Episode 8 — Templates, Flows, and the Claude Code Question

## The Starting-From-Zero Trap

[A] "You ever spend, like, three hours just whispering sweet nothings to an AI? Just to get the absolute perfect tone for a client email — you feed it examples, you tweak the adjectives, you massage the prompt. And finally, it spits out a masterpiece."

[B] "Yes. You feel like an absolute prompt engineering genius."

[A] "But then, the very next morning, you open a fresh tab, ask it to write a simple follow-up, and it spits out the most corporate, robotic garbage imaginable."

[B] "Ah, the classic 'dear sir or madam' email. You are instantly back at square one. Well, that amnesia is essentially the defining frustration of how we currently work with language models. We treat them like search engines — we open a window, we get what we need, and we close it. Just a one-and-done kind of thing. But by doing that, we are interacting with these incredibly powerful adaptive systems as if they have zero long-term memory. We are actively choosing to experience amnesia by design."

[A] "Anyway, welcome to another Deep Dive. Today, we are looking at a really fascinating source document that outlines an advanced, complete AI system tailored specifically for marketers and knowledge workers — essentially a blueprint for building an AI workspace that actually remembers what you did yesterday. Our mission is to extract the practical tips, the daily workflows, and the underlying philosophy from this text. We want to ensure your best work actually compounds over time. Because if you are starting every single project from a blank chat window, you are definitely doing it wrong."

[B] "Okay, let's unpack this. What's the core issue here?"

[A] "The core complication identified in our source material is what we might call the starting-from-zero trap. Think about the reality of a modern knowledge worker's workflow. You might have your AI set up perfectly with custom instructions. It generally knows who your clients are. But when a brand new client brief lands on your desk, you still produce that first draft totally from scratch — every single time. You just stare at the blank white box, the cursor is blinking, and you have to re-explain the entire context of the universe to the machine."

[B] "And the real tragedy there is that you have already solved this problem. Maybe last month you wrote a campaign brief that landed perfectly. Or an onboarding deck that the client loved so much they didn't change a single comma. But that high quality, that hard-won victory — it lives entirely in the final deliverable. It lives in the PDF on your desktop, or in a slide deck in the cloud. It doesn't actually live in the system itself."

[A] "Exactly. Because of that isolation, your specific knowledge, your expertise, your unique creative fingerprint — none of it is compounding within your AI tools. It is like hiring a wildly fast, incredibly talented assistant who unfortunately suffers from 100% memory loss every time they go to sleep. Every morning they walk in, ready to work, but they have absolutely no recollection of that massive stylistic win you achieved together yesterday afternoon."

## Extracting Templates from Your Best Work

[B] "So the multi-million dollar question the text poses is this: how do we guarantee that our next brief starts from our absolute highest standard, rather than starting from zero? And the answer comes down to one strategy."

[A] "Which is it?"

[B] "Extracting templates. The fundamental insight here is that every piece of genuinely strong deliverable work is actually just a strong structure. If you can extract that underlying structure once, you never have to pour that foundation again. The document lays out an exact step-by-step workflow for this."

[A] "What's the first step?"

[B] "It tells you to take your absolute best piece of client work from the last month — the one you were most proud of — open Claude Cowork, paste the full text of that winning document into the chat, and then use a very specific prompt. I want to read this directly from the source: 'Extract the structure from this. Identify the sections, the flow, the formatting patterns, and the tone approach. Strip all client-specific content. Save what remains as a template to templates/[type].md.'"

[A] "That command is performing a digital extraction. You are taking a one-off deliverable, a snapshot of past success, and transmuting it into a permanent, reusable asset. But I noticed the prompt specifically asks to save it as a .md file. Why not just save as a regular Word document or a PDF?"

[B] "If you save something as a Word document, there is a massive amount of hidden, messy formatting code running under the hood — font sizes, margin data, meta tags. Humans don't see it. But an AI has to process all of it, which wastes its processing power. A markdown file, on the other hand, is basically raw text with very simple, universal symbols for things like bolding or headers. Super clean. It makes it infinitely easier for the AI to read and perfectly apply to future tasks."

