Kashi's 2026 ChatGPT Prompting Guide

Most people use ChatGPT like a search engine with better grammar. They type a vague question, get a generic answer, and wonder why it doesn't feel that useful. The gap between mediocre output and expert-level output isn't the model — it's the prompt. This guide breaks down the principles I actually use every day to pull sharp, specific, usable work out of ChatGPT. No tricks. No jailbreaks. Just the craft of writing prompts that get results.

The Single Most Important Shift: Stop Asking, Start Assigning

Questions invite generic answers. Assignments produce specific output. When you phrase a prompt as a question — 'What are some ways to get more clients?' — ChatGPT gives you a listicle you could have Googled in 2012. When you phrase it as a job to do, something changes.

The assignment frame means you treat ChatGPT like a contractor, not a librarian. You tell it what you need, what format it should deliver in, and what constraints apply. You don't ask what it thinks. You tell it what to produce.

Compare these two prompts for the same goal:

What are some good ways to get coaching clients?

You are a B2B marketing strategist. Write a 5-step outbound sequence for a business coach targeting HR directors at companies with 50–200 employees. Each step should include: the channel (LinkedIn, email, or DM), the core message angle, and the exact first sentence of the message. Keep each step under 100 words. Tone: direct, no corporate fluff.

The second prompt gets something you can actually use Monday morning. The first gets advice you already knew. Make this shift and everything else in prompting becomes easier.

Role + Context + Constraint: The Three-Part Prompt Frame

Once you're thinking in assignments, you need a reliable structure to build them. The frame I use on nearly every serious prompt is: Role, Context, Constraint.

• Role — Tell ChatGPT who it is. 'You are a conversion copywriter with 10 years of experience writing sales pages for online courses.' This activates a specific lens and vocabulary.
• Context — Tell it what's true about your situation. Your audience, the offer, the channel, what you've already tried, what isn't working. The more specific, the better the output.
• Constraint — Tell it what's off limits or what the output must look like. Length, format, tone, things to avoid, things to include. Constraints make ChatGPT actually decide instead of hedging.

Here's what this looks like assembled into a real prompt for a coach writing a cold email:

You are a direct-response copywriter who specializes in cold outreach for service businesses.

Context: I'm a business coach targeting wellness retreat centers. My core offer is a 90-day private coaching engagement to help owners fill their retreats without paying for ads. The typical owner is doing $300k–$800k/year, burned out on social media, and skeptical of coaches.

Task: Write a cold email subject line (5 options) and a single cold email body (under 150 words). The email should open with a specific observation about their business or industry — not a generic compliment. No hollow phrases like 'I came across your profile and was impressed.' End with one low-commitment CTA (a question, not a calendar link).

Tone: Peer-to-peer, not salesy. Treat the reader like a smart operator.

That prompt structure makes ChatGPT do real work. You get five subject line options you can A/B test and an email that sounds like it was written by someone who actually knows the retreat industry — because the context you provided gave ChatGPT enough to pattern-match against.

Using System-Level Instructions to Set the Defaults

If you're using ChatGPT's Custom Instructions feature (under Settings), you're leaving power on the table if you haven't filled it out. This is effectively a persistent system prompt that runs before every conversation. You set it once and it shapes every response.

In the 'What would you like ChatGPT to know about you?' field, put who you are, who your audience is, and what you're building. In the 'How would you like ChatGPT to respond?' field, put your tone rules, format preferences, and what it should never do. Here's a real example of what I use:

About me: I'm a marketing operator and AI educator working with coaches and small business owners. My brand is sharp, practical, and no-hype. I run funnels, write copy, and build AI systems for client acquisition.

My audience: Online coaches, consultants, and service providers who are skeptical of tech but open to results.

How to respond:
- Write at an 8th-grade reading level unless I say otherwise
- No bullet points unless I specifically ask for a list
- Short paragraphs, active voice
- Never add a disclaimer unless the topic legally requires one
- If I ask for copy, write it — don't explain what you're about to do first
- If something in my prompt is unclear, ask one clarifying question before proceeding

That last instruction — ask one clarifying question if unclear — is underrated. It stops ChatGPT from guessing and producing something half-right. A one-question clarification loop almost always produces better output than a second full prompt.

Iteration Is Part of the Process, Not a Sign the Prompt Failed

One-shot prompting gets you 70% of the way on most tasks. The other 30% comes from directed iteration. This isn't a workaround — it's how professional users actually work. ChatGPT holds context within a conversation. Use that.

