The AI Playbook for Marketers: Build It Once, Stop Cold-Prompting Forever.

Welcome back to the front row.

Quick note on last week: I was recovering from an injury, so the issue didn't ship — the first one I've missed. I'm back, I'm fine, and today's issue is the proof.

Every fourth issue of NVM, we go deep. Not a roundup. Not three bullets and a take. A proper field guide on one topic, written for practitioners who need to actually do something with it, not just nod along. The regular format returns next week. This is the one you save.

Issue #12 is Deep Dive No. 03. The topic: the reusable AI infrastructure a growth marketer should build once; workspaces, packaged assistants, memory, and saved workflows; so you stop re-explaining your brand, your account, and your rules every single morning.

You've been paying for these tools by the month and using them by the prompt. That's the leak. This is the issue that closes it.

DEEP DIVE III: THE AI PLAYBOOK

What to Build Once So You Never Cold-Prompt Again

I ended Deep Dive No. 02 with one line: build a Claude Project, upload your brand guide; that's not a tip, that's the whole infrastructure.

Half of you replied asking the same thing. What else goes in it?

This is the answer. The whole build.

Here's the first thing to get straight: a prompt is disposable and infrastructure compounds. Every time you open a blank chat, paste your brand voice, re-describe your audience, re-explain your account structure, and then ask your question — you're doing setup work you already did last week. And the week before. The output is only as good as the context, and you're rebuilding the context from zero every time.

Infrastructure is the fix. It's the context you install once and reuse forever. And it lives in three layers, whether you've named them or not.

Layer one: workspaces — persistent containers that hold your files, instructions, and history in one place (Claude Projects, ChatGPT Projects). The context lives in the room, not in the prompt.

Layer two: assistants — packaged, reusable roles with a fixed job and their own knowledge (Custom GPTs, Gemini Gems, Claude Skills). You build the specialist once; you summon it on demand.

Layer three: memory and automation — the persistence that carries what matters across sessions, and the saved workflows that run without you (memory settings, subagents, scheduled tasks).

Start with the layer that matches your biggest leak. Most of you leak worst at layer one.

LAYER 1: WORKSPACES — CONTEXT LIVES IN THE ROOM

The mistake everyone makes first

You treat every chat as a fresh start. You open ChatGPT, paste the brand guidelines, and get to work. Tomorrow you open a new chat and paste them again. A workspace ends that. It's a container that holds your instructions, your reference files, and your conversation history — and every chat inside it inherits all of that automatically.

Claude Projects — Best for: brand-voice-sensitive work, long documents, agencies that need one shared context

A Project is a workspace that groups related chats, uploaded files, and custom instructions under a persistent context, so every conversation inside it starts already knowing your brand. You can upload up to 20 files per chat at 30 MB each, and — this is the part people miss — project content is cached and doesn't count against your per-message usage limits. You can load a heavy brand bible once and stop paying for it in tokens every message.

Where it genuinely outperforms everything else:

• Holding brand voice across long, instruction-heavy work without re-priming
• Team context: on Team and Enterprise, Projects can be shared so everyone works from the same instructions and knowledge base — one source of truth for the whole agency
• Keeping a living account context (structure, rules, past decisions) that every new chat inherits

Where to use something else: if the job is 30 fast creative variations, you don't need a workspace — open a quick chat. Workspaces earn their keep on recurring, context-heavy work, not one-off volume.

Set it up right now:

1. Create a Project named for the client or account (e.g., "ACME — Google Ads").
2. Upload three files: your brand voice guide, your account structure/rules, and your last 90 days of performance notes.
3. Paste this into the Project's custom instructions:

"You are my performance marketing partner for [ACCOUNT]. Voice: [2–3 sentences]. Audience: [specifics]. Rules: no exclamation points, cite the data file before recommending, flag any claim you can't verify. Default to MENA context unless I say otherwise. When I ask for copy, produce it in the brand voice above."

