Read Prompts That Ship free: the opening pages

Part 1, Setup and the Context-First Method

You have done this. You open ChatGPT, you type "write a LinkedIn post about our new feature," and you get back something that sounds like every LinkedIn post ever written. "Excited to announce." Rocket emoji. Three hashtags you would never use. You try again with "make it punchier" and get the same post with shorter sentences. You close the tab and write it yourself, again, while a coworker tells you AI writes all their content now.

Here is the diagnosis before anything else: the tool is doing exactly what you asked it to do. You asked for a LinkedIn post about a feature. It gave you the average of every LinkedIn post about every feature it has ever seen. That is not a malfunction. That is the correct output for that input.

Generic output is a symptom of generic input. There are five specific failure modes that produce it, and once you can spot which one is happening, the fix takes minutes. This section shows you all five, with the actual weak prompt, the actual weak output, and the corrected version side by side. By the end you will be able to look at any bland AI draft and name the exact failure in under a minute.

Failure Mode 1: No Audience Context

This is the most common one. The model has no idea who is reading the thing you asked for, so it writes for everyone, which means it writes for no one.

Weak prompt:

Write an email announcing our new reporting dashboard.

Weak output:

Subject: Introducing Our Exciting New Reporting Dashboard!

We're thrilled to announce the launch of our brand-new reporting dashboard! This powerful new feature gives you deeper insights into your data, helping you make smarter decisions faster. With an intuitive interface and robust analytics, you'll wonder how you ever lived without it. Log in today to explore!

Nothing in that email is wrong, and nothing in it would make anyone click. "Deeper insights," "smarter decisions," "intuitive interface." These are the words that appear when the model has to guess who cares and why.

Strong prompt (audience layer added):

Write an email announcing our new reporting dashboard.

Audience: operations managers at mid-size ecommerce companies (200 to 1,000 SKUs) who currently build weekly reports by exporting CSVs into spreadsheets. Their main frustration is that pulling a single cross-channel report takes them 2 to 3 hours every Monday. They are skeptical of "analytics" promises because they have been burned by tools that required a data team to set up.

Strong output:

Subject: Your Monday mornings, minus the spreadsheet

You know the routine. Export from Shopify. Export from Meta. Export from Google. Paste, clean, pivot, repeat. Three hours later you have a report that is already out of date.

The new reporting dashboard pulls all of it automatically. Your Monday report is already built when you open your laptop, and it takes about 10 minutes to set up. No data team, no SQL, no implementation project.

Same tool. Same feature. The difference is that the second version knows whose Monday it is talking about.

Action step: Take the last AI output you rejected as "too generic." Rewrite the prompt with one addition: a two-sentence audience description covering who they are and what frustrates them about the status quo. Regenerate. Compare.

Failure Mode 2: No Voice Samples

Most marketers tell the AI their brand voice with adjectives: "professional but friendly," "bold," "conversational." The problem is those adjectives mean nothing. "Professional but friendly" describes roughly every B2B brand in existence. The model cannot hear your voice from a label. It can only hear it from examples.

Weak prompt:

Write a tweet about our spring sale. Brand voice: playful, irreverent, smart.

Weak output:

Spring has sprung and so have our DEALS! Get up to 40% off sitewide. Your cart will thank you. #SpringSale

That is every brand's idea of "playful." It is nobody's actual voice.

Strong prompt (voice samples added):

Write a tweet about our spring sale. Match the voice in these three examples from our account:

1. "Our accountant said we can't keep discounting like this. Anyway, 30% off through Sunday."
2. "Things that pair well with our candles: rain, your ex's apology text going unread, a Tuesday."
3. "New customers ask if our stuff really lasts. Our returns rate is under 2%. You do the math."

Strong output:

Spring cleaning tip: throw out your old stuff, buy our stuff instead. 40% off, and yes, our accountant has seen this tweet.

Now it sounds like the brand, because the model had something to imitate. Voice is pattern, and patterns need samples.

Pro tip: Keep a running doc of 10 to 15 of your best-performing posts, emails, and headlines. This becomes your voice sample bank. You will paste from it constantly, and in Section 1.3 you will formalize it into your Context Pack.

Failure Mode 3: No Constraints

When you do not specify format, length, structure, or what to exclude, the model fills the gaps with its defaults. Its defaults are listicles, three-part structures, and a word count that drifts long. Constraints are not restrictions on quality. They are how you get output that fits the actual job.

Weak prompt:

Write Google ad headlines for our project management software.

Weak output:

1. Transform Your Team's Productivity Today
2. The Ultimate Project Management Solution
3. Streamline Workflows Like Never Before

All three are over Google's 30-character headline limit, and all three are interchangeable with any competitor's ad.

Strong prompt (constraints added):

Write 10 Google ad headlines for our project management software.

Constraints:

• Each headline must be 30 characters or fewer. Count carefully.
• No exclamation points.
• Do not use the words "ultimate," "transform," "streamline," or "solution."
• 5 headlines must include a specific number (e.g., setup time, task volume, a time saving).
• 3 headlines must name the audience (agencies, client services teams).
• Our differentiator: clients can see project status without needing a login.

Strong output (sample):

1. Client Portal, No Login Needed (30)
2. Set Up Your First Project in 5 Min (30)
3. Built for Agencies, Not Enterprises (30)
4. Clients See Status. You Keep Working (30)

Constraints force specificity. Specificity is what makes output usable.

What you omit	What the AI defaults to
Length	30 to 50 percent longer than you need
Banned words	"Discover," "elevate," "unlock," exclamation points
Structure	Intro, three points, conclusion
What to avoid	Every cliché in its category

Warning: The model will not tell you it ignored a constraint. It will happily return a 34-character headline when you asked for 30. Build verification into your process: for character-limited work, paste outputs into a counter before you use them. The workflows in Part 2 include these checks by default.

