Most founders use Claude as a smarter Google. The problem is not the tool, it is the setup. Here is the four-layer build order that turns Claude from a clever stranger into a colleague who has read the brief.
Published on 14 May 2026
You know the routine. Open Claude, paste last week's newsletter draft, ask for a tidy-up. The output arrives polite, professional, and structured. It also reads like every other AI-generated post you have scrolled past this month. So you spend 20 minutes downloading context, then 15 minutes editing the result back into something that actually sounds like your business. Hit send, close the tab, and wonder why you bothered.
Tomorrow, the same dance. Another chat, another context dump, another round of edits that quietly eats whatever time the tool was supposed to save.
I started calling this the cold-start problem after I noticed it costing me roughly 20 minutes per session. Multiply that across a week's worth of chats and it adds up to something like a working day per month spent re-explaining your own business to a machine that should already know.
But the time cost is only the surface problem. Underneath, something more damaging is happening. Claude is returning the average answer, the output its training data thinks a business roughly like yours should produce. For marketing, that is exactly what you do not want. The entire point of marketing is to not sound like everyone else. Yet the tool defaults to the median voice every time you start from scratch.
Most founders reach the same conclusion: AI does not work for their kind of business. They are wrong, because the tool works fine when it has been properly briefed.
There is no shortage of content telling founders what Claude can do, or how to write better prompts. Dozens of guides, YouTube tutorials, and Substack posts cover prompting techniques, use cases, and clever tricks. What almost none of them teach is the build order: what to set up first, second, and third so that the value compounds rather than resets with every new conversation.
That gap matters because prompt engineering is a treadmill. You optimise one prompt, then need to optimise the next. The context window closes. You start again. Setup, by contrast, is a one-time investment that pays off in every session that follows. The difference is the same as between giving a freelancer verbal instructions every Monday morning versus handing them a proper brief on day one and letting them get on with it.
The fix takes one afternoon. Maybe 90 minutes if you are focused, maybe 3 hours if you want to be thorough. After that, every Claude chat inherits the context automatically. You stop being the AI's memory, because the system takes over that role.
What follows is the four-layer architecture I run in my own consultancy, in the order I built it. The methodology, the build sequence, and what good output looks like at each stage are all here. The specific artefacts (your brand voice file, your audience mapping, the full Bollocks Filter checklist, the positioning decisions in our Discovery Diagnostic) are not, because those depend on your business. This guide gets you to understanding the system. Applying it to your specific context is the next step, and you can work that out as you go.
Layer 1 - The Project. A container per business area, set up once and lasting as long as you use the tool.
Layer 2 - The context files. Brand voice, audience, style guide. Plain markdown documents loaded into the Project's knowledge base. This is where 80% of the value lives. About 60 to 90 minutes the first time, then iterates over months.
Layer 3 - Skills. Reusable playbooks for tasks you repeat weekly. Each Skill is a folder with a brief and some examples. Build a quality gate first, then an atomiser. About 15 minutes per Skill once Layer 2 is in place.
Layer 4 - Connected tools (MCP). Claude reading from and writing to your actual systems: task manager, drive, CRM, knowledge base. Optional, and not where you should start.
The most common mistake is starting at Layer 3 because Skills are the cool new feature. A founder builds a clever LinkedIn-post Skill on top of an empty Project, gets a marginal improvement, and concludes the technology is not ready. The technology is fine, but the context was missing.
Each layer reads everything below it. A Skill (Layer 3) reads the context files (Layer 2) inside the Project (Layer 1). Without Layer 2, the Skill has nothing to draw on. Without Layer 1, the Skill has nowhere to live. The sequence is not a suggestion; it is load-bearing.
A Project turns Claude from a stateless chat interface into a workspace. Everything inside it, every conversation, Skill, reference file, and prompt, operates within the same context. Each new chat already knows what you are working on, who you are, who you serve, and how you sound.
One Project per business area. Marketing gets its own, Sales gets another. You can add Operations and Customer Success once the first two have proved their worth. The temptation is to bundle everything into a single mega-Project. Resist it. Different business areas need different reference material, and a bloated Project muddies which context Claude should load and when. Smaller, focused Projects outperform a single large one.
To create one, open the Projects tab in Claude.ai or the desktop app and add a 3-sentence system instruction in plain English:
You are working with [name], who runs [business name], a [size + sector]. Speak in [register], use British English. Default to specific examples rather than abstract advice.
That takes about five minutes. The temptation is to write a comprehensive instruction document at this stage. Resist that too. The Project instruction is scene-setting. The detailed brief lives in Layer 2.
This is where the real value sits, and it is the layer most founders skip because writing a brand voice document does not feel like setting up a system. It feels like writing.
