Prompting Basics for Claude and ChatGPT
Lloyd Evaroa, Founder, Really Good Marketing
Tried Claude or ChatGPT and got back something flat, generic or just not quite there?
Getting sharp, usable output from AI comes down to how you structure your request. The eight techniques below give ChatGPT and Claude and likely most LLMs the explicit constraints they need to get you a higher quality response, faster.
The theme here: the idea, then a weak prompt sharpened into a stronger one.
Explain the reason behind the ask
A large language model is, underneath everything, just an enormous collection of knowledge. Giving it a reason gives it a frame of reference, and that's how you get access to the right slice of that knowledge instead of an average of everything it knows.
Take that cold lead. "Write a follow-up email" gets you a bland nudge. "Write a follow-up, I need to reopen this without sounding pushy, because if I push too hard this early I could kill a deal that's still alive, but if I leave it another week I've probably lost it anyway" gets you something that actually reads the situation.
“Write a follow-up email.”
“Write a follow-up, I need to reopen this without sounding pushy, because if I push too hard this early I could kill a deal that's still alive, but if I leave it another week I've probably lost it anyway.”
I recently watched Sandeep Swadia build this into the first letter of his PRIME framework, purpose before anything else: give Claude a precise goal, not just an instruction. Same idea, different route to it.
Hand it a role
This one's a small change but it has big impact. Give your AI a persona to work from before it starts. "Take on the role of a sales manager who's talked dozens of dead leads back to life, help me write this follow-up." A role narrows the model down to the right register straight away, instead of leaving it to average across everything it could be.
“Help me write this follow-up.”
“Take on the role of a sales manager who's talked dozens of dead leads back to life. Help me write this follow-up.”
Give it something to aim at
A real example does more work than any description. For the cold lead, that might be the last follow-up that actually got a reply, or a message from someone else whose tone you admired. Hand it over and say aim for this.
Vicky Zhao calls this one-shot and few-shot prompting, and her test case makes the point well: ask for a "playful" summary and you get something bland, but hand it one line of the tone you're after and the whole output shifts. She takes it further too, giving a few examples of what she wants and a couple of what she doesn't, together, in the same prompt.
“Write a follow-up in a warm, professional tone.”
“Write a follow-up in this tone: [paste the last follow-up that actually got a reply]. Match that voice.”
If you don't know what you want yet, say that too
Not sure whether to nudge again or let this one go. That's fine, say so, and ask for options first: "give me three different ways to reopen this, one low-key check-in, one with a genuine reason to reach out, one direct ask for a decision." Read the three, pick the one that fits, then switch modes: "help me tighten this one up, ready to send."
“Write the follow-up now.”
“Give me three different ways to reopen this: one low-key check-in, one with a genuine reason to reach out, one direct ask for a decision.”
Vicky names this split directly, explore versus execute, and says most people blur the two without realising it. Vague is the right call when you're exploring, you want options, not precision. Ask for the finished piece once you have certainty otherwise you will end up with many rounds of revision.
Take it one step at a time
Same split applies to the research behind the follow-up. First, what actually happened: "here's the quote and the email thread, what did they ask about, what did the quote cover, how long has it actually been since they went quiet." Once that's straight, then the output: draft the follow-up based on it.
“Draft a follow-up for this lead.”
“Here's the quote and the email thread. What did they ask about, what did the quote cover, and how long has it actually been since they went quiet?”
Will Francis runs this as a loop on anything that has to be accurate: get it searching or reading first, then get it to check what it found, rather than trusting whatever it already assumes. The same logic applies here, get the facts of the deal straight before it writes a word to the client.
Get it to interview you first
If you had to pick just one of these tactics, pick this one. Before it drafts the follow-up, get it to ask you three questions first, how long since the quote went out, whether they raised any objection, what you're actually willing to flex on if price turns out to be the sticking point. It's a genuine game changer, and on paid plans it saves tokens too: less back and forth.
“Write the follow-up now.”
“Before you draft this, ask me three questions to help clarify the task so you're better informed.”
Will puts it plainly: letting the AI just get on with it "makes no sense, it can literally 10x the quality of the results if you just give it the prompt to ask you a few questions first." Sandeep builds the same move into PRIME as the I, interview, and reckons it's the step people skip most.
When it goes off track, fix the start, not the middle
Say the first draft comes back sounding like every other "just checking in" email ever sent.
Stack a fourth message onto the thread trying to course-correct the tone.
Go back to your very first prompt, fold in what you've learned since, and send that instead.
Nine times out of ten, in my own experience, that alone sorts it. Vicky points at the research behind why, a study out of the University of Washington, MIT and Google finding that these models weight the start and end of a conversation far more heavily than the middle. Correct three messages deep and there's a real chance it's already lost. Edit the beginning and it isn't.
Ask it to mark its own work
Once there's a draft, hand it a role again. "You're a sales manager reviewing this before it goes out, tell me what would make the lead delete it without reading." Or go further, Will's move: grade it against the five things most likely to get a reply, explain why it isn't a ten on each, then rewrite, and keep going until it is.
“Is this good to send?”
“You're a sales manager reviewing this before it goes out. Tell me what would make the lead delete it without reading, then rewrite it.”
Feels a bit strange asking it to mark its own homework. It isn't capable of cheating the way a person might be tempted to. It just takes the long way round and thinks harder because you asked it to.
Bringing it all together: the mega prompt
Once you understand these eight building blocks, you don't always have to use them one by one. When you need a high-stakes response fast, stack them into a single mega prompt that hits every constraint at once:
You are an expert B2B sales strategist who specialises in reviving quiet leads without sounding pushy.
A prospective client went quiet three weeks after I sent a project quote. I need to reopen communication: if I push too hard I'll kill a deal that's still alive, but if I wait another week the deal is lost.
Match the tone of this successful past follow-up: [paste the email that got a reply].
- Don't write the draft yet. First, ask me three brief questions about the account history and price flexibility to get your facts straight.
- Once I answer, give me three distinct approaches: a low-key check-in, a value-add update, a direct decision ask.
- After I pick an approach, draft the email, review it against common objections, and refine it to cut any generic corporate fluff.
It's usually bad practice to introduce something new in your conclusion, but I'm breaking that rule here, from my own experience. You start out prompting properly, then over time, because the output's good enough anyway, your prompts get shorter, and shorter, and shorter. What that actually costs you is more time in round two, three, four, chasing the result you wanted first go.
That interview-first loop, ask before you draft, review before you send, is the same shape I build into ongoing systems for clients through AI Operations, so it runs on its own instead of needing you to retype the prompt every time.
If you want to go into more depth on this topic, happy to grab a coffee and talk it through with you.
Sources
- Sandeep Swadia, 7 Ways to Use Claude So Well It Feels Illegal
- Vicky Zhao, ChatGPT Prompting Techniques No One Talks About...But Should!
- Will Francis, 3 things to add to ChatGPT or Claude prompts