Failure Mode 4: One-Shot Prompting

This is the process failure underneath the others. You ask for the finished asset in one prompt, get a mediocre draft, and conclude the tool is mediocre. But no good marketer works that way. You do not sit down and produce a finished campaign in one pass. You research, then position, then draft, then edit. AI performs the same way: much better in a sequence than in a single leap.

The one-shot approach:

Write a campaign for our new meal kit aimed at busy parents.

You get a soup of headline ideas, body copy, and taglines, all surface-level, all based on guesses about what busy parents want.

The sequenced approach (how the workflows in this book run):

1. Prompt 1, insight extraction: "Here are 12 real reviews from our customers [paste]. Pull out the exact phrases parents use to describe (a) the problem before they found us, (b) the moment it clicked, (c) what they tell friends about it."
2. Prompt 2, angle generation: "Using these customer phrases [paste], generate 8 campaign angles. For each: the core tension it exploits, the emotional promise, and a one-line hypothesis for why it would convert."
3. Prompt 3, angle selection pressure test: "Here are my 8 angles. Act as a skeptical CMO. Rank them by differentiation and tell me which two are most likely to be ignored as generic."
4. Prompt 4, asset drafting: "Write the campaign landing page hero section for angle #2. Voice samples: [paste]. Constraints: headline under 10 words, subhead under 20, no exclamation points, must include the phrase customers actually used: [phrase]."
5. Prompt 5, iteration: "Here is the draft. Diagnose: which sentence sounds most like a competitor wrote it? Rewrite only that sentence three ways."

Each step feeds the next. By step 4, the model is writing from real customer language, a pressure-tested angle, your actual voice, and hard constraints. The output is unrecognizable compared to the one-shot version, and the total time is about 20 minutes.

Example: A solo founder selling ergonomic desk chairs ran both approaches for a product launch email. One-shot version: 450 words of "invest in your comfort" filler. Sequenced version: opened with a phrase pulled from an actual review ("my 3pm backache just stopped showing up"), converted at roughly double her usual launch email rate. Same tool. Different process.

Failure Mode 5: No Examples of Good

The last failure mode is the least intuitive. Even with audience, voice, and constraints, the model does not know what "good" looks like for your specific task unless you show it. You would never brief a freelancer without a reference. The same applies here.

Weak prompt:

Write a meta description for our pricing page.

Weak output:

Discover our flexible pricing plans designed to fit businesses of all sizes. Find the perfect plan for your needs and start your free trial today!

Strong prompt (example added):

Write a meta description for our pricing page.

Here is a meta description from our best-converting page as a reference for structure and tone: "See exactly what [Product] costs. No sales calls, no hidden fees. Plans from $29/mo, cancel anytime."

Match that structure: direct cost statement, objection handled, concrete number, low-commitment close. Under 155 characters.

Strong output:

Pricing that fits on one line: $49/mo, every feature included, cancel in two clicks. No demo required to see the numbers.

One example did more than three adjectives ever could. This is why every workflow in this book includes a "paste an example of good" slot in the prompt template.

The Five Failures, Diagnosed in Your Own Work

Here is the diagnostic table. When an AI output disappoints you, run it against this list before you touch the prompt again.

Symptom in the output	Likely failure
Could be for any company in your category	No audience context
Sounds nothing like your brand	No voice samples
Wrong length, wrong format, full of clichés	No constraints
Shallow, scattered, tries to do everything at once	One-shot prompting
Technically fine but structurally wrong for the channel	No examples of good

Action step: Pull up your last three disappointing AI outputs. For each one, name the failure mode from the table. Most outputs fail on two or three at once, and "no audience context" is almost always one of them.

Why This Keeps Happening

Two reasons, and both are fixable.

First, the blank text box invites blank thinking. The interface asks "what do you want?" and you answer with the task ("write an email") instead of the brief ("write an email to lapsed customers who left because of price, offering a win-back discount, in our dry self-aware voice, under 120 words, like this example"). The gap between task and brief is where generic output lives.

Second, most advice about AI prompting is written for demonstrations, not for work. "Act as a world-class copywriter" is a demo trick. It produces plausible-sounding copy in a demo and unusable copy in a real campaign, because "world-class copywriter" still contains zero information about your audience, your offer, or your voice. Persona prompts are not a substitute for context. Context is the substitute for context.

The Fix, in One Sentence

Every prompt you write for a real marketing task needs four things before you ask for a single word of output: who it is for, what you are offering, how you sound, and what good looks like. That is the Four-Layer Input Template, and it sits inside a larger method: Context Pack, Task Spec, Constraints, Examples. That is the Context-First Prompting Method, and it is the operating system for every workflow in this book.

Here is what to do right now, before the next section:

• Write down your primary audience in two sentences: who they are, and what frustrates them about the current options.
• Collect 3 to 5 samples of your best on-brand writing into one doc.
• List 5 words or phrases your brand would never use (your banned list).
• Find one piece of your own marketing you consider genuinely good, to use as a "reference of good."
• Save all of the above in a single document. That is the seed of your Context Pack.

In Section 1.3 you will turn that document into a reusable asset you paste at the start of every session, so you never brief from scratch again. First, though, you need your tools set up properly, because configuration choices affect how much context your tool can hold and how well it remembers your voice. That is the next section.

The short version of everything above: the AI you are using is a mirror. Show it a vague, context-free request and it reflects the internet's average back at you. Show it your audience, your voice, your constraints, and your best work, and it starts reflecting those instead. The tool was never the problem. The brief was. Now you know how to write one.

The sample ends here.