It is writing, and that is precisely the point. The quality of these files determines the quality of everything Claude produces from this point forward. By month 3, these 3 documents will be the most valuable artefact in your business after the people in it.
This file gives Claude enough context to sound like you rather than the averaged-out version of a founder in a similar sector. The trap is writing rules. Adjectives like "Friendly. Bold. Story-driven." produce TED Talk transcripts. What works is samples. You cannot articulate the rule for "my cadence alternates long sentences with very short ones". You can only demonstrate it.
What to produce: a brand-voice.md file with 4 short paragraphs. Who you are (a personality, not a positioning statement). Who you serve (one phrase from how they describe their own situation). How you sound (sample paragraphs, not adjectives). What you refuse to do (over-claiming, snark, hashtag spam, false urgency). Followed by 3 or 4 real pieces of writing you would put your name to, each labelled with where it came from: "newsletter intro, March 2026", "sales email that closed Q4".
The operating principle: more samples beat more rules. A file with 1 paragraph of description and 10 samples produces better output than a file with 5 paragraphs of rules and no samples.
Brand voice is how you sound. The audience file is who you are talking to. Without it, Claude defaults to a marketing-persona caricature: "Sarah, 42, CEO of a 30-person professional services firm, frustrated with manual processes." Sarah does not exist, you have never met her, and Claude ends up writing copy for a fiction. That copy does not land.
What works is real customer language. The exact phrases your clients use when they describe their problem over coffee:
"We're competing on price and can't break the hourly billing trap."
"All our leads ghost after we send the quote."
"My team can't articulate our value."
"I am the bottleneck, and I know it."
When Claude reads those, it starts in the customer's world rather than the consultant's. The copy that follows uses their framing, not yours.
What to produce: an audience.md file. One segment to start, not three. The most common buyer you work with. One paragraph each on who they are (a situation, not a demographic), what they are trying to do, what they are stuck on, what they are afraid of, and what they are not telling their team. All in their language. If you have recorded sales calls or real client emails, lift phrases verbatim. "We are being eaten alive by a competitor we don't respect" is gold. "Increasing competitive pressure in the market" is what an LLM would have written without the file.
This file is the surface artefact of a deeper exercise we run as the Discovery Diagnostic: 23 positioning decisions a founder-led business has to make to grow past its current ceiling. The audience file is one of those decisions written down, and while you can absolutely build it yourself, the diagnostic makes it sharper and faster.
This catches the AI tells that slip past even a well-written brand voice file. The patterns every model defaults to that you keep editing out by hand:
Write down the things that annoy you, or you will be editing them out for as long as you use the tool.
What to produce: a style-guide.md file with 3 sections. Spelling and punctuation (UK English, your stance on Oxford commas, which dashes you allow). Banned phrases and patterns (the specific words and sentence shapes you have deleted three times this month). Cadence rules (vary sentence length, single-word landings welcome, do not smooth everything to the median).
You do not need the comprehensive version on day one. Five rules is plenty. Add new ones as you catch yourself repeatedly making the same edit. By month 3, this file will be one of the most useful artefacts in the system.
Skills are the playbooks. Reusable instructions for tasks you do more than once a week. Each is a folder containing a brief and a few examples. When triggered, Claude reads the brief alongside your Project's context files, then produces the output. If you want the full deep dive on what Skills are and how to build them, the practical guide covers that ground in detail.
Keep Skills narrow. The temptation is to build one giant "Marketing Skill" that handles everything. A handful of sharp, focused Skills will beat one broad one every time. A narrow Skill has a clear input, a clear output, and a clear stop-condition. A broad Skill becomes a saved prompt with delusions of grandeur.
Here are the three I would build first, in this order.
Layer 2 tells the model what to avoid before it writes. This Skill catches whatever slips through after the fact. The Style Guide is the door Claude should not walk through. The Quality Gate is the bouncer at the end of the night, checking nothing snuck past.
You need both because the Style Guide alone never quite reaches the quality bar. With my style guide loaded, some AI tells still leak into a minority of drafts. The Skill that writes the copy gets you 80% there. The Skill that checks the copy catches the remaining 20%. Anyone who claims a model handles voice end-to-end without a review step is selling something.
My version is called the No Bollocks Filter: a 10-point checklist auditing any client-facing copy. The patterns at the top are the ones that persist regardless of briefing: em dashes as default punctuation, contrastive antithesis ("it's not about X, it's about Y") where nothing genuinely contrasts, staccato triplets ("Speed. Scale. Savings."), and empty transitions ("but here's the thing...") where nothing surprising follows. I ran the fourth draft of a recent newsletter through the filter on a hunch the day before it went out. It caught 51 em dashes. That is why the gate exists.