## Templates Capture Shape, Not Words

[A] "But okay, I have to push back on this entire premise for a second. If I just template my best work and then force the AI to use those same templates over and over, won't my marketing start to sound completely robotic? I feel like I'd just be building a factory assembly line that churns out identical cookie-cutter content, entirely stripped of any real human creativity."

[B] "Yeah, that is the natural fear. But what's fascinating here is that this methodology produces the exact opposite result. The reason lies in what the template is actually capturing — it is not capturing the words you used. It is capturing the shape of your thinking."

[A] "The shape. Let's define that because that sounds a bit abstract."

[B] "When the system extracts the structure, it is analyzing the mechanics of your success. It looks at the section order and why that order works. It analyzes your sentence patterns — are you using three punchy short sentences to build tension, followed by a long flowing sentence for the resolution? It looks at what comes first: do you lead with a counterintuitive insight, a provocative question, or a hard data point? And it looks at what comes last — ending on a hard call to action or leaving an open loop. That invisible architecture — that is the shape."

[A] "It's like the architectural blueprint of a house. The blueprint doesn't dictate what color you paint the walls, what art you hang, or what furniture you put in the living room. It just ensures the house won't collapse. You're not going to end up with a door on the ceiling. By having that blueprint locked in, you actually have more mental freedom to be creative with the interior design because you aren't wasting hours trying to figure out how to pour the concrete foundation."

[B] "And the return on investment for building these blueprints is staggering. The source notes that just three months of maintaining this discipline — extracting the shape of your best work whenever you have a win — yields a library of about 15 to 20 highly refined templates. And according to the text, every single time you deploy one of those templates, it cuts exactly 40 minutes from your next brief. Three months from now, when a new client brief arrives, you don't stare at the blinking cursor. You just tell the system: 'Read templates/campaignbrief.md and aboutme/aboutme.md. Here's the raw info for the new brief. Write the first draft.' And the output is on format, on brand, and structurally sound on the very first try. You've effectively cured the amnesia."

## Three Advanced Session Patterns

[A] "But capturing those templates is only half of the system. Because even the best template will degrade if you don't manage the active working session properly. The text addresses this by breaking down three advanced session patterns that knowledge workers need to internalize."

[B] "Let's get into those."

[A] "Pattern one: conversation pruning. The rule here is strict — you have to start a new session every 20 messages. It's a hard limit. To do this, you copy the best output from your current session, open a brand new chat window, and paste it as the very first message with the prompt: 'Here's where we got to. Continue from here.' It's a hard reset."

[B] "But I genuinely struggle with this 20-message rule. When I'm deep in a brainstorming session with an AI, the ideas are flowing. Forcing myself to stop, copy everything, and paste it into a blank window — it feels like slamming the brakes when I am finally cruising."

[A] "I totally get that. But think about what is happening under the hood. AI models have what is called a context window — essentially their active short-term memory. Every single time you type a new message and hit enter, the AI does not just read your new sentence. It has to reread and reprocess the entire history of that conversation from the very first hello, all over again, to formulate its next response."

[B] "Wait, really? So by message 40, it's rereading thousands and thousands of words just to answer a simple question."

[A] "And two dangerous things happen when a conversation drags on like that. First, the AI's attention gets diluted — it starts to lose track of the precise template shape you gave it at the beginning because that instruction is buried 40 messages up. It starts to drift. Second, all that rereading wastes computational power — what the source refers to as token waste. I actually had this happen last month. I was doing a massive rebrand strategy for a boutique coffee roaster, maybe 80 messages deep throwing ideas back and forth. And I asked it to summarize our final tagline options, and it suddenly started suggesting taglines about organic tea. It completely hallucinated a new product line because the context window was just overflowing with random brainstorming garbage."