The most effective iteration moves are directional corrections, not rewrites of the whole prompt. You tell it what was wrong about the last version and what to change — not re-explain everything from scratch.

• Too generic? Say: 'The second paragraph is too vague. Rewrite it with a specific example from the wellness coaching industry.'
• Wrong tone? Say: 'This is too formal. Rewrite in a casual, peer-to-peer voice — like one operator talking to another.'
• Too long? Say: 'Cut this by 40% without losing the three core arguments. Remove anything that restates what was already said.'
• Structure off? Say: 'Flip the order — lead with the outcome, then explain how it works. Right now it buries the most compelling part.'
• Missing a key element? Say: 'Add a transition sentence between section 2 and 3. Right now it feels like two separate pieces.'

Iteration is also where you refine copy into something that sounds like you, not like a chatbot doing an impression of you. Take the output, mark what landed, tell ChatGPT to keep those parts and fix the rest. Three rounds of this and you often have something genuinely strong.

Prompt Patterns That Work Across Dozens of Use Cases

Beyond the core frame, there are a handful of prompt patterns that show up in high-quality outputs across very different tasks. These are worth learning as reusable tools.

The 'Devil's Advocate' pattern is one of the most useful for thinking work. You write a plan, argument, or strategy, then ask ChatGPT to attack it as a skeptic would:

Here is my plan for launching a group coaching program: [paste your plan]

Now respond as a skeptical potential buyer who has bought online programs before and been burned. What objections do you have? What's missing? What would stop you from buying? Be honest and don't hold back. I want 5 specific objections, not general ones.

The 'Steel Man' pattern works in the opposite direction — you ask ChatGPT to make the strongest possible case for a position before you critique it. This is useful when you're building counter-arguments or writing content that addresses objections fairly.

The 'Format Mimic' pattern lets you give ChatGPT an example of the style you want, without explaining it:

Here is an example of copy in the style I want:

[paste 2–3 sentences of your own writing or writing you admire]

Now write a 3-email welcome sequence for a new subscriber to my AI coaching newsletter. Match the tone, sentence length, and voice of the example above. Do not be more formal or more casual than the example.

The 'Think Step by Step' instruction is not hype — it genuinely improves reasoning outputs on complex tasks. Add it when you're asking ChatGPT to analyze something, solve a multi-variable problem, or build a strategy. It slows the model down in a way that improves accuracy.

What to Stop Doing Right Now

Most prompting mistakes aren't about missing a trick. They're about habits that consistently produce bad output. Cut these and your results improve immediately.

• Vague verbs: 'Help me with my marketing' tells ChatGPT nothing. Replace 'help me with' with a specific deliverable every time.
• Skipping the audience: If ChatGPT doesn't know who this is for, it defaults to 'general reader' — which means nobody in particular.
• Asking for too many things in one prompt: One prompt, one output. If you need five things, build five prompts or chain them sequentially in a conversation.
• Accepting the first output uncritically: The first draft is a starting point. It's not a finished product. Read it like an editor, not a recipient.
• Restating the whole prompt when iterating: ChatGPT remembers the conversation. Just direct the change you want — don't re-explain the whole context.
• Using ChatGPT as a fact source without verification: ChatGPT is excellent at reasoning, structure, and language. It is not reliable for current events, specific statistics, or anything where being wrong has real consequences. Verify independently.

Build a Prompt Library, Not Just Skills

The highest-leverage thing you can do once you start getting good outputs is save them. Not the output — the prompt that produced it. A prompt library turns a one-time win into a repeatable asset.

I keep mine in a simple Notion database with columns for: task type, the prompt text, the model used, and a 1–5 rating on how well it worked. When I need something similar, I pull the closest match and adapt it. This takes 20 seconds and consistently outperforms writing a new prompt cold.

Organize by function, not topic. Useful categories for operators and coaches:

• Cold outreach (email, LinkedIn DMs, direct mail scripts)
• Content creation (newsletter intros, social hooks, blog outlines)
• Sales copy (landing page sections, offer descriptions, objection handling)
• Client work (discovery call prep, proposal frameworks, follow-up sequences)
• Thinking tools (strategy critiques, competitive analysis, decision frameworks)
• Operations (SOPs, onboarding docs, delegation templates)

Within six months of consistent use, your prompt library becomes one of the more valuable things in your business. Every good prompt in it represents a task you no longer have to figure out from scratch — and that compounds fast.

The bottleneck is almost never the AI. It's the clarity of your input. Get sharper at specifying what you want and ChatGPT becomes significantly more useful — without any change to the model, the plan, or the tool. That's the skill worth building.