Your AI is not stupid.
It is badly briefed.
Most "AI hallucinations" are context bugs, not model bugs. Give a model a folder of contradictory files and it will guess, confidently. Give it a proper foundation and the same model writes in your voice, to your buyer, with your facts. Have a play below.
You can't arrange your way out of a messy context.
Shuffle these notes as much as you like. While the files contradict each other and half are out of date, the model has nothing solid to stand on. Structure has to come first.
Once it's structured, try breaking it. Drag any card out of the model and watch the answer fall over. That is what each piece is holding up, and why half a foundation is not a foundation.
Five context sets. That's the job.
The five groups the cards sort into are not a metaphor. They are what actually gets written, structured and loaded into your Claude Project, custom GPT or whatever your team already uses.
Brand
Who you are, in a form a model can apply rather than admire. Named colours, defined type, the logo rules nobody can find. It stops inventing a house style every time it writes.
Voice & tone
How you sound, with examples of good and bad output annotated so the model can tell the difference. Plus the banned-word list, so nothing ships saying "leverage best-in-class synergies".
Product
What you sell, what it costs today, what it will not do. Dated, so a 2017 feature never turns up in a 2026 proposal.
Audience
Who buys, what pressure they are under, what makes them reply. The difference between copy aimed at everyone and copy aimed at the person actually reading it.
Messaging & proof
The approved lines, the real numbers, the case studies you have permission to use. So claims come with evidence attached instead of being generated to sound good.
Guardrails
The rules about what it may never invent: statistics, customer names, prices, features, quotes. This is the one that stops a plausible fabrication reaching a buyer, and the one nobody writes themselves.
Same model. Same prompt. Different brief.
"In today's rapidly evolving digital landscape, our best-in-class platform empowers forward-thinking organisations to unlock transformative value across the entire ecosystem."
- Invents statistics that sound plausible
- Sounds like every competitor you have
- Gets your own pricing wrong
- Written for nobody in particular
- Needs rewriting before anyone sees it
"Finance teams lose the 20th of every month to reconciliation. Acme cut that to a morning. Here is what it cost them and what changed."
- Refuses to invent a number it cannot source
- Sounds like you, because the rules are written down
- Quotes current pricing, dated
- Written to the person who actually buys
- Usable on the first pass
One to two weeks. Then it's yours.
Gather and sort the mess
You send me what exists, however scruffy. Brand files, old decks, the pricing spreadsheet, the Slack thread where the tone got decided. I read all of it and find the contradictions.
Write what's missing
Most of it is not written down anywhere. The voice rules, the persona, the guardrails. That gets drafted, you correct it, and it becomes the source of truth.
Build and load it
Structured into the five context sets and loaded into your Claude Project, custom GPT or both. Set up so your team uses it without thinking about it.
Test it against real work
We run your actual tasks through it, find where it still drifts, and tighten. You get a short guide for the team and the files themselves, which are yours to keep.
AI setup questions
Want one of these for real?
Tell me what your team is using AI for and where it keeps going wrong. I will tell you what the context foundation would need to cover and what it would take. No deck required.