The Prompt Guide
How to actually get work done with Claude
A working guide for people who want to use AI for real work, not party tricks. Setup, the rules that matter, the patterns that work, and playbooks for web design, finance, business, and real estate. Built by someone using these tools every day.
What's in here
Why Claude vs other AIs
All current frontier models are good. The differences show up when you push hard on a specific kind of work. Here's how the major models actually behave in practice — with Claude as the reference point because it's what most of this guide is tuned for.
| Model | Where it shines | Where it stumbles |
|---|---|---|
| Claude | Long, careful reasoning. Writing in your voice. Refactoring code. Pushing back when you're wrong. Honest about uncertainty. | Slower. Less aggressive about citing recent web content. Will sometimes ask clarifying questions when you wanted action. |
| GPT-4 / GPT-5 | Broad capability across categories. Strong default when you don't know which model fits. | Sometimes confident when wrong. Style can feel generic without prompt steering. |
| Gemini | Multimodal (images, video, long PDFs). Math and structured data. Speed. | Less polished prose. Inconsistent on long-form reasoning. |
| Grok | Distinct voice. Less filtered for edgier creative work. | Less reliable for analytical/professional output. |
| Perplexity | Web-grounded answers with real citations. Best for "what's the current state of X." | Not for creative or open-ended generative work. |
The Ask 'em AI bet: different questions deserve different models. The same prompt sent to four AIs side-by-side often reveals more than any single model would. That's what the home page is for — try it.
Setup essentials
Five minutes that will save you hours later. Skip these and you'll keep wondering why prompts that work for other people don't work for you.
- Pick the right plan for your work. Free tiers are fine for trying things out. Paid (Claude Pro, ChatGPT Plus, Gemini Advanced) gets you bigger context windows, faster models, and far fewer "you've hit the limit" messages mid-thought. If you're using AI more than a few times a day for real work, paid pays for itself in a week.
- Find the model picker. Most chat tools let you choose between a fast/cheap model and a smarter/slower one. Default to the smartest one for anything that matters — the speed difference is often less than you think and the quality difference is huge. Switch to the fast model only for quick lookups.
- Turn on conversation memory or Projects (where available). This lets the model remember context across sessions. For Claude, Projects let you upload reference docs (your style guide, your code, your business plan) and have every prompt in that project use them as context.
- Set custom instructions / system prompt. Tell the model who you are and what you typically want, once. Example: "I'm a real estate agent in Hot Springs, AR. When I ask for help with marketing copy, default to a warm but professional tone, target buyers 40-65, and never use the words 'luxury' or 'stunning' — they're overused in our market." This applies to every chat going forward.
- Bookmark the keyboard shortcuts. Shift+Enter for new line, Enter to send, Ctrl/Cmd+K to open new chats in most tools. Mouse-driven AI use is the slowest way to use AI.
Pro move: Most chat tools support "regenerate" or "edit and resend" the previous message. When an answer is close-but-not-quite, edit your last prompt with one more constraint — don't write a follow-up. Your prompt history becomes a refined draft instead of a sprawling thread.
The eight rules of better prompts
These compound. Use one and your answers improve. Use all eight and you'll get output that feels like it came from a thoughtful colleague.
- Be specific about what you want. "Tell me about Austin" gets you a tourism brochure. "Give me a 200-word overview of Austin's tech scene focused on AI and semiconductors, written for a venture investor evaluating relocation" gets you something useful. Specificity isn't optional — it's the entire game.
- Tell the AI who you are and who the answer is for. "I'm a beginner; explain it for someone who has never seen Python" produces a completely different answer than "I'm a senior engineer; assume Python expertise." This single sentence often does more than the rest of the prompt combined.
- Ask for the format you want. "As a numbered list," "as a table with three columns," "as a 200-word paragraph," "as JSON with keys X and Y." If you don't specify, you get whatever the model defaults to — which is usually prose when you wanted a list.
- Provide examples when you can. If you want the AI to do something tricky, show one or two examples of input and the desired output. This is "few-shot prompting" and it dramatically improves quality. Especially powerful for matching a writing voice or a structured format.
