This page is referenced from every module of the CMO and Creative Director track. Read it once now, bookmark it, and come back when a module says "use the LLM workflow." It assumes you have used ChatGPT once or twice but have not built a real workflow around any of these tools. By the end, you will have a sharper instinct for which model to open for which task, how to talk to it so it produces work you can ship, and where it will quietly lie to you if you let it.
The course site itself was built using Claude, the assistant from Anthropic. That bias is disclosed up front. The recommendations below still hold, because each model has a real seat at the table for different jobs.
1. Which LLM to use, and why
There are five tools worth knowing in mid-2026. Each has a free tier you can actually do work in, and each has a paid tier in the twenty-dollar-a-month range. You do not need more than one paid subscription. Pick the one that matches the work you do most often.
ChatGPT (OpenAI)
Best at: brainstorming, conversational ideation, image generation built into the same chat window, the broadest plugin and integration ecosystem, and the deepest "people I know already use this" network effect. The voice mode is genuinely useful for thinking out loud on a walk.
Avoid for: anything that needs a long, careful, multi-thousand-word draft held together with internal consistency. It tends to drift in long sessions, summarize when you asked it to expand, and lose track of constraints stated thirty messages earlier.
Free tier: usable. You get GPT-5 with daily message limits, image generation with daily limits, voice mode, and memory across chats. For a student doing one marketing exercise per week, the free tier is enough.
Paid tier: ChatGPT Plus at twenty dollars a month removes the message caps and gives priority access during peak demand. Pricing at https://openai.com/chatgpt/pricing.
How to access: chatgpt.com, or the iOS or Android app.
Claude (Anthropic)
Best at: long-form writing that has to stay coherent, document analysis (paste a forty-page PDF and ask questions), code generation, careful reasoning where you want the model to push back rather than just agree. Claude is the model used to build this very course site, and was chosen specifically because it holds its place across long, structured documents better than the alternatives.
Avoid for: image generation (it does not generate images natively), real-time web search by default on the free tier, and casual chitchat where you want a shorter, snappier response. Claude over-explains by default.
Free tier: usable. You get Claude Sonnet 4.7 with a daily message cap, file uploads, and Projects (a folder that holds context across chats). For most students, the free tier handles a week of marketing work without hitting limits.
Paid tier: Claude Pro at twenty dollars a month. Removes the daily cap, adds access to Claude Opus 4.7 for harder reasoning tasks, and gives you Projects with much larger context windows.
How to access: claude.ai, iOS app, Android app.
Gemini (Google)
Best at: anything that touches Google's ecosystem. If you live in Gmail, Docs, Sheets, and Drive, Gemini's integration into those tools is the closest thing to a real assistant. Gemini also has the freshest data of any of the chat models because it pulls live from Google Search by default, which means it is the right tool for "what was announced last week" questions.
Avoid for: anything that needs the model to hold a single, specific voice. Gemini tends toward bland, helpful-sounding prose. It is also the most aggressive of the models at refusing things, which becomes annoying when you are doing legitimate competitive research.
Free tier: usable. You get Gemini 2.5 Flash in the chat interface, deep Google Workspace integration if you have a Workspace account, and a generous quota. The Workspace integration alone is worth keeping it in your back pocket.
Paid tier: Google AI Pro at twenty dollars a month (bundled with extra Google One storage). Adds Gemini 2.5 Pro, longer context windows, and Gemini in Docs and Sheets at a higher rate limit.
How to access: gemini.google.com, or the sidebar in any Google Workspace app.
Perplexity
Best at: research where you need real citations, a focused web search that returns sources rather than vibes, and quick factual questions where you do not want to read ten blog posts. The default model is fine. The "Deep Research" mode is the strongest free-tier research tool of any of the five.
Avoid for: drafting. Perplexity is a search engine with a chat skin. Long-form writing is not what it is for. It is also not the place to brainstorm or do creative work.
Free tier: usable, and the most generous of the five for research. You get unlimited basic searches, a few Deep Research runs per day, and live web access on every query.
Paid tier: Perplexity Pro at twenty dollars a month. Unlocks unlimited Deep Research, your choice of underlying model (GPT-5, Claude Sonnet 4.7, Gemini 2.5 Pro), and file uploads.
How to access: perplexity.ai, iOS app, Android app.
Mistral (Le Chat)
Best at: European data residency, multilingual work (French, Spanish, Italian, German), and a fast, no-fuss interface for short tasks. Mistral is the EU-headquartered option, which matters if you are working with EU clients who care about where their data goes.
Avoid for: cutting-edge reasoning, long documents, image generation parity with the others. Mistral's models are competent but not category-leading in 2026.
Free tier: usable. Le Chat at chat.mistral.ai is free and has no hard daily cap on most models, which is rare.
Paid tier: Le Chat Pro at fifteen dollars a month (cheaper than the US incumbents). Adds higher-quality models, document upload, and image generation.