Best practice: one Project per client or account, never one giant Project for everything. Context bleeds when you mix accounts. Clean rooms, clean outputs.

ChatGPT Projects — Best for: mixed workflows where you also want images, browsing, and volume in the same room

ChatGPT Projects are smart workspaces that group chats, reference files, and custom instructions so the model stays on-topic across a long-running effort. The reach advantage is real: Projects are available to all free and paid tiers globally, and Business, Enterprise, and Edu users can share them with teammates. If your team lives in ChatGPT, this is your container.

Where it genuinely outperforms everything else:

• Keeping creative volume, image generation, and browsing inside one persistent context
• Onboarding a whole team into a shared workspace without a paid-tier barrier at the entry level

Where to use something else: for tight brand-voice drafting, Claude holds a style across long output more faithfully. Use ChatGPT Projects for range; use Claude Projects for voice.

Best practice: whichever workspace you pick, the rule is the same — the brand guide goes in the room once, not in the prompt every time.

LAYER 2: ASSISTANTS — BUILD THE SPECIALIST ONCE

A workspace holds context. An assistant packages a job. It's a fixed role — with its own instructions, knowledge, and default behavior — that you build one time and call whenever you need that exact task done the exact same way.

Custom GPTs — Best for: repeatable, templated tasks you run weekly

A Custom GPT is a saved ChatGPT configuration with its own system prompt, knowledge files, and tools. You can attach up to 20 files at 512 MB each, and since June 2025 you can select any available model when building one, with a "Recommended Model" fallback for teammates whose tier lacks the configured model. The GPT Store has hosted hundreds of thousands of user-built GPTs since it launched in January 2024 — but the ones worth having are the ones you build for your own recurring work.

Where it genuinely outperforms everything else:

• Encoding a recurring deliverable — "RSA generator for my account," "weekly report summarizer" — so it runs identically every time
• Handing a teammate a task-shaped tool instead of a prompt they'll paste wrong

Where to use something else: for one-time thinking, a Project or a plain chat is faster. Build a GPT only for the task you'll run more than five times.

Gemini Gems — Best for: Google Workspace teams, Drive-native knowledge, Arabic + English brand assistants

A Gem bundles a name, instructions, an optional default tool, and knowledge files into a reusable assistant. You can attach up to 10 knowledge files, and when you reference a file from Google Drive the Gem always uses the most recent version — so your brand guide updates everywhere the moment you edit the Doc. Inside Gemini for Workspace, Gems can be shared with your organization, and you can set a default tool (Deep Research, Canvas, image) that launches on every new chat.

Where it genuinely outperforms everything else:

• Drive-linked knowledge that stays current without re-uploading
• A shared org-wide brand assistant for teams already living in Docs and Sheets
• Running a parallel Arabic Gem and English Gem, each with its own localized style guide, so voice holds in both languages

Where to use something else: for stylistic precision on long-form English, Claude still drafts cleaner. Use Gems for integration and freshness, not for the finest voice control.

Set up a brand assistant right now:

1. New Gem → name it "[BRAND] Voice Assistant."
2. Instructions: paste your voice rules + three do/don't examples.
3. Knowledge: link your brand guide Doc from Drive (auto-updates).
4. Default tool: Canvas, for iterative drafting.
5. Save, then share to your org if you're on Workspace.

Best practice: build two assistants before you build ten — a voice assistant (writes in brand) and a research assistant (briefs before you write). Those two cover most of the week.

Claude Skills — Best for: turning a repeatable process into a callable capability

Skills package a specific, repeatable process — a checklist, a format, a workflow — into something you invoke by name instead of re-describing. Your own Playbook already covers the build (Set Up Claude Code for a Marketing Team and How to Build Your First Claude Skill in 10 Minutes). The infrastructure point: a skill is a process you'll never have to re-explain. That's the whole game.

LAYER 3: MEMORY & AUTOMATION — PERSISTENCE AND PIPELINES

Workspaces and assistants hold context you put there on purpose. This layer handles the context that should carry itself — and the work that should run without you.