- Iterate, don't restart. When the first answer isn't quite right, reply with "shorter, please" or "now in a more formal tone" or "drop the third point." The model has all the prior context. Restarting from scratch loses everything you just calibrated.
- Give it permission to disagree with you. "If you think my premise is flawed, say so before answering." This unlocks pushback. Otherwise the model often plays along with bad assumptions to be helpful.
- Show your work, ask for theirs. "Walk me through your reasoning step by step before giving the final answer" produces noticeably better answers on hard questions, because the model uses its own intermediate steps as context.
- End with a clear ask. If your prompt is more than a few sentences, finish with the actual instruction restated. "...Given all that context, draft three opening lines for the email." Models lose the instruction in long prompts; restating it at the end fixes it.
Patterns that work
Reusable shapes. Memorize these and you'll have a starting point for almost any task.
| Pattern | Example |
|---|---|
| Role + task | "You are an experienced literary editor. Critique this paragraph for clarity and pacing, then rewrite it." |
| Task + format + constraints | "Summarize this article in exactly 5 bullet points, each under 20 words, focused on what changed for consumers." |
| Step-by-step | "Walk me through this proof step by step. After each step, explain why it follows from the previous one." |
| Compare + contrast | "Compare React and Vue across three axes: learning curve, ecosystem, and performance for SaaS dashboards. Use a table." |
| Critique mode | "Find three weaknesses in this argument and suggest how to address each. Be specific." |
| Translate domains | "Explain microservices using kitchen analogies a line cook would understand." |
| Voice transfer | "How would Richard Feynman explain entropy to a high school student? Match his cadence and willingness to be playful." |
| Devil's advocate | "Argue against this position as forcefully as you can, then steel-man it back." |
| Scope first, then write | "Before drafting, give me a one-paragraph outline. I'll approve it, then you draft." |
| Persona swap | "Rewrite this email three times: once as a CEO, once as a trusted friend, once as a firm but fair vendor." |
Patterns to avoid
These look reasonable but produce thin answers.
- "Write me an essay about X." Too vague. Specify length, audience, angle, tone, and what you want the reader to do or feel after reading.
- "Is X true?" AIs can't reliably verify recent facts. Better: "What are the strongest arguments for and against X? Where is the evidence weak?"
- "Make it better." Better at what? Be specific: "Make it shorter. Cut to the strongest two arguments and drop the throat-clearing."
- "Rewrite this without using AI-style language." Models don't have a clear concept of their own style. Better: paste an example of the style you want, ask them to match it.
- Long context dumps with no question. Pasting 5,000 words and saying "thoughts?" gets you a vague summary. Always end with a specific instruction.
- "Tell me the truth about [contested topic]." Models reflect their training data. Ask for multiple perspectives, then form your own view.
- "Be more creative." Vague. Try: "Give me three options that diverge in tone — one playful, one formal, one ominous."
- Giving the model a personality just to be cute. "Pretend you're a pirate" or "respond as a wizard" wastes tokens and often degrades quality. Use roles only when they actually shape the answer.
Playbook: Web design vertical
🎨 For designers, devs, and people building sites
Claude is genuinely strong at front-end work — layout decisions, copy, accessibility audits, and producing clean HTML/CSS that doesn't look AI-generated. The difference between OK and great here is how much context you give it about your site's voice and constraints.
Use cases where Claude shines
- Drafting copy for hero sections, about pages, feature lists in your existing voice
- Reviewing a page for accessibility issues (color contrast, alt text, heading hierarchy)
- Refactoring a tangle of CSS into something maintainable
- Writing meta descriptions and OG tags that don't sound canned
- Generating content variants for A/B testing
- Producing structured data (JSON-LD) for SEO without you reading the schema docs
Try this prompt
You're reviewing a landing page for a [TYPE OF BUSINESS] targeting [AUDIENCE]. The page is below. Critique it on five axes: 1. Clarity of value proposition (does a visitor understand what we do in 5 seconds?) 2. Friction in the call to action (is what to do next obvious?) 3. Voice consistency 4. Mobile-first considerations I might have missed 5. SEO basics (title, meta, heading structure) For each, give a one-sentence verdict and one specific change you'd make. [paste HTML or screenshot description]
Try this prompt: copy in a known voice
Below are three pieces of writing from my existing site to establish voice. Then I'll give you a new page to write copy for. Voice samples: [paste 3 paragraphs from your existing site] Now write the hero section (h1 + 2-sentence subhead + CTA button text) for a new page about [TOPIC]. Match the voice exactly. Three options, then tell me which you'd pick and why.
