Phase 1 · ChatGPT · Level 1 · Foundations
Prompting ChatGPT well
By the end, you'll be able to…
- Apply the Phase 0 role-task-context-format framework specifically to ChatGPT
- Use ChatGPT's own quirks (its long memory within a thread, its eagerness to fill gaps, its default to long prose) to your advantage
- Iterate on a reply in the same thread instead of starting over, and know when to start fresh
Why it matters
You already have a general prompting framework from Phase 0. ChatGPT has its own personality on top of that: habits that reward a certain way of asking. Knowing those quirks is the difference between fighting the tool and steering it, and iterating in a thread is the single skill that makes it feel effortless.
Building on what you know
In Phase 0 you learned the durable framework: give a prompt a role, a task, context, and a format, and be specific. That works with every AI tool and it works with ChatGPT. This lesson adds the ChatGPT-specific layer: the handful of habits this particular tool has, and how to use them rather than trip over them.
Quirk 1: it holds the whole thread, so build, don't restart
ChatGPT can see everything earlier in the current conversation, and that span of remembered text is its context window. Within one chat it hasn't forgotten your first message when it writes its fifth reply. This is the most important quirk, because it changes how you work: you don't cram everything into one perfect prompt, you have a conversation. Set the task, read the draft, then nudge. Each nudge inherits everything already established.
The flip side: because it carries the whole thread, an unrelated new task in the same chat can get coloured by what came before. New topic, new chat: you learned the button in the tour, and this is why it matters.
Quirk 2: it fills gaps confidently unless told not to
Leave something unspecified and ChatGPT won't stop to ask; it'll pick a plausible answer and carry on as if you'd told it. Ask for "a reply to the customer" without the facts and it invents a name, an issue and an apology for something that never happened. This isn't a bug; it's the tool trying to be helpful. You manage it two ways: give it the real material, and add a short guardrail like "don't invent details I haven't given you; if something's missing, ask."
Quirk 3: it defaults to long, hedged prose
Left to its own devices, ChatGPT tends towards thorough, slightly over-explained answers with a caveat or two. That's fine when you want depth and annoying when you want a one-liner. The fix is simply to say so: name the length, the format, and the register every time it matters. "In one sentence." "As three bullets." "No preamble, just the rewrite."
Quirk 4: it copies examples brilliantly
If you want a particular voice or shape, showing beats describing. Paste one short example of the tone or structure you're after and ChatGPT will match the pattern more faithfully than any adjective could. One good example is worth a paragraph of instructions.
A worked example, start to finish
Watch the quirks combine on a real Fernway task: turning a decision buried in an email thread into a clear note for someone who wasn't copied in.
You're helping me write a short internal note. Task: from the email thread below, write a 4-sentence summary for a colleague who wasn't on it, covering what the disagreement was and how it was resolved. Plain, neutral tone. Don't add any detail that isn't in the thread. Thread: [paste the Fernway email thread].
Why this works: It sets a role, a precise task, the source material, an exact format, and a 'don't invent' guardrail, covering the framework and heading off Quirk 2 in one go.
ChatGPT returns a neat four-sentence note: the onboarding-versus-Harlow clash, Maya's compromise of moving the sales-shadowing to the 25th, and everyone agreeing. It reads cleanly and sticks to what's there. Now you iterate, not by rewriting the prompt, but with a short follow-up:
Good. Two tweaks: make it two sentences shorter, and add a final line stating the agreed dates so there's no ambiguity.
Why this works: Concrete feedback on the draft it already has is faster than re-explaining, and because the thread is still in view it keeps everything that was working.
It trims and adds the dates, keeping the rest untouched. If you wanted three versions to choose between, you'd just ask ("give me three versions of that, one more formal, one more casual") because choosing between options is often easier than judging one draft alone.
Rewrite it to match how I usually write updates. Here's a note I sent last week so you can match the voice: [paste two or three sentences of your own]. Keep the facts identical.
Why this works: Pasting a snippet in your own style lets ChatGPT copy the pattern (Quirk 4), which shapes tone far more reliably than describing it.
That's the whole skill: a clear first prompt, then a short series of specific nudges in the same thread until it's right.
When to start fresh instead
Iterating is the default, but sometimes a thread gets tangled: you've changed direction three times and the drafts are contradicting each other. When feedback stops landing and the replies feel muddled, don't keep pushing. Start a new chat, paste in the good version you'd reached, and carry on cleanly. A fresh start with your best draft beats wrestling a confused thread.
Try it now
Common mistakes
- Retyping the whole task instead of nudging. You lose the context the thread was holding. Give feedback on the draft you have.
- Leaving gaps and accepting the guesses. If you don't supply the facts or fence off invention, ChatGPT fills the space plausibly, and wrongly.
- Cramming five instructions into one message. "Shorter and warmer and more formal and add a stat" muddles the result. One nudge at a time.
- Trusting a fluent, well-framed reply as if the framing made it true. A great prompt improves the writing, not the truthfulness. ChatGPT will follow your structure and still state a confident wrong fact inside it, so read critically and verify anything load-bearing before it goes anywhere.
Keeping current
The framework and these quirks are durable, but ChatGPT's defaults shift as models change; newer ones may ask clarifying questions more often, or hedge less. For the latest on getting good results, see OpenAI's ChatGPT release notes and its prompt engineering guide. Accurate as of 13 July 2026.