7 Rules for Getting Real Work Out of AI (Claude, ChatGPT & Gemini)

7 Rules for Getting Real Work Out of AI (Claude, ChatGPT & Gemini)

The fastest way to get better output from AI is to stop prompting like it's a search engine. Most operators using Claude, ChatGPT, and Gemini type a question, take the answer, and quit when it's wrong. Better output comes from giving the AI a role, explaining the why, loading a brand voice file, and packaging recurring tasks into reusable skills. Here are the 7 rules I follow every day, in any AI chat I'm paying for.

Key Takeaways

• Structured prompts (Role, Context, Task, Output, Example) produce structured output. Wall-of-text prompts don't.
• The reason behind the ask changes the output more than the ask itself.
• The strongest brand voice file is one the AI helped you write — through an interview, backed by real docs, links, visuals, and ads.
• One job per chat. Old context bleeds — /clear (Claude Code) or a new browser chat between tasks.
• Giving AI permission to say "I don't know" cuts hallucinations in half.
• Packaging recurring tasks into a Skill (Claude), Custom GPT (ChatGPT), or Gem (Gemini) saves hours every week.
• Never paste passwords, keys, or customer data into an AI chat window.

Why does AI give you garbage output most of the time?

Because most people prompt it like they're typing into a Google search bar.

They give it a task with no role, no context, no output format, no why. Then they're surprised when the draft sounds generic, misses the point, or invents facts.

The tool isn't broken. The prompt is.

The 7 rules below are what changed my output — from AI slop to drafts I actually ship. If you want the broader mindset first, our post on how to talk to AI for your business sets the frame these rules build on.

What structure should every AI prompt follow?

Break the prompt into five short blocks:

• Role: who the AI is playing. "You're a copywriter for a family law attorney."
• Context: who it's for, the brand voice, the constraints. "Our audience is people going through divorce in Riverside County, ages 35–55."
• Task: the specific ask. "Write a follow-up email to a lead who didn't book after their free consultation."
• Output: the shape you want back. "Plain text. 5 sentences max. No subject line."
• Example: paste a real one that hit the mark. "Here's a follow-up I sent last month that got a booking — match that voice, swap in this lead's details."

Structured prompts get structured output.

"Make it good" is not a spec. Show it a good one and it'll copy your homework.

Why does explaining the "why" change AI output?

Because the reason behind the ask changes the decisions the AI has to make.

"Write a follow-up email to a lead who didn't book" gets you a polite template. Goes straight to trash.

"Write a follow-up email to a lead who didn't book. She got two other quotes, mine was the highest, and she told my tech she's 'thinking it over.' Address the price gap head-on — we're the only shop in town that pulls a permit on every job and guarantees the install for 2 years. End with one specific question. No pushy close" — now you get an email you'd actually send.

Same task. Completely different output. The why does the work.

How do you use AI to build your own brand voice file?

Skip the "sound like me" instructions. Do this once and never again.

Open a chat and paste this:

"Interview me until you can write in my brand voice. Ask me 20 questions about my business, my customers, my products, the words I own, the words I'd never use, and how I sound in an email vs on a sales call. Then compile everything into a brand voice guide with real examples."

Answer every question. Then feed it your assets:

• Documents: real emails, sales scripts, past newsletters, blog posts, LinkedIn posts, sales-call transcripts.
• Links: your website, YouTube videos, podcast appearances, customer reviews, interviews.
• Visuals: logo file, brand colors (hex codes), fonts you use, product photos.
• Ads: Facebook, Google, and print — the ones that worked and the ones that flopped. Both teach it something.

The more you feed it, the sharper the guide. The AI only knows what you show it.

What comes back is a permanent asset. Save it as a file.

Every new session starts the same way: "Load my brand voice guide before you write anything."

The AI stops guessing at your voice. You stop rewriting drafts.

Claude Skill built for exactly this: /brand-voice

If you use Claude, there's a skill built for this — three moves under one skill:

• /brand-voice:discover-brand — hunts through your connected tools and pulls every scrap of brand material you've already written into a discovery report.
• /brand-voice:generate-guidelines — takes that report, plus any docs, transcripts, or examples you feed it, and builds the actual voice guide.
• /brand-voice:enforce-voice — reads any new draft and flags where it drifted off your voice. Spell-checker for brand.

Interview once. Assembly and QA on autopilot from there.

Why should you let AI plan the task before it builds?

Because a bad plan is easier to fix than a bad draft.

For anything bigger than a one-liner, ask the AI to write the plan first — anchored on your why. Have it list the steps. Then have it execute one at a time.

You spot the misfires when they're 3 bullets long, not when they're a 700-word draft.

No plan means rewriting whatever it hands you.

How do you stop AI from hallucinating?

Give it permission to be uncertain.

Tell it up front: "If you don't have enough information, ask. Flag anything you're guessing at."

Most AI mistakes come from it faking confidence to sound helpful. Take that pressure off and it stops making things up.

You'll get questions instead of guesses. Guesses get shipped. Questions get answered.

Why start a fresh chat between AI tasks?

Because old context bleeds into new answers.

If you drafted a newsletter in a chat and then ask the same chat to write a follow-up email, the AI carries every assumption from the first task into the second. Same brand voice — good. But also every scope constraint, every specific number, every tangent. That drags the new draft off course.

One job per session. New job, new chat.

