Table of Contents

• Why you get generic text (and it is not the model’s fault)
• The flow: brief → research → outline → draft → edit → channel

  • 1. Brief
  • 2. Research
  • 3. Outline
  • 4. Draft
  • 5. Human edit (non-negotiable)
  • 6. Channel adaptation
• How to keep the average from showing
• Editorial calendar with AI (without turning it into a factory)
• SEO: AI does not replace search intent
• Client data and cost: two limits the flow does not solve
• Frequently asked questions

  • Can I publish the first draft if it is “almost fine”?
  • Does it work for ads and landing pages?
  • What if the team cannot write well, even when they edit?
  • Is a fixed project with my documents worth it?
  • How do I measure whether the flow works?

Why you get generic text (and it is not the model’s fault)

The model does not know your market. It was not on the call where the client said “we already tried that with the agency.” It does not know that in your country nobody searches that keyword in English, or that your offer does not include implementation, or which case you cannot tell by name.

“Write me a post about inbound for SMBs” produces the average: definitions, 7 tips, “in the digital age,” and a close about “taking the next step.” That is not a failure. You asked for an article on a topic, with no owner.

The text stops being generic when the prompt stops being a topic and becomes a brief: audience with role and country, a problem with a symptom, what you do not cover, banned phrases, one real example, and what the reader has to achieve (understand, compare, book a call). If you cannot fill those boxes, you are not missing a prompt. You are missing the brief.

The flow: brief → research → outline → draft → edit → channel

Do not ask the same chat to do everything in one shot. Each stage has a different output. If you mix them, it invents research, dresses opinions up as data, and hands you an “SEO article” that answers no search.

1. Brief

You write the brief, even if it is ten lines. The AI stress-tests it; it does not invent it. Fill the brackets before you send it:

Piece: [blog article / LinkedIn post / newsletter / short]
Audience: [role, industry, country or region]
Problem they already have: [concrete symptom, not a category]
What they already tried: [agency, tool, internally]
Promise of this piece: [what they will be able to do or decide at the end]
Our offer (if it applies): [what we sell and what we do not]
Constraints: do not invent statistics; do not mention [competitors];
do not use these phrases: [list]; do not talk about [topics out of scope].
Proof we can actually use: [anecdote, our own number, anonymized client quote].

Prompt so the brief does not stay soft:

Here is the brief. Do not write the piece.
Point out: (1) the real question this person is asking,
(2) what they would have to believe to take a call with us,
(3) five concrete objections from someone in LATAM,
(4) which sections are surplus if we want a single idea,
(5) what I am missing so this does not sound like a template.
Ask me only what blocks the next step.

If the objections could be copied onto any SaaS (“it’s expensive,” “I don’t have time”), the brief is still abstract. Sharpen it until they smell like your vertical: “the owner will not let go of the Ads account,” “legal will not approve case studies,” “the team is one and a half people.”

2. Research

AI organizes what you already have and maps questions. It is bad at hunting “the truth” if you do not attach a source or you let it cite from memory.

Prompt for intents, not loose keywords:

Topic: [X]. Audience: [role] in [Mexico / Colombia / Argentina / regional].
List 12 queries the way that person would type them into Google,
grouped by intent: understand, compare, implement, buy.
For each one: what they expect to see on page 1 and what would be an
insufficient answer. Do not invent volumes. Mark the 3 we can cover
with our own proof and the ones we should leave out.

Then paste call notes, sales comments, tickets. Ask for clusters, contradictions, and gaps. Do not ask for “the 10 trends of 2026” if you are not going to verify them. Numbers go the other way: you bring the data and you ask it to use it or to say it is not enough. A percentage with no source does not go into the draft.

3. Outline

The outline is the last cheap moment to correct course. A weak structure does not get saved in the edit.

With the brief and these notes [paste], build an outline for a [1,600–2,000 word article].
A single H1, already defined: "[title]".
H2s that advance the promise, not labels like "Introduction" or "Conclusion".
Under each H2: the idea in one sentence, the proof (data, example, counterexample)
and what the reader should be able to do when they finish that section.
No filler lists. If a section does not change a decision, cut it.
Return the outline, not the article.

Read it out loud. If you can swap two H2s and nobody would notice, one is surplus. If it does not answer the query in the first paragraphs, reorder before you write.

4. Draft

Now write: by section, with a ceiling and prohibitions.

Write the section "[H2]" only, in direct second-person English.
Length: [N] words. Tone: direct, concrete, no infomercial phrases.
Use this example of ours: [paste].
Banned: "in a world where", "it is important to note", "without a doubt",
"success is within reach", travel analogies and superhero analogies,
statistics that are not in the notes.
If you are missing a data point, leave [MISSING: what data] and continue.
Do not close with a moral.

A whole article in a single prompt invites filler. If one section comes out tight and the next one does not, rewrite that prompt: almost always you are missing an example or you have leftover abstraction.

5. Human edit (non-negotiable)

You are the editor, not the “human who approves.” If the outline was good, twenty minutes are enough:

• Does the first screen answer the search, or does it warm up the engines?
• Is there a sentence only you could have written? If not, the text belongs to nobody.
• Did a number or a “study” appear that you did not put there? Cut it or verify it.
• Does it sound like your offer or like an agency whitepaper?
• Can you cut 15% without losing a decision? Cut it.

Editing is not swapping “utilize” for “use.” It is giving the text an owner again. If you do not have half an hour to edit, you did not have time to publish. You had time to generate.

6. Channel adaptation

An article is not “distributed” by clipping the first paragraph. Each channel asks for a different unit.