To build it: create a folder called quality-gate with a SKILL.md brief listing the patterns to flag, an example of each one, and 3 to 5 before/after rewrites showing the kind of edit you want. The brief tells Claude when to trigger and what to output: "Flag what fires, explain why, suggest a rewrite."
Every founder doing content marketing has the same rhythm. Produce one thing of substance per week (a newsletter, a video, a long blog post), then break it into distribution pieces: LinkedIn variations, an email summary, image prompts, sometimes a podcast script. The atomising is the part nobody enjoys. It is structural lift, not creative work, and it quietly consumes hours that could go toward the next piece of substance.
A well-built atomiser Skill takes the long-form piece as input, loads the brand voice and style guide automatically (because they are in the Project), and produces 3 or 4 shorter pieces in distinct registers. LinkedIn is denser than a newsletter. Image prompts use different vocabulary. Stating that explicitly prevents Claude from defaulting to a single tone across all outputs.
For me, atomising went from a half-day job to 90 minutes including taste-checks. The structural lift disappeared. The creative editing remained, which is where the value was anyway.
Less glamorous than the other two, but probably the biggest time saver I have built.
Founders accumulate informational debt. PDFs, meeting transcripts, exported Notion pages, articles forwarded by clients, screenshots of useful posts. None of it is urgent enough to read immediately, but all of it is mildly relevant. Unsorted, it is useless. You either spend an hour on Friday triaging the pile, or you let it sit there and lose the value.
A triage Skill picks up new files from a folder, summarises each one, decides where it belongs in your system, and files it. You read the summary, not the original. What to produce: a folder with a SKILL.md brief explaining the input (a folder of mixed content) and the output (a one-paragraph summary per file plus a routing decision: keep, file under topic X, archive, discard). Include a taxonomy of where things go: "Brand voice updates to the style guide. Customer language to the audience file. Competitor moves to market notes."
After 2 or 3 rounds, the Skill is reliable enough that you stop reading the originals. I gained at least 2 hours a week back.
My first Skill took a week. I overthought it, treated it like a software specification with options and edge cases. The most recent one took 15 minutes. Treat them like a standard operating procedure for a capable person. A Skill is a folder with a brief and some examples. The brief gets shorter the more you trust the model. Build something rough, run it on real work, and let the next 5 outputs tell you what needs adding.
This layer is optional, and I would not recommend starting here.
Once the first three layers are solid, MCP (the Model Context Protocol) turns Claude from a drafting partner into something that can act on your real data. You connect Claude to your actual tools (task manager, drive, CRM, a vector store of your past writing) and it reads from and writes to them inside the conversation. This is where the AI coworker concept stops being theoretical.
The reason it comes last is that value compounds on top of the previous layers. A Claude connected to your CRM but lacking a brand voice file produces outreach that does not sound like you. A Claude with your brand voice but no tool connections writes good copy that you paste manually into systems all day.
My own setup splits into four folders: Operate/ for live execution (session logs, invoices, the wrap-up protocol), polything-context/ for business reference (brand, ICP, services, frameworks), skills/ for the playbooks, and an obsidian-vault/ for personal thinking (journal, projects, decisions). All four are indexed by a semantic search layer (a ChromaDB-backed MCP I call Sovereign Brain) that gives Claude a single query interface across roughly 50,000 chunks. When I ask "what do we know about [client name]'s AI workflow needs?" Claude reads the specific notes rather than guessing.
That is the long-run version. You do not need any of it on day one. Run Layers 1 to 3 for a month. During that period, notice the moments you manually copy and paste between Claude and other tools. Those become your shortlist. Add one connection at a time. Pick the one that causes the highest friction first (usually the task manager or the drive folder where reference material lives). Five carefully chosen connections are easier to maintain than fifteen poorly chosen ones.
If you only ever build Layer 1 and Layer 2, you will get 80% of the value. A Project with a good brand voice file, an audience file, and a style guide produces drafts in your voice from any prompt, with no Skill required. Layers 3 and 4 stack the savings across hundreds of similar tasks. Layer 2 is where the real leverage sits.
The order is designed so that each layer reads everything below it. Skip a layer, and everything above it empties out. Build them in sequence, and each one amplifies the previous. This is why the guides that jump straight to "build a Skill" or "connect your tools" leave founders frustrated. The foundation was never laid.
Consider an accountancy founder whose junior associates keep discounting to win work because nobody in the firm can articulate why they are worth more than the practice down the road. A brand voice file captures the firm's positioning in writing for the first time. Junior staff use it without the founder needing to be in every pitch meeting. Within a month, the team stops discounting because they have the language to hold a price.