Reviewing the AI's own draft? Also a new chat — otherwise it defends the work like a proud toddler instead of catching the problem.

Here's how to actually do it in each tool:

• Claude Code (the CLI): type /clear — resets the session on the spot.
• Claude.ai, ChatGPT, Gemini (in the browser): click the "New chat" button in the sidebar.
• ChatGPT shortcut: Cmd + Shift + O on Mac, Ctrl + Shift + O on Windows.

Old context bleeds. Kill it between tasks.

How do you turn a recurring AI task into a reusable skill?

Package it.

If you're typing the same setup instructions three times in a week and not packaging them, that's on you.

Every AI tool now has a customization layer:

• Claude has Skills
• ChatGPT has Custom GPTs
• Gemini has Gems

A skill is a saved procedure. You define the job once. The AI runs it the same way every time.

Here's how to build one, in four steps:

1. Pick a task you do over and over. A weekly newsletter. A follow-up email after a free consultation. A job scope for a specific service.
2. Open a chat and say: "Help me turn this into a repeatable skill. Interview me about how I do this task — what starts it, the steps I take, the format I want back, the rules I always follow, and anything I never want. Then write the skill instructions for me."
3. Answer every question. Be specific. File names. Tone rules. Length limits. Which brand voice guide to load. What the finished output should look like.
4. Save the output as a Skill (Claude), Custom GPT (ChatGPT), or Gem (Gemini). Give it a name you'll remember.

Next time you need that task done, you fire the skill and it runs the whole procedure. Same inputs. Same output shape. No re-typing the setup.

I have skills for newsletters, LinkedIn posts, blog articles, and job scopes. Each one loads my brand voice guide, my rules, and my file-naming system before it writes a word.

20 minutes to build one. Hours saved every week. If you're not doing this, you're paying yourself minimum wage to type.

For a deeper look at where AI actually earns its keep in a service business, see our breakdown of how to win with AI in 2026.

What should you never paste into an AI chat?

Passwords. API keys. Customer names. Phone numbers. Payment info. Anything that could hurt you or a customer if it leaked.

AI tools ask for this all the time. Don't do it.

If the AI genuinely needs a piece of sensitive data to help you, use a placeholder. "The API key is [REDACTED]" or "The customer's name is [Customer]."

Any tool worth using can figure it out from the placeholder.

Bonus: talk to your AI instead of typing

I hate typing. So I don't.

I use the Whispr app to talk every prompt straight into the chat bar. About 3x faster than my thumbs. Hands stay free.

When I'm brainstorming — thinking out loud, not writing — I flip to voice mode. Claude, ChatGPT, and Gemini all have one. You talk. It talks back. Ideas come out faster because you're not stuck in edit-mode while they form.

Try both. See what sticks.

One more thing: these models are getting stupidly better every 30 days

Faster. Smarter. Doing things last month's version couldn't touch.

If you sat out the last two years because ChatGPT wrote you one bad haiku, you're getting lapped.

Learn the rules now. Every upgrade, you get more output with the same prompts. Compounding, on autopilot.

Bottom line

Most people use AI like a search engine. Type a question, take the answer, quit when it's wrong. Then they complain the answer sucked.

Yeah — because they spent 4 seconds on the prompt.

The operators actually getting results out of it treat it like a new hire — with a role, training, a plan, a written playbook, and permission to say "I don't know."

Same tool. Better output. Way fewer rewrites.

If you want an AI system in your business that actually books jobs — not one that drafts emails you'll rewrite anyway — start with a Free Growth Assessment.

Frequently Asked Questions

Do these rules work for ChatGPT and Gemini, or only Claude?

All three. The rules are about how you prompt, not which model you're using. Every major AI tool now supports the same core moves — custom instructions, saved contexts, and reusable skill packages. The names change (Skills in Claude, Custom GPTs in ChatGPT, Gems in Gemini) but the structure holds.

What if I don't have any writing samples to feed the AI for a brand voice file?

Start with what you have. Even one email, one voicemail transcript, and one customer review is enough to seed the interview. As you build more content, feed the new pieces back in and have the AI update the guide.

How long should my Role and Context blocks be?

One or two lines each. The scaffold is about clarity, not volume. If the Role is "copywriter for a service business," don't pad it to a paragraph. The Task and Output blocks are where you spend the words.

Can I skip the plan step if the task is short?

Yes. For a one-liner — a subject line, a headline, a single tweet — plan-first adds friction. Use it for anything with more than one moving part.

Does asking AI to flag guesses actually reduce hallucinations?

Yes, but not to zero. Most hallucinations come from the AI trying to sound helpful when it doesn't have the answer. Once you signal that "I don't know" is an acceptable response, the AI is more likely to ask a clarifying question than to invent one.

What information should I never paste into an AI chat?

Passwords, API keys, customer names, phone numbers, payment info, or anything that could hurt you or a customer if it leaked. Use a placeholder instead. Any tool worth using can figure out the task from a redacted stand-in.

Sources

• Anthropic, "Best Practices for Claude Code," Anthropic Engineering (2025) — https://www.anthropic.com/engineering
• Anthropic, "How Anthropic Teams Use Claude Code" (2025) — https://www-cdn.anthropic.com/58284b19e702b49db9302d5b6f135ad8871e7658.pdf
• Claude Code Documentation, code.claude.com/docs (accessed 2026)

FAQs