The complaint every marketer has
Every marketer I speak to who has tried using AI for content has the same complaint. The output sounds like an article. It uses phrases nobody in the sector actually uses. It contradicts what the brand would actually say. It reads as safe, glossy and slightly foreign.
This is not because AI is a bad writer. It is because you have asked an intelligent stranger to write about a business they know nothing about.
If you hired a very capable copywriter on Monday morning, sat them at a desk, and said “write a LinkedIn post for our client’s brand,” you would not expect the result to sound like the brand. You would expect it to sound like a capable copywriter’s guess. That is exactly what a public AI produces, and for the same reason.
The gap is not the tool. The gap is context.
The mental shift from prompt engineering to context engineering
Two years ago, the loudest advice about getting good AI output was prompt engineering. Write better prompts. Add role-play. Ask the model to think step by step. Add examples. Iterate.
Some of that was useful. Most is now obsolete. Models got significantly better at handling loose, natural language. The clever prompt template that used to matter now barely moves the needle on output quality.
What actually moves the needle in 2026 is context engineering. Not “how do I phrase the request.” “How much of my specific reality does the model have when I make the request.” Those are different questions, and the second is where the leverage now lives.
An agency practising context engineering is not fussing over the wording of the prompt. It is asking what the model needs to know to produce the right answer. Client brand voice. Product catalogue. Compliance constraints. Decisions from the last six months. History of what worked. All of that as inputs.
Given that context, most prompts collapse to one sentence. “Draft this week’s LinkedIn post about the topic we agreed.” The context does the work.
Five categories of context most agencies do not load
In order of impact:
Brand voice. How the client’s brand writes. What words it uses and refuses to use. What length of sentence it favours. How direct or hedged it sounds. Under-loaded. Highest leverage. Turns generic output into recognisably-them output.
Product and service facts. What the client actually sells. What each item costs. Which product is aimed at which customer. Which claims may and may not be made about each. Without this, every product-specific piece contains a small factual error you must catch. With this, product-specific content becomes a copy-paste job.
Decision history. What the client has already decided about pricing, positioning, messaging, target markets. Without this, the AI happily contradicts a decision made in April. With this, the AI enforces the decision for you.
Regulatory posture. Which sector rules apply. Which words are landmines. Which disclosures are mandatory. Which claims trigger a legal review. For regulated clients this is the difference between drafts you ship and drafts you scrap.
Historical results. What the client has already tried in market. What worked and what did not. Without this, the AI happily suggests the exact campaign that failed six months ago. With this, its suggestions build on what you already know.
Give an AI these five categories of context and the same “draft this LinkedIn post” prompt produces categorically different output. Not marginally different. Recognisable. Factual. Non-contradicting. Ship-ready.
What this looks like practically
At GKnect Digital, every client has a dedicated context bundle. Not a folder of Word documents. A structured pack that the assistant loads the moment I open a conversation about that client.
For one financial services client the bundle contains: the FSP number and the exact risk warnings that must appear in every consumer-facing piece; a brand voice specification derived from about fifty pieces of their published content; the current product suite with each product’s key claims and mandatory disclosures; a running record of decisions from the last six months; the compliance officer’s flagged phrases to avoid.
The result: when I ask the assistant to draft a Facebook ad, it writes copy that sounds like the brand, features the correct product benefits, includes the mandatory risk warning, avoids the landmine list, and does not contradict a decision made in April. The draft still gets reviewed. It just starts in the right place instead of five drafts away.
For our own brand, the same principle. GKnect Digital has one brand voice pack. Gknect Managed IT has a different one. My personal writing has a third. Each surface loads its own. Consistency across a hundred pieces of content stops being an act of will and becomes a property of the system.
The three levels of context maturity
Most agencies sit somewhere on this ladder. Being honest about which rung you are on is the shortcut to the next one.
Level 1: cold prompting. Open a chatbot, ask a question, close the tab. No context loaded. Every session starts from zero. Output generic. Heavy editing required. Most agencies here.
Level 2: working prompts. You have a document with your favourite prompts, maybe some “role-play” instructions. Better than Level 1. Still fragile. Still relies on your discipline to paste the right stuff at the right time. Some agencies here.
Level 3: loaded context. Your assistant automatically has the client’s brand voice, product facts, decision history and regulatory posture the moment you engage it. You paste nothing. Context is a property of the environment. Very few agencies here. The ones that are produce shippable output on first draft.
Level 1 to Level 2 is a productivity gain. Level 2 to Level 3 is a category change. It converts AI from a helpful writing companion into a co-author who already knows the client.
Three failure patterns when agencies try to build this
Confusing custom instructions with loaded context. ChatGPT’s custom instructions field holds a few paragraphs. That is a slightly better prompt. Real context is measured in pages, different per client, not one global setting.
Writing the context from scratch. They sit down to draft “our client’s brand voice” as a Word document, get three sentences in, decide it feels forced, give up. The correct move is to extract voice from what the client has already written. Every established client has a corpus. Use it.
Treating context as a one-time setup. Context is a living record. Decisions change. Products retire. Regulatory positions update. Without maintenance, context drifts and eventually contradicts itself. Fifteen minutes a month adding what changed. Not heavy. Has to happen.
What to build this quarter
Three concrete pieces of work will get most agencies most of the way.
Extract brand voice from what each client has already written. Point an AI at fifty pieces of their published content and ask it to describe the voice patterns. Iterate the description with your editorial team until it reads as “yes, that is them.” Save it. This is the voice pack.
Assemble a product-and-service brief per client. Every product, every service. Key facts, price ranges, target customer, permitted claims, mandatory disclosures. Factual, not marketing.
Maintain a decisions log per client. A single running document. Every meaningful decision gets a one-line entry with a date. What we decided about the entry-level product in August. Two minutes per decision. Six months from now, this is the most valuable single artefact for that account.
Load all three into every AI session you run about that client. Not manually per session. Once, as the default context for anything about that account.
Do that, and the “AI content sounds generic” complaint stops. Not because you got cleverer at prompts. Because you stopped asking a stranger to write about someone else’s business.
Where this ends
Prompt engineering was a phase. Context engineering is the craft. The agencies that treat their client operating context as an asset (captured, maintained, loaded on demand) will produce AI-augmented work no competitor with the same public tools can match. The agencies that keep cold-prompting will keep getting generic outputs and concluding AI is not for them.
The tools are the same for everyone. The context is not. That is where the moat now sits.
Want brand voice extracted from your client's own writing?
If this landed, the practical next step is the extraction. GKnect Digital runs a brand voice audit as a productised service. We point our stack at fifty pieces of your (or your client’s) existing content and produce a voice pack you can load into any AI workflow. Reach out via gknectdigital.co.za.
Part 4 of a 4-part series on how Marcos Diez uses AI to run GKnect.