Blog. Canonical piece: search, depth, internal links, one promise per URL.

LinkedIn. One idea, as an incident or a decision. No H2s and no “in this post we will see.” Close with a question a peer can answer with their own case, not with “what do you think?”

Newsletter. Why now and what to do this week. They already gave you permission: you can be more direct. One link, if it adds something. An index of links gets archived.

Short (Reels, Shorts, TikTok). A visible problem in 3 seconds, one mechanism, one example. The article is not the script. The script is a scene:

From this section [paste], write a 35–45 second script.
Structure: concrete problem / why the obvious thing fails / one actionable step.
No on-screen text that repeats what is being said. No "follow me for more" CTA.
Close with a line that only makes sense if you watched the whole video.

Bridge prompt, once you already have the edited article:

This is the final article [paste or attach].
Give me: (1) a LinkedIn post of 180–220 words with a single thesis,
(2) a 140-word newsletter with a single next step,
(3) three short hooks that do not repeat the post title.
Do not recycle the introduction. Each channel has to work on its own.

How to keep the average from showing

Three levers the model does not pull on its own.

Voice with samples, not with adjectives. “Warm and professional tone” means nothing. Paste 400 words of yours and ask: “imitate constraints (phrases I would never use, sentence length, irony), do not copy the topic.” If there is no sample, you write the first paragraph and let it continue.

Local proof. An example from Bogotá, Monterrey, or Rosario beats any metaphor. Prices in the reader’s currency. Objections from WhatsApp, from tax, from “the cousin who does the website.” If the text could come out the same in Madrid or Miami, it is still generic.

Banned list. Paste it every time: revolutionize, supercharge, ecosystem, holistic, journey, hack, scale your personal brand, the power of, don’t get left behind. When the draft sneaks them in, delete them and rewrite.

A dirty trick: have the model attack the draft.

You are a reader who already hired two agencies and it went badly.
Mark the sentences that would sound like a pitch, the promises with no proof,
and the paragraphs that any marketing blog could have written.
Do not rewrite yet. Just annotate.

That list is your edit map.

Editorial calendar with AI (without turning it into a factory)

It is useful for seeing the board, not for filling it. Once a month, paste 3 to 5 pillars, offers, published URLs, and sales questions. Ask for four weeks, one canonical piece per week and its derivatives. Each canonical attacks a different intent. If two outlines merge, they merge.

Do not ask for “30 post ideas”: it gives you 30 titles and no priority. Instead:

With this inventory [paste URLs + pillars + sales questions],
propose 6 pieces for the next 4 weeks.
For each one: search or conversation intent,
why now, how it differs from an existing piece,
and what proof of ours is missing before we write.
If proof is missing, do not schedule it: put it on a "research" list.

Review it on Monday in fifteen minutes. Without proof, it drops. What sales keeps hearing goes up. AI does not decide the priority: it makes it visible.

SEO: AI does not replace search intent

A model can stuff the keyword into your H1. Without evidence it does not know if that query is your buyer’s question, or if page 1 already solved the topic with a format you are not going to beat.

Before you write, by hand:

1. What the person wants to do, not which words the query has: understand, compare, implement, or hire.
2. Which format is already winning. If page 1 is calculators, a 2,000-word essay will not move them. If they are listicles, do not publish another unless you have a selection criterion the others do not have.
3. What you can say that the others cannot. A number of yours, a process, a market constraint. Without that, you are version 40: faster and weaker.

Let the model classify, not guess rankings:

Target query: "[query]".
These are titles and H2s of the first 5 results [paste].
Tell me what intent they cover, what question they leave open,
and whether our angle [paste] is complementary or redundant.
If it is redundant, propose a more specific query or recommend not writing.

Two rules against junk content: do not publish just because the model said there is volume; do not inflate 800 useful words to 2,000. If the site looks like an interchangeable batch, you do not “win SEO”: you train people not to trust you.

Client data and cost: two limits the flow does not solve

The flow breaks when, to “give it context,” you paste emails, a contract, or the CRM. It is not a shortcut: it is a security problem. Before you upload “everything about clients” into a chat, read AI security and privacy in the company. If the data identifies a client or you would not publish it, it does not go into the prompt.

The other limit is money. Brief and outline tolerate a cheap model; the long draft and research with sources ask for more care, not always a higher price. If you pay for volume you do not edit afterward, the problem is the flow. The breakdown is in real AI prices in 2026.

Frequently asked questions

Can I publish the first draft if it is “almost fine”?

No. “Almost” usually means the structure holds and the voice does not. If there is no time to edit, publish fewer pieces.

Does it work for ads and landing pages?

The same flow, with a harder brief: offer, objection, proof, next step. A generic landing loses more money than a generic post, because someone is paying for the click.

What if the team cannot write well, even when they edit?

AI is not the problem. Someone has to tell an argument from an ornament. Train or hire editing, not more generation.

Is a fixed project with my documents worth it?

Yes, for brief, outline, and adaptation, if they are yours (voice, canonical pieces, banned list) and not client files. Not for research with fresh data: it becomes a museum of your biases.

How do I measure whether the flow works?

Not by number of URLs. In 60 days: editing time (it should go down), conversations that mention an angle of yours, and pieces you reuse because they are still true.

---

AI saves you the dumbest part: going from notes to a text with a shape. It does not save you from deciding what a URL is worth or what you can prove. If the flow starts at “write me a post” and ends at publish, you have a full calendar and an empty brand. If it starts with a brief that stings a little and ends in an edit with your name on it, the model is paying you back in time.