How to Build an AI Content Engine That Doesn’t Sound Like AI

Why it sounds like AI

Because you gave it a topic. A model asked to “write about email marketing” produces something close to the average of everything ever written about email marketing. The average is competent, structurally fine, and completely forgettable. That is not a model failure — it is precisely what you asked for.

The tells everyone now recognises follow from the same cause:

• Openings that describe the landscape before saying anything (“In today’s digital world…”).
• Balanced both-sides paragraphs that never commit to a recommendation.
• Lists of considerations rather than decisions.
• Adjectives where numbers belong: significant, powerful, robust, seamless.
• A conclusion that restates the article and adds nothing.

None of that is fixed by a cleverer prompt. It is fixed by having something to say and supplying it.

The model is not the author. It is the fastest writer you have ever hired, with no opinions and no memory of your business.

The brand voice file

One document, reused for every piece. This single file changes output quality more than every prompt trick combined, and most teams never write it because it feels like overhead rather than work.

What goes in it:

# Voice

Who we sound like: [a specific person or publication you admire]
Reading level: a smart client who is not in our industry.
Sentence length: mostly short. Vary it. No three-clause sentences.
Person: second person ("you"). First person for opinions ("I think").

# What we believe

- [Opinion 1 you would defend in a meeting]
- [Opinion 2 that some competitors would disagree with]
- [Opinion 3 about what most people in this industry get wrong]

# How we argue

- Claim, then evidence, then the limit of the claim.
- Name the trade-off. Never present something as free of downsides.
- If we cannot evidence it, we say so rather than implying it.

# Banned

Words: leverage (verb), unlock, seamless, robust, game-changer, hustle,
       "in today's fast-paced", "it's important to note", "delve"
Structures: - Opening paragraphs that describe the landscape
            - "There are many factors to consider"
            - Conclusions that restate the article
            - Fake questions to the reader we then answer ourselves

# Formatting

- Answer under each heading in the first two sentences.
- Tables for comparisons. Lists only for genuine lists.
- Indian English spelling. Rs. for INR pages, USD for US pages.
- Never bold entire sentences.

# Examples

Good: [paste 3 paragraphs you actually wrote and liked]
Bad:  [paste 2 paragraphs that sound like everyone else]

That last section does the heaviest lifting. Three paragraphs of your real writing teaches voice better than two pages of adjectives describing it. If you have never written anything you liked, that is the actual problem — and it is worth an hour with a voice recorder, transcribing how you explain your work to a client.

The source-of-truth rule

One rule, written into every content instruction, prevents the majority of serious problems:

Facts come from the sources I give you. If a fact is not in the sources, flag it rather than writing it.

This turns the model from an author into an editor of material you control. In practice it means each brief carries its own evidence pack:

• Your own data, exports and past reports.
• Named external sources you have actually read, with links.
• Your pricing and offer document, so it never invents a price.
• Transcripts of real client conversations — the single best source of authentic language.
• An explicit list of claims you are not allowed to make.

Where the model has no source, it should leave a marker like [NEEDS SOURCE: click-through improvement figure] rather than producing a plausible number. Plausible invented numbers are the most dangerous output in marketing, because they survive review — they look exactly like real ones.

The five-stage workflow

Sequential, and the first two are yours. If you skip them you get volume, not content.

1. Decide the point (human). One sentence: what does this piece argue, and who changes their mind? If you cannot write that sentence, do not commission the piece.
2. Assemble the evidence (human). Data, sources, client quotes, screenshots, your own experience written as rough notes. Fifteen minutes here saves an hour of editing.
3. Draft (AI). Brief plus evidence pack plus voice file. Ask for the structure first, approve it, then the prose. Reviewing an outline is far cheaper than reviewing 2,000 words built on the wrong shape.
4. Edit (human). Cut hard. Add the one thing only you know — the client story, the mistake you made, the number from your own account. This step is what makes it yours and it cannot be delegated.
5. Package (AI). Meta title and description, schema, internal links, social variants, image alt text. Mechanical, high-volume, perfect for automation.

Notice where AI sits: stages three and five. The thinking at either end stays human, and that is not a compromise — it is where all the differentiation lives.

Four review gates

Nothing publishes without passing all four. Make it a checklist someone physically ticks.

The usefulness gate is the one people quietly drop, and it is the one that separates content that generates enquiries from content that fills a calendar. If a reader finishes and has nothing to do, you published a topic, not an article.

The volume trap

Here is the mistake I see most often. A team gets an AI workflow running, output capacity jumps, and they immediately publish four times as much. Six months later organic performance is flat or worse and nobody understands why.

What actually happened:

• Average page quality dropped, because the review step could not scale with output.
• New pages began competing with existing pages for the same queries.
• Internal links spread thinner across more destinations.
• The pages that generate enquiries got no attention, because everyone was making new ones.

Google’s stated position is that it rewards helpful, reliable content regardless of how it was produced, and acts against content mass-produced mainly to manipulate rankings. The practical reading: AI assistance is fine, unreviewed volume is the risk.

Repurposing without the slop

Repurposing is where AI genuinely shines, provided you respect one rule: each format gets a different angle, not the same text reflowed.

• LinkedIn — take the single most contrarian point and defend it. Not a summary.
• Email — the practical action, one story, one link. Written to one person.
• Short-form video — the mistake people make, then the fix. Fifteen seconds.
• Carousel — the framework or table, one idea per slide.
• Reply-in-community — answer the actual question asked, link only if genuinely relevant.

Give the model the source article, the voice file, and the specific angle for that format. Asking it to “repurpose this for social” produces exactly what everyone else’s repurposing produces.

How to tell if it is working

Set up a cohort comparison from the start: everything published in the three months before you adopted the workflow versus everything after. Then compare the things that matter.

• Enquiries per published page. The number that ends every argument.
• Engagement rate and scroll depth versus the pre-AI cohort. If people leave faster, quality dropped regardless of volume.
• Search Console clicks per page at 30 and 90 days.
• Editing time per piece. Should fall over time as the voice file improves. If it rises, your workflow is degrading.

The measurement setup is covered properly in How to Track AI Marketing Automation So It Shows Up in Revenue, and if you want the content pipeline running automatically, the tooling is in Claude Code for Marketers.

Frequently asked questions