Avoid
- Asking for "modern, sleek, minimalist" design copy — you'll get clichéd output. Show, don't tell, with samples.
- Letting Claude generate emoji-heavy marketing copy unless you actually want that. It will reach for emojis without being asked.
- Trusting it to know current CSS feature support — cross-check anything past container queries.
Playbook: Finance vertical
💰 For personal finance, small business books, and analysis
Claude is good at the explanatory and analytical side of finance — summarizing statements, modeling scenarios, drafting communications. It is NOT a substitute for a CPA, fiduciary, or legal advice. Use it to think faster, not to act on its conclusions blindly.
Use cases where Claude shines
- Translating dense statements (10-K excerpts, fund prospectuses, mortgage docs) into plain language
- Building scenario tables ("if rates do X, monthly payment is Y")
- Drafting polite-but-firm communications to lenders, vendors, tax authorities
- Categorizing transactions when you paste a CSV
- Explaining tax concepts for your specific situation (with the disclaimer that you'll verify with a pro)
- Writing investor updates or month-end summaries from your numbers
Try this prompt: scenario modeling
I'm considering [DECISION, e.g., refinancing my mortgage]. Current state: [numbers]. Three options on the table: [list with rates, terms, fees]. For each option: - Total cost over 5, 10, 30 years (assume I stay) - Total cost if I sell in 7 years - Break-even point in months - One non-obvious risk Then recommend one, and tell me what would have to be true for that recommendation to be wrong.
Try this prompt: making a CSV usable
Below is a CSV export of my business expenses for Q3. Categorize each row into one of these buckets: Software, Travel, Marketing, Contractor, Office, Other. Output a markdown table with totals per bucket and a flag for any row that looks unusual or possibly miscategorized for me to double-check. [paste CSV]
Hard rule: Never act on AI tax, legal, or investment advice without a human professional reviewing it. Use Claude to prepare smarter questions for that meeting, not to skip the meeting.
Playbook: Business vertical
🏢 For operations, communications, and decisions
This is where most knowledge workers will get the highest return. Email triage, meeting prep, doc review, decision frameworks — the model can shave hours off a week without anyone being able to tell the work was AI-assisted.
Use cases where Claude shines
- Inbox triage: paste a thread, ask "what's the actual ask, what's the deadline, what should I respond?"
- Meeting prep: "I'm meeting [PERSON] about [TOPIC]. Here's their LinkedIn / their last email / our prior history. What should I ask, what should I avoid, what does success look like?"
- Drafting tough conversations: difficult feedback, vendor disputes, partnership disagreements
- Reviewing contracts at a comprehension level (not legal review): "what does this clause actually obligate me to?"
- Brainstorming with structure: "give me 10 options, then your top 3 with reasoning, then your single recommendation"
- Summarizing long documents into the 5 things that matter
- Writing job descriptions, performance reviews, OKRs in your company's voice
Try this prompt: the decision framework
I need to decide whether to [DECISION]. Context: - [What's true today] - [What I know about the options] - [What I'm worried about] - [What I want to be true in 12 months] Don't recommend yet. First: 1. Identify the 2-3 most important factors 2. Tell me what I haven't said that I should have considered 3. List 3 questions I should answer before deciding Then, once I respond, give me your recommendation with confidence level.
Try this prompt: the email triage
Here's an email thread. Tell me: - What's the sender actually asking (one sentence) - What's the deadline or implicit timing - What would a great response look like (3 bullets, no draft yet) - Anything in this thread I might be missing or misreading [paste thread]
Playbook: Real estate vertical
🏡 For agents, investors, and STR operators
Real estate is content-heavy, communication-heavy, and full of repetitive structured tasks — exactly where AI compounds value. Listing copy, market briefs, buyer/seller emails, neighborhood research, comp analysis prep. Local market knowledge stays with the human; the model handles everything around it.
