Average content is free now. That is why it is worthless.

Volume is not the win

For fifteen years the content marketing advice was the same: publish more, publish consistently, win by accumulation. That advice worked because writing was expensive. The barrier was never the idea, it was the cost of producing it. That cost went to zero for everyone at the same time, not just for you.

The consequence is uncomfortable and worth saying plainly: if the output can be reproduced without you, it is not an asset. It is inventory. A blog of forty articles that a model wrote from forty titles is not worth forty times one article. It is worth roughly nothing, because anyone with a subscription and a free afternoon can replicate it. And it now also competes with the search engine itself, which summarises the same generic answer without sending the click to anybody.

Carlos, running three to five clients in parallel, feels the obvious temptation: multiply deliverables because he finally can. Sofia, who has tried twelve tools and mastered none, feels the other one: keep hunting for the tool that will finally give her the edge. Both instincts fail for the same reason. The tool is identical for everyone. What is not identical for everyone is what you feed it.

The right question is not "how much content can I produce". It is this: what is in this piece that only I could have put there? If the answer is nothing, do not publish it. Not because it is unethical, but because it will not work.

What the model cannot know

Claude knows how people write about your category. It knows nothing about your business. That border is the entire map of this page. On the model's side sits form: structure, grammar, register, format conventions, the ability to synthesise a thousand pages without getting tired. On your side sits substance, and substance is four concrete things.

One: research with actual customers. What they said on the onboarding call, which question support keeps getting, the exact moment in the demo where the doubt shows on their face. Two: objections in the customer's words, not yours. Not "price is a barrier", but "I don't know how to justify this to my boss if I can't show him a number in the first month". The second version is copy. The first is a meeting summary. Three: your positioning, meaning what you actually compete against and what you decided not to be. Four: your data, the numbers that come only from your product and your operation.

None of those four appear by writing a better prompt. A better prompt fixes form. Substance has to be brought in, and brought in so that it does not depend on you remembering to paste it every single time.

Loading context without pasting it every time

Pasting your positioning document at the start of every chat works right up until the day you forget once, and then you publish something that sounds like anyone. The fix is not personal discipline, it is structure. There are three mechanisms and they do different jobs.

1. Projects for the stable material of a brand or a client: positioning, ICP, voice guide, pricing, what can be promised and what cannot. One project per client, not one project for everything. Carlos with four clients needs four contexts that do not contaminate each other.
2. Skills for knowledge that only applies when it applies. A Skill is a folder with a SKILL.md file that Claude loads on demand. It works by progressive disclosure: the description sits in context, the full file is read only when the task calls for it. Which means you can have ten Skills (brand voice, SEO checklist, newsletter format, sector legal rules) without paying the cost of keeping them all open all the time.
3. MCP for whatever changes on its own. The Model Context Protocol is the open standard that connects Claude to external tools and data: Drive, GitHub, Slack, databases. This is where you stop copying and pasting CRM exports and the context updates because the data updated. The spec is at modelcontextprotocol.io.

The practical rule: if you are going to repeat an instruction more than three times, it is not a prompt, it is a system, and it belongs in a Project or a Skill. Systems over hacks, here too.

Brand voice is a system, not an adjective

"Friendly but professional" is not a voice guide. It is a wish. You can ask any model to write friendly but professional and it will hand back exactly the same internet average it would hand anyone else. Voice only becomes operational when it is extracted from real examples and written down precisely enough that you could contradict it.

The process that works has four steps and is done once per brand.

1. Collect eight to fifteen pieces that unmistakably sound like the brand to someone inside it. Emails that worked, support replies, a couple of posts. Real text, not the landing page an agency wrote three years ago.
2. Ask Claude to extract patterns, not adjectives: average sentence length, whether it uses rhetorical questions, whether it opens with the problem or the conclusion, which words never appear, how it addresses the reader, how it delivers bad news.
3. Turn that into checkable rules. "Sentences under twenty words" can be verified. "Human tone" cannot. Add a banned list with the sector's filler phrases and three before-and-after examples.
4. Save it as a Skill and make it apply to every piece. That is where the drift stops. The usual failure is not that the model does not know your voice, it is that it knew it at message four and had lost it by message fourteen.

Research that is actually research

This is where the 1M-token window stops being a spec-sheet number and becomes a change of method. A million tokens is, in practice, roughly a month of a company's documents in a single session. Applied to marketing, it means you can hold all of a quarter's customer conversations at once, instead of reading them ten at a time and misremembering the rest.