[B] "That is the exact failure point. By pruning the conversation at 20 messages, you wipe away the messy back-and-forth brainstorming, provide fresh context, and it's like clearing the cache. You are essentially saving your progress in a video game, resetting the console so it doesn't overheat, and loading right back into the exact spot you left off — running at full speed."

[A] "Which leads us to pattern two: parallel sessions. The absolute golden rule here is do not mix tasks in the same window. Research happens in one session. Writing happens in a totally different session. Because research sessions naturally become chaotic — you are throwing a million things at it, feeding the AI multiple articles, asking it to summarize conflicting viewpoints. If you then ask it to write a polished, templated campaign brief in that exact same chaotic window, it is going to pull in all sorts of weird, half-baked ideas from the research phase. You're cross-contaminating your own workspace."

[B] "You take the final, crystallized summary of the research from your first session, and you use that as the opening message in a brand new writing session. And that brings us to pattern three: save and reference. When a session produces something highly reusable — like a strong positioning statement or a foundational creative brief — you do not just leave it buried in the chat history. You save it directly to a dedicated Outputs folder. For example: OUTPUTS/clientname/positioning.md. So the next time you sit down to do work for that specific client, maybe a week later, your very first prompt is: 'Read OUTPUTS/client/positioning.md. Here's the new brief for today.'"

[A] "Conversation pruning, parallel sessions, and save and reference are all designed to enforce a single non-negotiable standard: no briefing from scratch, ever. You are always building on a foundation that has already been approved, polished, and saved. We stop treating AI like a disposable one-off writing tool and start treating it as a dynamic, evolving repository of our own best thinking."

## Claude Cowork vs Claude Code

[B] "But there is a crucial limitation to acknowledge. These session patterns require active, conscious human management. You are the one copying and pasting at message 20. You are the one keeping the writing and research windows separate. You are the one saving files to the Outputs folder. Which brings up a fascinating tension: what if someone wants a system to just run on a schedule? What if they want an AI to do the work while they are asleep, taking themselves out of the loop entirely?"

[A] "That requires a totally different approach. This leads directly to a major point of confusion — the text addresses the great tool divide between Claude Cowork and Claude Code. People are constantly asking: is Claude Code going to replace Claude Cowork? And the definitive answer is no."

[B] "No, absolutely not. But understanding the difference between the two is crucial to ensure you aren't fighting your own tools. Let's define the boundaries clearly. Claude Cowork is designed for people running a knowledge business. If your final outputs are documents — writing, strategy analysis, client management, content calendars, basically the entire marketing production workflow — Cowork is where you live. It relies on the template system and the chat interface we just spent all this time talking about. Everything assumes the end product is words on a page meant for a human to read."

[A] "Now contrast that entirely with Claude Code. Claude Code is a terminal-based coding agent. It does not have a friendly browser interface with a chat box. You install it via your computer's command line. Its job is to read local files, run system commands, execute complex tests, and build software. It is built for a professional whose final outputs are functional code, not strategic documents. Using Claude Code to write a marketing brief is like using a heavy-duty industrial forklift to meticulously arrange the antique furniture in your living room."

[B] "That is a terrifying visual. The machine is incredibly powerful, but you are applying it to completely the wrong job. You only pivot to Claude Code if you have built something technical — say, a Python script that automates a task on your desktop that you want to version control or deploy to a server. Or if you need a multi-step process to run on a schedule without your involvement. If I need a script to automatically scrape a competitor's pricing page every Tuesday at 3 a.m., cross-reference it against my own prices, and dump the flagged discrepancies into a database while I am sound asleep — that is a job for Claude Code."

[A] "The learning curve to use the terminal is steep. But if you need true hands-off automation, that's where you go. But if you are synthesizing research for a 10 a.m. client meeting or extracting the shape of your best copywriting to build a reusable template — Cowork is the right tool for the job every single time."

## Model Tier Mapping and Exercises

[B] "Okay, so we know how to capture the shape of our best work. We know how to manage our daily AI sessions so they don't degrade. And we know how to choose the correct tool. The source material actually ends with a set of deeply practical exercises."

