Write it down once: the context files that stop generic output

The most common complaint about AI-written work is not that it is wrong. It is that it is generic — technically fine, and recognisably not yours. It could have been written about any company in your sector.

That is not a model limitation. It is a missing input. The model does not know what you sell, who you sell it to, what you are allowed to claim, or how you sound, because nobody ever wrote it down.

Four documents, an afternoon

Not a knowledge base. Four short markdown files that every prompt and every skill reads from. Each one is a page, maybe two.

1. What this business actually is. What you sell and to whom, in the words a customer would recognise. Pricing shape. Where you operate and in which languages. Who the real alternatives are, including "they do nothing". And the part everyone omits: what you do not sell, and what you cannot deliver. Capacity limits and exclusions stop the model writing cheerful promises you cannot honour.

2. Who this is for, and who it is not. Derive it from your last twenty won and lost deals, not from aspiration. Then write the disqualifiers before the qualifiers — they are more useful, less contested, and much faster to agree in a room. "Who is this not for" is a question a management team can answer in ten minutes; "who is our ideal customer" can absorb a quarter.

3. What we are allowed to say. Every claim you currently make, each marked evidenced, partly evidenced, or unsupported. Delete the third category from your marketing while you are there.

Then the crucial line: any figure not in this file may not be used. If a draft needs one, it writes a placeholder and stops. This one rule removes most invented statistics from customer-facing work, because the model no longer has to choose between a plausible number and an awkward gap.

4. How we sound. Not adjectives — examples. Take five pieces of real work you were happy with and five you were not, and say briefly why. Add maximum lengths, and the phrases you personally cannot stand. The contrast between the good and weak examples does more than any description of tone.

Where the Belgian bit goes

Two things belong in file 1 and nowhere else, because if they live in someone's head they will be forgotten at volume:

Which language each audience gets. Employee-facing documents follow the language of the operating unit — that is a legal question, not a style one, and a document in the wrong language can be treated as never having existed. Write the rule into the file so every draft inherits it.

Whether the output is customer-facing. In Belgium the practical distinction that matters is between an internal draft and something that leaves the building with your name on it. Say which is which, once.

Why this is worth an afternoon

Everything downstream inherits it. A prompt improved by 10% helps one task; a context file improved by 10% helps every task, forever, including the ones the person who wrote the file never imagined.

It is also the difference between one colleague being good at AI and the team being good at it. Their skill is in their chat history. The files are shared.

Keeping them honest

Four rules, each of which exists because of a specific failure:

• A date and an owner at the top of every file. Without an owner it goes stale; without a date nobody knows whether it has.
• Nothing aspirational. These describe what is true today. A file describing the company you intend to be produces drafts about a company that does not exist.
• Someone other than the author reads all four. Authors cannot see their own assumptions — this is the single cheapest quality step available.
• A review date in a calendar, not in someone's memory. Quarterly is enough for most; monthly while you are still calibrating.

Start with one

If four feels like a project, write what we are allowed to say first. It is the one that prevents the most damage, it is the easiest to get agreement on, and it pays back the same week.