For agency founders and operations leads
Four layers, the tools I run in each, and the order I build them in.
Why the tools only work in the right order
Video is added before launch
Free, and it opens as a web page. One work email, no sales call.
Keynotes, talks, live sessions. I say on stage what I put in the list, and nothing in it is a position I hold back for paying clients.
Three or more, and prompting is not your problem. Your client context has no owner, no current version and no visible conflicts. Every AI workflow inherits that, and every person patches it differently.
Someone who knows the account gets a great result. They know which details matter, they add everything the model is missing from memory, and the output works. Nobody writes down what they added.
Then the whole team starts using AI. One works from the pitch deck, one from the last call, one from an old brand PDF that is still on the server. Every workflow now begins with a different version of the client.
The output still looks polished. It stops being dependable. Contradictions disappear inside confident copy, quality depends on who is prompting, and review rounds grow because nobody can see which assumption shaped the result.
Sooner or later somebody asks it in a review, and nobody in the room can answer with confidence. The client exists differently in every deck, prompt, call note and person's head.
That is exactly what the list is ordered around. Layer one is the client context: who owns it, where it lives, which version is current. The tools above it only start paying off once that layer holds, which is why they come second in the list and not first.
No dramatic AI failure anywhere in this day. That is exactly why it stays expensive — every step looks reasonable on its own.
The account lead shares the latest deck and some call notes. It feels like enough, because everyone in the room fills the rest in from experience. The AI gets the documents. Your seniors silently add months of conversations, rejected ideas and political nuance.
She rewrites the prompt, deletes an old positioning line and adds the product detail nobody documented. Ten minutes later the output looks strong. The comfortable conclusion is that the prompt works. What actually happened is that one experienced person supplied the missing system from memory, and none of it became shared client context.
He uses the saved prompt and the brand PDF he can find. That file still describes the old audience. The output is polished, fast and subtly wrong. Nothing crashes, no warning appears. Same tool, same prompt, different source reality.
The client is not annoyed about AI. They are annoyed because the agency looks like it forgot a decision. And the decision did exist — just not in a form every person and every workflow could reach.
More instructions, more pasted examples, and the contradictory sources underneath stay untouched. Tomorrow the same pattern repeats on the next account.
Like technical debt, it builds quietly. You pay the interest in longer prompts, extra reviews, repeated briefings and senior people fixing work that was supposed to save them time.
The list does not clear that debt for you. It does something smaller and more useful: it shows you which of the four layers is charging you the most interest right now, so you fix that one first.
That is why I check the client context before I look at any tool.
Before AI can be dependable, four things have to be written down: what is confirmed, what contradicts, what is missing, and who decides. That layer is easy to skip, and skipping it is the default. Skip it, and the model has to guess the client from documents that disagree with each other.
Briefs, decks, notes and approved examples go into one diagnostic. No new documentation project.
Line by line: what is evidenced, what contradicts something else, and what nobody ever decided.
You end up with one explicit next step, and you know what it unblocks.
Foundation before automation. That is the whole rule I work by. Automation does not fix bad inputs — it spreads them faster.
Most tool lists are a pile of logos. I grouped mine by the job each layer does, and put them in the sequence I build them in.
Everything else reads from this layer. If there is no current version of the client that someone owns, every tool above it inherits the confusion.
Buying team accounts is not the same as having a system. Here I name the environments worth standardising on, and how I keep client boundaries intact when several people work in them.
This is where a prompt turns into something you can maintain. Small Skills with one job, defined inputs and a defined output, plus the MCP connections that let them read approved sources instead of pasted fragments.
Governance is the layer I find missing most often. Then nobody notices when a Skill keeps working from a client file that was replaced months ago. Review rules, a named owner, and a way to check output against the source instead of against a feeling.
The tools I run in each layer, the job each one does, and the point where I stop trusting it.
Why it is free: copying a tool list is easy. The install order, and naming who owns context, access, Skills and governance, is the work a list cannot do for you.
Mechanical engineering taught me one lesson I never got rid of: if the inputs and tolerances change every time, the output is not a system. It is luck.
I studied at RWTH Aachen and worked on Industry 4.0 at Bosch before agency operations became my day job. I found the same pattern there around AI, so I stopped treating prompting as the main problem and started engineering the layer around it. Today that is what I do full time: context, ownership, access, governance and tests.
If the input changes every time, the output is not a system.More about Christoph ↗
Yes. Your work email, and the list opens. I put no upsell inside the document and there is no card step.
Four layers: client context, approved AI environments and access, Skills and MCP connections, governance and tests. For each tool I name the job it does and the point where it stops being the right answer.
The list is the smallest part of it. I grouped and ordered it, because a tool in layer three is wasted money while layer one still has no owner.
No. The layers matter more than the vendors. Where a category has real alternatives, I name the trade-off so you can pick on your own constraints.
Usually not. A brand guide says what you intend. This list is about where that intent lives, who owns it, and which environments can reach it.
No. It helps anywhere more than one person or AI workflow has to work from the same client reality.
Not for the first two layers. Those are ownership and access decisions. From the Skills layer on, someone technical helps.
Your email address, and the chance to be useful to you before I am relevant to you. If you later want the foundation built for one real account, I have a paid offer for that. The list works without it.
Four layers, in the order I build them. Compare it to your setup and find the layer worth fixing first.
Free, straight to your inbox. One work email, no sales call.
P.S. Your team already uses AI. You are already paying for the missing layers, in review rounds and senior time. The tools are the easy part. The order is the part nobody sends you.