Or a retrofit founder whose marketing materials read like specification sheets when prospects need to hear stories about real homes getting warmer. The atomiser Skill takes a technical document (heat-pump performance data, U-value tables) and produces a story-driven LinkedIn post about a real customer in a real house. The brand voice file ensures the output avoids jargon without the founder having to remember to ask. That could save a day of writing every week.
A SaaS founder with a 25-person team, struggling because everyone explains the product differently and customers feel it. One audience file in a shared Project means the entire team starts using the same language. Onboarding emails, sales decks, support messages, and product updates begin to feel consistent. What used to be a painful, months-long alignment effort now happens by updating the documents everyone draws from.
Different businesses, different goals, same architecture. The setup compresses what would otherwise be an open-ended cultural alignment job into a writing job that takes one afternoon.
Week 1: the brand voice file will feel too short, and you will want to add more rules. Resist it. The file gets sharper through use, not pre-emptive perfection. By the fifth time Claude produces something almost right but not quite, you will know exactly what is missing. That is when to add the rule.
Week 2: you will notice you have stopped pasting your context into new chats. For a while, you will keep doing it out of habit (just in case). Eventually, you trust that the Project loads the files every time, and you reclaim another 15 minutes per session.
Week 3: you will be building your second and third Skills. They take a fraction of the time the first one did. You might start getting Claude to build Skills for you from a short brief and a few examples.
Week 4: you will catch yourself explaining the system to a colleague. Describing how to set it up will feel easier than describing what you used to do before it existed.
It does not replace the human in the loop. You are still the guardian of the brand, the taste-maker, the editor-in-chief. After building the setup, your first drafts will be 80% in your voice. The remaining 20% is still yours: sharpening a specific line, adding a story you remembered as you read, cutting a sentence that flattens the rhythm, restoring a single-word landing you almost let slide.
That work is not less work, but it is better work. Editing a strong draft is a different activity from staring at a blank page or fixing a generic one. You also become a better delegator the more you use the system, because writing clear context files forces you to articulate what you want.
If I were starting from scratch tomorrow morning, here is what I would do.
0-10 minutes. Open a Project. Name it after the area. Write the 3-sentence system instruction. Save it.
10-35 minutes. Write the brand voice file. Four short paragraphs: who you are, who you serve, how you sound, what you refuse to do. Do not write rules. Show samples. Paste in 3 pieces of writing you would put your name to. Save as brand-voice.md and drop it into the Project.
35-55 minutes. Write the audience file. Pick your most common buyer. Write a paragraph in their actual language, lifted from real conversations or real emails. If you catch yourself reaching for phrases like "frustrated with manual processes", stop and ask: what would they actually say? Save as audience.md. Add it to the Project.
55-75 minutes. Pick a repetitive task and turn it into a Skill. The smallest, most-repeated task you do. A folder, a short brief, a single example of a good version, and a trigger phrase. Keep it small enough to fit on one page.
75-90 minutes. Run it. Use the Skill to produce one piece of real output. Edit it. Notice what you had to change. Those edits are signals: they go into the brand voice file or the style guide as those files mature.
You now have a Project, two context files, and one Skill. Total elapsed time: about 90 minutes. The value of those four artefacts shows up the next time you sit down to write something. And every time after that.
Prompt engineering is popular, but this infrastructure is what compounds. Once the system works, you can share it with the team so that every new chat starts from the same foundation. Everyone produces work in the same voice, at the same quality bar, without the founder in the room. If you want to understand why Skills are becoming central to how founders work with AI, the pattern is the same: the setup is the product, not the prompt.
The founders who win the next 12 months will not be the ones with the cleverest prompts. They will be the ones who built the right infrastructure beneath them.
If you have read this far, you are either going to build the system yourself this week or you want someone to build it with you. The 42-day Cowork experiment gives you an honest look at what this system produces after a month of daily use. And if you have tried Layer 2 but felt stuck on the audience file, the answers are likely already in your files, you just need to surface them.
If you would rather map the four layers with someone who has built this for their own business, book a Discovery Diagnostic and we will work through it together.
P.S. - If this resonated but you're not ready to talk yet, the Polything newsletter has more like this. One email a week, founder-focused, no fluff. Subscribe here.
Explore more insights and strategies to elevate your marketing approach.
A hands-on guide to building Claude Skills that remember your brand, audience, and standards. Real examples from three months of daily use in a founder-led consultancy.
Claude Skills turn organisational knowledge into reusable AI modules. For founders of £1-7m businesses, this means multiplying productivity without scaling headcount. Here is why it matters.
42 days with Claude Cowork, £0 revenue, 9 roles. Measured real work created, not hype, and the lessons for solo founders.