Use cases where Claude shines
- Listing descriptions in your voice (not the generic "stunning luxury retreat" template)
- Neighborhood/area write-ups for buyer relocation packets
- Drafting price-reduction conversations with sellers
- Personalized buyer-tour follow-up emails based on what they reacted to
- Translating buyer/seller communications between agents and lawyers
- STR descriptions tuned to what your specific guest segment cares about
- Market brief content for your newsletter (you provide the data, AI makes it readable)
- Drafting Zillow / Realtor.com / MLS descriptions in three different lengths
Try this prompt: a listing description that doesn't sound like every other listing
Write a 180-word listing description for the property below. Audience: families relocating to [CITY] for [REASON, e.g., remote work]. Voice: warm, specific, never uses the words "luxury", "stunning", "boasts", or "nestled". Property details: - [bedrooms, baths, sqft] - [3-5 standout features] - [neighborhood angle, what's nearby] - [one quirky thing — old fireplace, specific tree, view] End with one sentence inviting a private tour. Don't put my contact info — that's added by the platform.
Try this prompt: the buyer follow-up
I just showed [BUYER FIRST NAME] three properties: 1. [address] — they liked [thing], didn't like [thing] 2. [address] — they liked [thing], didn't like [thing] 3. [address] — their reaction They're prioritizing [criteria]. Their timeline is [timing]. Draft a follow-up email that: - References specific things they reacted to (not generic) - Suggests next step without being pushy - Includes one new property I should consider showing them based on their reactions - Feels like it came from a person who was actually paying attention Tone: warm, professional, lightly personal. No emojis.
Local-knowledge wedge: AI doesn't know your market. Your value as the human is the granular local truth: which schools are actually good vs. on paper, which streets flood, which builders cut corners. Feed that local knowledge IN to the model as context, then let it do the writing.
Beyond the chat box
The chat interface is the easy door in. Real productivity gains often come from one of these next layers.
Claude Projects
Upload reference docs once (style guide, codebase, business plan). Every chat in that Project uses them as context. Stops you from re-pasting.
Claude Code (CLI)
Claude in your terminal, with file editing, command running, and codebase awareness. Massive multiplier for anyone who works in code or text files. Free with Claude Pro.
The Anthropic API
Build Claude into your own apps and scripts. Pay per token, no monthly cap. Worth it once you have a workflow you're running multiple times a week.
MCP servers
"Model Context Protocol" lets Claude talk to your other tools — Gmail, Calendar, Drive, GitHub, databases. Set up once, use forever from any chat.
Custom GPTs / Claude Projects
Save a specialized assistant tuned for one task ("draft my listings", "review my CSS"). Skip the setup prompt every time.
Browser extensions
Right-click any webpage to summarize, translate, or ask questions about it. Different vendors offer this; pick one and stop opening new tabs.
Common pitfalls
- Treating AI output as final. First drafts are first drafts. Read carefully, especially numbers, names, and dates.
- Pasting confidential data into free public tools. If it matters, use the paid tier or API where data isn't used for training. Read the privacy policy.
- Asking the model "are you sure?" It will often back down regardless of whether it was right. Better: "what would have to be true for that to be wrong?" or "what's the strongest counter-argument?"
- Believing confident-sounding citations. Models can fabricate sources that look real. Always click through.
- Losing your own voice. Use AI for first drafts. Rewrite enough that the result sounds like you, not like the model. Otherwise readers can tell — and so can search engines.
- Endless prompt-tuning when one human edit would be faster. If you've sent four follow-ups and it's still not right, edit by hand and move on. AI is a tool, not a religion.
- Asking for help instead of doing the thinking. The best prompts come from people who already half-know the answer and use AI to sharpen it. If you don't know what good looks like, AI can't help you get there.
Where to go next
A few paths from here.
Try a side-by-side
Open the main page, paste a real prompt from your work, send it to 4 models, and see which voice fits your task.
Browse the resources
Curated tools, courses, and reading on the Resources page.
Check the FAQ
Privacy, pricing, "which model when" — on the FAQ page.
Watch the news
What's actually changing week to week in the AI space — on the News page.
Found a missing pattern, broken example, or playbook this guide should have? Email walhus@gmail.com. This guide is meant to be revised — the goal is real usefulness, not finished polish.