Memory — Best for: killing the repetition you don't even notice

Chat memory derives summaries of past conversations and carries them forward, viewable and editable at Settings → Capabilities → Memory. It shipped to Claude Team and Enterprise in September 2025 and reached Free in March 2026. ChatGPT's newer memory system updates automatically, keeping the details it judges most important so you don't repeat yourself.

The practitioner move: memory is a setting, not a strategy — so audit it. Open the memory panel monthly and delete anything stale or wrong. A model quietly carrying an outdated account fact into every answer is worse than one carrying nothing.

Where to be careful: memory that persists across clients can leak one account's context into another's work. For clean separation, keep client-specific context inside per-client Projects, not global memory.

Automation — Best for: the recurring work that shouldn't need you

The last layer is where infrastructure stops being passive. Subagents let you split a job across specialized roles — a researcher, a writer, an editor — each owning one task. Scheduled tasks run a workflow on a cadence without you opening anything. Your own Playbook has the pattern: 5 Claude Subagents Every Marketing Team Should Set Up. This is the same architecture that now builds this newsletter every Sunday — research, draft, edit, and deliverables split across subagents, assembled for review.

Best practice: don't automate a workflow you haven't run manually five times. You can't hand off a process you haven't yet made repeatable.

THE INFRASTRUCTURE MAP

Screenshot this. It's the whole build on one screen.

Layer	Claude	ChatGPT	Gemini	What goes in it	Setup
Workspace	Projects (20 files / cached)	Projects (all tiers)	— (use Gems)	Brand guide, account rules, performance notes	20 min/account
Assistant	Skills	Custom GPTs (20 files)	Gems (10 files, Drive-linked)	One voice assistant + one research assistant	15 min each
Memory	Settings → Memory	Auto memory	Personal context	Audited monthly; client context stays in Projects	5 min/month
Automation	Subagents + scheduled tasks	Custom GPTs as pipelines	Default-tool Gems	Only workflows you've run 5× manually	45 min each

The default build for one marketer: one Project per client, two assistants (voice + research), memory audited monthly, one automation once a workflow proves itself. That covers 90% of the repetition you're doing by hand today.

PUTTING IT TOGETHER: THE 90-MINUTE BUILD

Stop thinking about this as a project you'll get to. It's an afternoon.

Minutes 0–20 — The workspace. Create one Project for your main account. Upload brand guide, account structure, last 90 days of notes. Paste the custom instructions.

Minutes 20–50 — The two assistants. Build a voice assistant (writes in brand) and a research assistant (briefs before you write). Give each its knowledge files. Name them so a teammate could use them tomorrow.

Minutes 50–65 — Memory. Open the memory panel. Delete what's stale. Add the two or three durable facts about how you work that you're tired of repeating.

Minutes 65–90 — One automation. Take the one workflow you've already run by hand a dozen times — the Monday research sweep, the weekly report summary — and set it up to run itself.

The marketers getting outsized leverage from AI aren't the ones with the most subscriptions or the cleverest prompts. They're the ones who did this afternoon once. After that, every prompt starts from a running start instead of a blank page.

Turn this into action. The full click-by-click build lives in the NVM Playbook: The 90-Minute AI Infrastructure Setup — copy-paste instructions, the exact custom-instruction blocks, and the two assistant configs. Words into actions. That's what the Playbook is for.

RESOURCES — GO DEEPER

Bookmark these. Every one is a direct link to something you can use today.

