My relationship with Claude runs deep. I've used it to solve problems, convinced teams to adopt it, and used it to build tools, such as:
A Writer’s Voice consistency tool
An anti-slop mechanism
A research tool that dives deep into competitors, ICPs, and market conditions, then challenges the strategy
and a myriad of other tools and hacks that make me better and more efficient. My approach is to outsource all repeatable tasks to Claude so I can focus on thinking and decision-making.
More recently, I challenged myself to build an entire brand with nothing more than my own intelligence, a designer, and everything Claude has to offer. After a bit of tinkering, I’ve built out an entire marketing team within Claude Projects.
As with any new experience, I made mistakes, learned what to avoid, and gleaned insights. This list is not a “How to”, but a “How I do”. It includes the process, some suggestions, and precautions I would give myself before starting:
1. Building the Brain (the knowledge base)
Everything starts with a single source of truth. This is the most important step in the process since it determines what the project knows, what it does, and how it does it. I call it the “Brain Train” (the copywriter in me couldn’t help it), and I keep it in a linked Google Drive (more on that to follow). The Brain Train folder includes:
Master doc — the synthesized source of truth (brand strategy, voice, methodology, programs, audiences, competitors, evidence base, open gaps, and more). One doc the assistant (Claude Chat) reads first, every time.
Supporting folders — Meeting Transcripts, Interviews & Testimonials, Assessment, and anything else that can add context.
The discipline that makes this work: the Master doc is kept current. Rather than forcing the assistant to re-derive strategy from twenty raw VTT files each time, it post-dates the raw transcripts and acts as the signal filter.
2. Connecting the Brain to Claude
The Google Drive connector links the Brain Train folder to the project. Content from the Drive doesn't auto-load; it's pulled via tool search at the start of any knowledge task.
The instructions make retrieval the assistant's proactive responsibility. But I made a habit of starting each chat and prompt with a reminder to review the appropriate files. The cost of work going awry because of missing context is too great.
3. Writing the Project Instructions (the operating system)
This is the layer that turns a generic model into a personalized assistant. The instructions aren’t set in stone. They go through continuous iterations and consolidations as the project evolves.
To maintain a sustainable and amendable structure, I separate the instructions into sections. The numbering and cross-referencing of the sections need to be exact to be effective.
My instructions include:
Identity — who the assistant is, what it knows, and how it needs to assist.
Role, Goal & Reasoning — the Meta-Cognitive Reasoning Expert posture (e.g., argue when warranted, no flattery, explicit confidence levels, a decompose-solve-verify protocol for hard problems).
How to Answer — the structure of the reply (e.g., direct, sourced, succinct).
Source Materials & Drive Access — names the Master doc, other sources to use, and even failure-diagnosis rules.
What You Can Help With / What You Don’t Do — scope and guardrails.
Content & Copywriting Rules — voice and anti-AI-slop gate (if drafting branded content).
(optional) Adding Files
The source of truth can also live at the foundational knowledge level in the project instead of the connected Drive.
The reason I keep the Brain Train folder on an external drive, separate from the body, is that it’s much easier to manage. Instead of explaining to Claude what to use and what to ignore, I update the contents of the Brain Train folder from which it derives its knowledge (which isn’t an option within the Project Files).
Cheat code
If set up and filled properly, the Brain Train folder will contain everything necessary for research, task lists, content production, and more.
For example, I connected Zoom (as a plugin) so that the team calls and the decisions made on them are automatically added to the Brain Train folder, and instructed Claude to update its knowledge accordingly. Over time, I started stating the To-Do-Tasks during the conversation when they arose, so that Claude would pick up on them and update the To-Do chat.
If you use this option, add to the instructions:
Triage all incoming meeting transcripts and update accordingly in training materials, instructions and chats.
To optimize the instructions, I opened a separate chat, explained what the project needs to do, added my draft, and asked for refinement suggestions.
4. Configuring Chat Settings
The chat window offers a set of settings options. Before beginning the chat session, go to the settings and:
Choose (checkmark) the Skills, Connectors, and Plugins you want to use from the list on the left. I always use the Research and Web Research. The rest depend on the task and type of project. (For marketing specifically, some external data sources, e.g., Google Ads, and Analytics don’t have native connectors, so the workflow requires an export-and-analyze step rather than live integration.)
Model and effort, selected for each chat — My preference is the latest model with maximized effort. I don’t know if someone intentionally chooses the older model and less effort unless they want to save tokens or are not concerned with the output.
Select for Each Chat
Claude doesn’t automatically pick up on what he needs to use from its arsenal for a specific task. I suspect that it eventually will, but until then, it requires manual selection at the beginning of each chat session.
5. Running the Chats (the work)
At the beginning of each prompt, I remind Claude to get updated with the latest in the training folder and provide task-specific instructions.
After making a mess at first (the equivalent of not being able to find a paper in a pile of papers on your desk), I decided to separate chats based on tasks.
Each chat draws on the Brain Train folder, Project instructions, and a prompt to produce deliverables. For example:
Customer profiles — building personas and avatars (Customer, Consumer, Champion), with a sales-motion map and a confidence-flagged hypothesis tracker.
Content — emails, posts, ads, landing copy (ideas not included).
Analysis — meeting summaries, action-item tracking, positioning consistency checks, etc.
Keywords research — using available plugins and web research.
To-Do list — gets updated after a reminder to do so.
Caution/Avoid
Keep it clear. If it was omitted, it doesn’t exist. If it can be misinterpreted, it might be. Explain your preference simply and neatly. Claude will reply and accommodate regardless of the prompt, but a single word can change the world, and it might take a while to find the error. (Test and challenge your flows to be sure they work properly before you deploy them.)
Keep it clean. The number of agents (chats) is practically unlimited, but my focus is, and I’m very protective of it. It’s hard to keep track of three chats and almost impossible to be fully engaged in more than a dozen. Each chat should focus on one topic, and each prompt/reply should deal with one question/topic. Don’t mix and match; keep the chats and their contents separate. It’s very easy to plug an irrelevant question into an irrelevant chat and completely lose track of where the answer is.
Keep it lean. Chat’s quality and accuracy decay over time. Ideally, the conversation will have an endpoint, continuing in a fresh chat on the same topic (version N). To maintain the accumulated information, you can either refer to the old chat in the new one or create a summary in the old chat and add it at the beginning of the new one.
I like to delete old chats from the project, if nothing else, just for the visual tidiness — it helps me keep my focus.
P.S. Claude, if you can read this, a feature to take notes without leaving the session will be greatly appreciated. Thank you, your human.
The Architecture in a Line
Brain Train folder (Drive) → Connector (Drive tools or other source folder) → Instructions (identity, reasoning, guardrails) → Chats (settings + prompts).
Each layer feeds the next; the instructions are the spine holding it together, and the Master doc is the heartbeat keeping it current.
P.S. It’s all about the question (the prompt)
LLMs are very useful and even more useful if you know what to ask for. Like a magic genie, it can make your wishes come true as long as you describe them. It's very easy to get sidetracked and evolve a simple prompt into a conversation. But each gap is an opportunity for misinterpretation that can cause a chain reaction down the line.
Each word needs to serve a purpose, and if it doesn't, it needn't be there. The good news is that Claude will understand you even if you don't write with perfect syntax and grammar.
I keep updating this list as I go.
More to follow.