How to access: chat.mistral.ai.
The honest take
If you are picking one paid subscription, pick Claude or ChatGPT. Pick Claude if your marketing work centers on long-form writing, strategy documents, brand voice, and careful analysis. Pick ChatGPT if your work centers on rapid brainstorming, image-heavy creative, and you want the largest ecosystem of tools to plug into.
Use Perplexity as a free secondary tool, always, for research. There is no good reason to do market research in ChatGPT or Claude when Perplexity exists and is free.
Open Gemini when you need something pulled from your own Gmail or Docs. Open Mistral if you are working in French or need to keep data inside the EU.
That is it. You do not need to pay for more than one of the main models. Anyone telling you otherwise is selling you something.
2. How to structure a prompt
A high-leverage prompt has five parts. Most students skip the first three and wonder why the output is generic. The five parts:
- Role assignment. Tell the model who it is for this conversation. "You are a senior brand strategist with fifteen years at boutique agencies."
- Context loading. Give the model the situation, the audience, and any constraints it needs to know. Paste the relevant document if there is one.
- Task definition. Tell the model what you want it to do. One verb, one object. "Write a positioning statement." Not "help me with positioning."
- Constraints. Word count, tone, what to avoid, what to include. The narrower the constraints, the better the output.
- Output format. A bulleted list, a markdown table, a single paragraph, a numbered ranking. Specify it.
The order matters less than the presence of all five. Here is what bad and good look like in practice.
Bad prompt (a real one, from the wild)
Write me a marketing plan for my business.
This will produce a 600-word document of business-school filler that mentions "target audience," "social media presence," and "engagement metrics" without saying anything about your actual business. The model has no role, no context, no constraints, and no format. It defaults to a generic template.
Good prompt (the rewrite)
You are a marketing strategist who specializes in independent home-services businesses with two to twenty employees. You have read Crossing the Chasm, Obviously Awesome, and The Mom Test.
Context: I run a residential window-cleaning company in Calgary, Alberta. We have been operating for six years, have eight technicians, and do about $1.4M CAD in annual revenue. About 80% of revenue comes from repeat customers on a quarterly cleaning rotation. The remaining 20% is one-off jobs from referrals or our website. Our website gets about 1,200 unique visits a month and converts about 1.5% of them into quote requests. We charge a premium (about 25% above the local market average) and our reviews reflect that we are worth it.
Task: Draft a one-page marketing plan for the next 90 days. The single goal is to lift the quote-request conversion rate on the website from 1.5% to 3%. Assume the budget is $5,000 CAD and that I personally have about six hours a week to spend on marketing execution.
Constraints: No paid advertising. No new website redesign. Tactics must be things I can run with my existing team. Do not recommend "increase social media presence" without specifying what posts, on what cadence, for what goal.
Output format: A markdown document with these sections, in this order: (1) The one hypothesis we are testing, (2) The three most underweighted tactics ranked by expected lift on conversion rate, (3) The weekly cadence for the next 12 weeks as a table, (4) What we will measure and how.
This produces a document you can actually argue with. It might still be wrong on the specifics (you will edit), but the model now has enough constraints that it cannot retreat to generic filler.
The shift from bad to good is not about length. It is about specificity. Every constraint you remove forces the model to guess, and its guesses are always the safest, most generic possibility. Constraints are how you get sharp work.
3. The five workflows that actually matter
There are dozens of ways to use an LLM. Most of them are noise. These five are the ones a working CMO or Creative Director will use every week.
Workflow 1: LLM as research partner
Use it for: gathering market research, customer insights, competitor scans, category audits, "what is the language people use to describe this category" questions.
A good session looks like this: open Perplexity Deep Research, give it a tightly scoped question ("how do residential window-cleaning customers in Calgary describe what they want when they leave a five-star review"), let it run, then take the citations it returns and read three of them yourself. The LLM's summary is the index. The cited sources are the actual research.
Never trust the LLM for: pricing data (always wrong or outdated), customer counts (made up), market sizes (made up), or anything where the precise number matters. Use it to find the source, then verify the number on the source.
Workflow 2: LLM as first-draft writer
Use it for: producing a first draft you will then edit. Press releases, blog posts, sales emails, landing-page copy, internal memos.
A good session looks like this: give the model the brief, the audience, two or three examples of writing in the voice you want, and one example of writing you specifically do not want. Ask for the draft. Read it. Cut everything that does not earn its place. Rewrite anything that sounds like a model wrote it (you will know). Send the edited version to a colleague.
The rule is absolute: never publish a raw LLM draft. Not because it will be terrible (it usually is not), but because raw LLM writing has a recognizable cadence and your audience will pattern-match to it within two sentences. The edit is what makes the work yours.