From the NVM Playbook:

• The 90-Minute AI Infrastructure Setup — the click-by-click build for this issue
• Set Up Claude Code for a Marketing Team From Scratch — one config file, team-wide context
• 5 Claude Subagents Every Marketing Team Should Set Up — the automation layer, built
• 10 Claude Skills That Quietly Replace Half Your Marketing Stack — assistants that retire subscriptions
• 12 GitHub Repos Every AI Marketer Should Bookmark — the open-source shortlist

Open source, straight from GitHub:

• anthropics/anthropic-cookbook — official Claude recipes, including sub-agent patterns you can lift directly
• modelcontextprotocol/servers — connect your assistants to real tools and data (memory, filesystem, GitHub, Postgres)
• promptfoo/promptfoo — grade your AI output consistently before you ship it
• n8n-io/n8n — own your automation layer instead of renting it
• ai-boost/awesome-prompts — curated prompts from the top-rated GPTs, for seeding your own assistants

PROMPT OF THE WEEK

Building on this issue: the prompt that generates the brand-guide file you'll load into every workspace above.

Act as a brand strategist. Interview me to build a one-page brand context file I can load into an AI workspace.

Ask me, one at a time: (1) who the audience is, in specifics; (2) the voice in 3 adjectives + one sentence; (3) three phrases we always use and three we never use; (4) the one thing our content must always do. After my answers, output a single tight file titled "[BRAND] Context" with sections: Audience, Voice, Do/Don't, Non-negotiables. Under 400 words. No filler.

What this gives you: the one file that makes every Project, GPT, and Gem you build instantly brand-aware — created in five minutes of answering questions instead of an afternoon of writing.

THIS WEEK

Claude's chat memory reached the Free tier in March 2026. The persistence that shipped to Team and Enterprise back in September 2025 is now available to everyone — viewable and editable in Settings → Capabilities → Memory. If you're on Free and haven't opened that panel, you've been leaving context on the table.

Gemini Gems now pull knowledge live from Google Drive. Link a Doc as a Gem's knowledge file and it always uses the most recent version. Edit your brand guide once; every Gem referencing it updates. For Workspace agencies, this is the cleanest single-source-of-truth setup available right now.

ChatGPT Projects are available on every tier, globally. Free and paid users all get persistent workspaces; Business, Enterprise, and Edu can share them. There's no longer a paywall excuse for your team cold-prompting — the container is free.

Custom GPTs can run any model since June 2025. Builders pick from any available model, with a Recommended Model fallback for teammates on lower tiers. Your task-shaped tools are no longer locked to one model's strengths.

TOOL OF THE WEEK

Claude Projects — Free tier and up | claude.ai

Not new. Almost certainly underused. It's the fastest place to feel the difference between prompting and infrastructure, because the caching means your loaded context is effectively free per message. The use case most marketers miss: a per-client Project that holds the brand guide, the account rules, and the running history — so every new chat opens already knowing the account instead of asking you to explain it.

How to use it right now: create one Project for your most active client. Upload three files. Paste the custom-instruction block from Layer 1. Your next twenty conversations about that account start smarter than your last two hundred did.

Why it matters this week specifically: with memory now on every tier and Projects shareable across teams, the gap between marketers who built their infrastructure and those still cold-prompting is about to get visible in output quality. The setup cost is one afternoon. The compounding is permanent.

MY TAKE

I keep meeting marketers who are convinced their problem is prompting. They collect prompt packs. They tweak wording. They chase the magic phrasing.

It's the wrong layer.

The best prompt in the world, typed into a blank chat that knows nothing about your brand, your account, or your rules, produces a confident stranger's answer. The most average prompt, typed into a workspace that already holds all of it, produces something you can almost ship. The leverage was never in the sentence. It was in the room the sentence was typed in.

So stop optimizing the disposable thing. Build the thing that compounds. One Project. Two assistants. Memory you actually maintain. One workflow that runs itself. An afternoon of work that makes every day after it faster.

Last time I said building a Claude Project was the whole infrastructure. I was pointing at the door. This is the house.

Build it once. Then go do the work only you can do.

That's Issue #12 of NVM. If this gave you one thing to act on this week, share it with someone who needs it. Reply to this if you have a question — I read every one. See you next week. — Mostafa

NVM · Front row seats to AI and digital marketing

• Subscribe → nvm.marketing/newsletter
• Join the community → nvm.marketing/community
• Write for NVM → nvm.marketing/blog