

This is the highest-value daily use for most people. It is also the easiest to verify, because you know what you meant.

4. Make a decision table

I am choosing between [option A] and [option B] for [situation].
Build a table with the criteria that actually matter for this decision, score each option
1-5, and explain the score in one line.
Then tell me which criterion would flip the decision.

You will often disagree with a score. That is the point: the table shows you which assumption you disagree with.

5. Rehearse a conversation

Role-play a conversation with my manager about moving to a four-day week.
You play my manager. Stay realistic, not hostile.
Start with their first question.

Then ask for feedback after ten exchanges. This is cheaper than rehearsing in the shower and more useful than a list of “tips for difficult conversations”.

How to tell a good answer from a confident one

Three checks, in order:

1. Provenance. Did the facts come from you, or did the model supply them? Ask: Which of these claims came from my message, and which did you add? Anything in the second pile gets checked.
2. Specificity. Does the answer contain details that could only come from your brief, or could it have been written for anyone? Generic answers usually mean a generic brief.
3. Friction. Does it disagree with you anywhere? A useful answer often pushes back: “this brief has two goals that conflict.” An answer that agrees with everything is usually optimizing for pleasantness, not accuracy.

If an answer fails all three, do not ask it to “be more accurate”. Give it better input, or use it for a different part of the job.

The habits that make it useful

• One thread, one job. Start fresh when the topic changes.
• Paste the source. Do not ask it to remember a document it never saw.
• Ask for the shape. “A table with four columns” beats “organize this”.
• Constrain the failure you keep seeing. If it adds fake quotes, forbid quotes. If it pads, set a word limit.
• Keep a prompt file. When something works, save it with the fill-in brackets. Your own library beats any listicle, including ours.
• Verify what you will publish or send. You are the last checkpoint. The model is not.

Where to get official answers

• Product and account help: help.openai.com
• The product itself: chatgpt.com
• Terms and policies: linked from the footer of the official site

This site is an independent guide. It is not affiliated with OpenAI, and it does not provide accounts, access, or support. If you need help with your account, the official help center is the only place that can actually resolve it.

Next

• A prompt is a brief, not a wish — the five parts of a prompt that gets a usable answer.
• Use it like an editor, not a vending machine — a writing workflow with a review step.
• When it will not cooperate — a triage table for the common failures.
• Open the workbench — write a rough request and turn it into a brief.