[A] "Exercise one is straightforward application: extract three templates from your recent work using that specific extraction prompt we discussed. Test each of those templates with a new brief this week, and adjust the template once based on the results."

[B] "Exercise two is about automation through what the text calls a skill. And it's worth drawing a clean line here, because templates and skills are related but distinct. Templates capture the shape of your output — the structure of a great campaign brief, the flow of a winning pitch. Skills encode the process — the exact steps the AI follows to produce that output. You need both: the template tells the AI what to build, the skill tells it how to build it. With that distinction clear: you pick one task you do the exact same way every single week — say, a weekly performance report for your team. You write a specific skill for it: define the exact prompt, the inputs required like a messy CSV file of raw metrics, and the exact output format like a clean, bulleted executive summary. The goal is to run that specific skill three times to iron out the bugs. Once you do that, you are done inventing it forever. It transitions from a creative writing task into a simple button you push every Monday morning."

[A] "And exercise three is about mapping your clients to a model tier — designating which clients need Opus-level work and which clients run cleanly on Sonnet. But if Opus is Anthropic's most capable, genius-level model, why would I ever downgrade? Why wouldn't I just use the absolute smartest model for everything?"

[B] "Think of Opus like a massive, multi-story university library. To answer your prompt, the AI has an enormous reasoning engine that has to run through a million virtual aisles, cross-referencing deep complex concepts before it gives you an answer. If you have a highly complex, nuanced, strategic problem — you absolutely want that massive library. But if you just need the AI to take a messy list of 50 client names and alphabetize them, or format a standard weekly update, that massive reasoning engine actually gets in its own way. It takes longer. Sonnet, by contrast, is like a highly specialized desk clerk — fewer aisles to check, much faster execution for straightforward tasks. Mapping your clients to a model tier acknowledges that not every deliverable requires deep, exhaustive reasoning. By putting this rule in your system, you save time, ensure faster responses, and reserve the heavy lifting models for high-stakes work."

[A] "And the ultimate metric the source leaves us with: run one real client engagement through this entire compounding system this week — not a drill, a real project with real stakes. The only metric that matters is measuring the time it took using this system against how long it used to take you. Because if the templates are built correctly and the sessions are pruned properly, the time saved will be undeniable. And more importantly, the quality of the final deliverable will actually be higher, because you didn't start from zero — you started from your best previous baseline."

## The Competitor Blueprint

[B] "We've covered a massive amount of ground today. We explored the trap of starting from zero and the frustration of amnesia by design. We learned how to escape it by turning our one-off victories into compounding templates that capture the shape of our work, not just the words. We broke down the vital habits of conversation pruning at 20 messages and keeping research and writing in parallel sessions. And we clarified the great divide between managing documents in Cowork and deploying true hands-off automation in the terminal with Claude Code."

[A] "But before we wrap up, I want to leave you with a final thought to mull over. We just spent this entire time talking about extracting the shape of your own best work to raise your personal baseline. But what if you applied this exact same template strategy outwardly?"

[B] "Outwardly? What do you mean?"

[A] "What if you ran this extraction process on your biggest competitor's best work? Think about the mechanics of it. If you took their most successful viral marketing campaigns, pasted them into Cowork, and asked the AI to extract the tension, the flow, the pacing, and the open loops — could you reverse engineer the underlying structural blueprint of why their marketing resonates with your shared audience?"

[B] "That is mind-blowing."

[A] "And if you could extract their blueprint, could you map that against your own templates to see exactly what structural elements you are missing? Here is where it gets really interesting: you aren't copying their words. You aren't stealing their intellectual property or plagiarizing their content. You are simply using the AI to analyze the invisible architecture of their success. It's a structural analysis."

[B] "That is an absolutely brilliant application of this system. Well, thank you so much for joining us on this deep dive. We highly encourage you to take that homework, build your templates, and test the system this week. Let's finally stop starting from zero every morning, and let's cure our AI assistants' amnesia once and for all. We will see you next time."