That enables four jobs that used to be done by eye. First, the competitor teardown: their pages, their pricing, their message, and the question that matters, which is not "what do they say" but "what do they promise that we cannot, and what do they avoid mentioning". Second, review mining: hundreds of reviews in your category, grouped by objection and sorted by frequency, with the verbatim phrasing preserved. Those verbatim phrases are the raw material of copy, and they are exactly what gets destroyed when you ask for a summary.

Third, synthesising interview transcripts, looking for what repeats across people who have never met each other. When four customers describe the same problem with the same metaphor, that metaphor is your headline. Fourth is vision: Claude reads images, PDFs, charts and screenshots, so the industry report in PDF and the analytics dashboard in a screenshot go into the same analysis as everything else.

One warning about method. Always ask for verbatim quotes with their source. A summary without traceability is an opinion formatted as data, and in customer research that is worse than having nothing, because it sounds convincing.

Repurposing done honestly

The spam pattern is recognisable from a distance: one idea, the same shape, blasted to six platforms on the same day. The thread that is the article chopped into pieces. The LinkedIn post that is the thread with more line breaks. Nobody reads it twice because there is nothing to read twice.

The alternative is not publishing less, it is adapting properly. One piece of real thinking with research behind it can be rewritten for each platform respecting what that platform rewards: in one place the full argument, in another the concrete case with the number, in the newsletter the part that only makes sense to someone who already knows you. Change the unit of value, not just the format. Claude does that adaptation well and fast, as long as you give it the complete original piece and not a summary of the original piece.

And there is an honest limit: if the idea was not enough for one long piece, it is not enough for five short ones. Repurposing multiplies what is there. It does not create what is missing.

Editing is the real skill

Treat every first output as a draft. Not as a result to approve or reject, but as raw material that has not been through you yet. That reframe is the difference between someone who uses the model well and someone who publishes its output.

The tells of unedited generated text are fairly stable. Sentences that open by hedging ("it's important to note that", "in today's world"). Three-item structures where the third item adds nothing and only completes the rhythm. Symmetrical empty claims of the "it's not about X, it's about Y" variety. Closings that summarise what you just read. Adjectives where a figure belongs. And the clearest tell of all: paragraphs you can delete whole without the piece losing anything.

The minimum edit is short and always the same. Delete the first paragraph, it is almost always warm-up. Strip every hedge ("perhaps", "in a way", "can help to") and check whether the sentence is still true; if it stops being true, the problem was the claim, not the hedge. Replace every generality with an example of your own. And read it out loud: if you would not say it on a call, do not publish it.

The measurement problem

Producing faster is not a result. It is a change in production cost, and it only pays off if it turns into something visible from outside. The common error is measuring output (pieces published, words, posts per week), which is exactly the metric AI inflates effortlessly.

1. Take the baseline before you start. Without the number from the previous ninety days you will not be able to claim anything. Record traffic by source, conversion by page, and real production time per piece.
2. Instrument conversion, not visits. Which content precedes a signup or a sales conversation. A piece that brings a thousand visits and zero conversations is a cost with a good appearance.
3. Measure time honestly. Include editing time and verification time. A draft in three minutes that gets edited for an hour is not a draft in three minutes.
4. Leave some pieces untouched as a control. A share of the calendar done the old way. Without a control, any lift gets attributed to the new tool by default, and it is probably seasonality.

If after ninety days the only measurable improvement is that you produce more pieces, the system is not working. It is manufacturing work.

What not to automate

Two categories, and both are easy to apply.

Any claim about your product you have not verified. The model completes with whatever sounds plausible in your category, because that is what it was trained to do. Ask it for a product page and it will hand back integrations that sound reasonable, benefits that sound standard, and figures with the right shape. Publishing that unchecked is not a style problem, it is commercial risk, and in some sectors legal risk. Every figure, integration, guarantee, timeline, or competitor comparison goes through human verification. Always.

Anything a customer would be upset to learn was generated. This test is more useful than any abstract ethics debate. The reply to a complaint, the apology email, the message of condolence, the testimonial, the response to a review written by a person with a name. The value of those messages is that somebody sat down and wrote them. Automating them does not save time, it destroys the only thing they were doing.

Everything else, the structural repetitive work that eats sixty percent of a freelancer's day, is exactly where this pays off. Clarity over noise: this is not about using more AI, it is about knowing where it stops being useful.