A lot of prompt engineering advice treats it like a secret skill: use the word "expert," add "think step by step," use role-play. Most of that is cargo cult. For SEO work, prompting comes down to three things:

1. Understanding what the AI can and can't do 2. Telling it exactly what you want 3. Knowing when to iterate vs. start over

Skill 1: Know the boundaries

Claude (and other LLMs) are good at:

Claude is bad at:

A bad prompt for SEO work:

"Analyze these 500 domains and rank them by domain authority."

Without a data tool, Claude can't look up those domains or their DA scores, so it would make up numbers.

A better prompt:

"I have a list of domains and their current DA scores [paste list]. Rank them by growth potential, considering their niche and backlink profile. Explain your reasoning for the top 5."

Here you're giving Claude the data and asking it to analyze what it can see, instead of fetching what it can't.

Skill 2: Be very specific

This prompt is bad:

"Write a blog post about keyword research."

This prompt is good:

"Write a 1,500-word blog post in my voice (direct, specific, practical) targeting the keyword 'keyword research tools for SEO professionals.' Include:

Do NOT include: hedging language, 'it's important to note', em dashes, tricolons."

The second prompt will produce something useful and the first will produce something generic, because the second tells Claude exactly what I want, including what not to do.

Skill 3: Change the direction when you iterate

Bad iteration loop: "Write a blog post about SEO" [Claude produces generic post] "Make it better" [Claude adds more words, doesn't fix the problem] "Make it more specific" [Claude mentions specific tools, but the structure is still weak]

Good iteration loop: "Write a blog post outline about keyword research" [Claude produces outline] "This outline is too linear. Reorganize around specific client problems instead of general topics" [Claude restructures] "Good. Now expand Section 3 into 800 words of practical steps" [Claude does] "This sounds like AI. Rewrite emphasizing my personal experience and specific numbers" [Claude rewrites with texture]

In the second loop, each request changes the direction based on what's actually wrong, instead of asking for surface tweaks.

Skill 4: Show examples instead of describing

Bad prompt:

"Write it in a conversational tone"

Good prompt:

"Write it like this previous post of mine [paste excerpt]. Same rhythm, same specificity, same voice."

Claude is better at matching examples than following abstract descriptions. If you want conversational writing, show it some; if you want specifics, show it specific examples.

For SEO, this means:

Instead of: "Write content that ranks well" Do this: "Write content like [this page that ranks #1 for the target keyword]. Match its structure, depth, and approach."

Skill 5: Know when to start over

If you've iterated five times and the core output is still wrong, stop iterating. Start a new conversation.

Bad workflow:

Better workflow:

New conversations are cheaper (faster API response) and produce better output (no context carryover from failed attempts).

The meta skill: know what you want

This one matters most. Most people don't know what they want from the AI; they only know something is wrong with what came back.

Before you prompt Claude, write down in a sentence:

Your prompts get much better once you've done that.

Example from my work:

Bad starting point: "Write a blog post about link building"

Good starting point: "I want to convince freelancers that outreach volume isn't where link programs fail anymore. Vetting is. They should finish the post thinking 'I need to audit what I'm buying before I renew it.' The post should acknowledge that vetting takes time but it's cheaper than paying for links that don't hold."

With that much clarity up front, the first draft lands much closer to what you meant.

The 10-minute prompt session

My usual route to a first draft:

1. Open a new Claude conversation 2. Dump my raw notes, numbers, and ideas (5 min) 3. Ask Claude: "Given this context, what's the core argument I should make?" (2 min) 4. Claude responds with a thesis 5. I say: "Yes, that. Now write the post" 6. Claude writes it 7. I rewrite the sections that need my own experience 8. Draft done

That gets me to a solid draft in about 10 minutes. It doesn't get me to a published post, because every factual claim still gets checked against a source before anything goes live. The session saves drafting time and doesn't replace review.

The key is step 3: let Claude help you clarify what you actually want before you ask it to write.

The bigger picture

Prompt engineering comes down to:

Most "prompt engineering courses" teach you to say "act as an expert" and add magic tokens to your prompts. In my experience that matters far less than understanding the model's limits and knowing what you want.