Never trust the LLM for: brand voice without examples (it defaults to corporate-blog English), specific factual claims (it will smooth over the gaps with plausible-sounding lies), or anything legal or financial that will go to a real audience.
Workflow 3: LLM as critic
Use it for: pasting your own writing and asking the model to grade it against criteria you specify. This is the most underused workflow in this list and probably the most underweighted one.
A good session looks like this: paste your draft, then give the model a rubric. "Grade this on (1) clarity of the single message, (2) specificity of claims, (3) presence of jargon, (4) whether it tells the reader what to do next. Be harsh. Do not flatter me." Read the critique. Disagree with about a third of it (the model is sometimes wrong). Apply the other two-thirds.
The trick is asking for harshness explicitly. The default behavior of every major model is to be encouraging. You have to tell it to stop. Phrases that work: "You are reviewing this draft as a skeptical senior editor. Your job is to find what is weak, not to make me feel good. Be specific." Phrases that do not work: "give me feedback."
Never trust the LLM for: feedback on legal language, feedback on whether a claim is true (it is grading the rhetoric, not the facts), or feedback on whether your audience will respond. It is a critic, not a focus group.
Workflow 4: LLM as analyst
Use it for: pasting CSV data and asking for analysis, segmentation, trends, or anomalies. Most modern chat interfaces (ChatGPT, Claude, Gemini) accept CSV uploads directly. You drop the file in and ask questions in plain English.
A good session looks like this: paste a CSV of your last 200 customers (anonymized; no PII), ask the model to segment them by some dimension, then ask follow-up questions. "Which segment has the highest average order value? Which has the highest repeat rate? Which two segments are most different from each other in their buying behavior?" The model will do the math, name the segments, and surface things you would not have noticed scanning the spreadsheet.
For real numerical analysis with multiple steps, ChatGPT and Claude both have a "code interpreter" or "analysis" feature that runs actual Python on your data instead of just eyeballing it. Use this whenever the question involves any calculation. The model's plain-English math is unreliable. Its Python is correct.
Never trust the LLM for: anything where the number has to be exact for a financial model, regulatory filing, or board deck. Even with code interpreter, you check the work.
Workflow 5: LLM as Socratic partner
Use it for: thinking out loud, pressure-testing positioning, working through a decision where you have not made up your mind. This is the workflow that feels least like "using AI" and most like talking to a smart, infinitely patient colleague.
A good session looks like this: open a chat, write a paragraph or two about the problem you are wrestling with, and ask the model to ask you questions. "Ask me three questions that would help me figure out whether this positioning works. Do not give me answers yet." Then answer the questions. The model will follow up. Forty-five minutes later you have a clearer position than you would have arrived at staring at a blank document for the same length of time.
This is also the workflow where switching models matters most. Claude is good at this because it tends to push back. ChatGPT is good at this because it has a conversational ease. Try both, find your preference.
Never trust the LLM for: the actual decision. The Socratic partner is there to clarify your thinking, not to replace it. If the model gives you an answer in this mode, ignore it and ask the next question.
4. What to never trust an LLM for
Every model in 2026 still has known failure modes. You can use them around the failures, but you have to know where they are.
Recent data. No model's training data is current. Even with web search enabled, the model can grab a stale page from 2024 and present it as today's truth. Publication dates, current pricing, recent news, recent product launches, "what happened last week" questions. Verify the date stamp on anything you got from an LLM. If the answer depends on freshness, go to the source.
Citations. This is the single most dangerous failure mode for a marketer doing research. LLMs hallucinate citations regularly. They will invent a study with a plausible-sounding title, a plausible-sounding author, a plausible-sounding journal, and a plausible-sounding URL. The study does not exist. The author does not exist. The URL 404s. This happens across every model. If you cite a source the LLM gave you without opening the URL and reading the actual page, you will eventually get caught. Open every URL. Read the page. Confirm the claim.
Numerical precision in financial models. The model will get the structure of your spreadsheet right and the arithmetic wrong. It is doing pattern-matching, not math. For anything where the number has to be correct (a board deck, a budget, a financial projection), use the code-interpreter or analysis mode that actually runs Python, then check the inputs yourself.
Brand voice without examples. The model has no idea what your brand sounds like. If you ask for "copy in our brand voice" without pasting two or three real examples of your voice, you will get generic corporate-blog English. The fix is showing, not telling. Paste three real examples. Then ask for the draft. The output gets dramatically better.
Legal language. Contracts, terms of service, privacy policies, anything binding. The model will produce something that reads like a contract. It is not a contract. It will miss jurisdiction-specific requirements, miss new regulations, and confidently include language that does not mean what it appears to mean. Use the model to draft a first pass for your lawyer to mark up. Never use it as a replacement for a lawyer.
Anything that ships without your editing pass. This is the meta-rule. If a draft is going out to a real audience (clients, prospects, the public, a board), you read it line by line and rewrite anything that does not sound like you. The cost of skipping this is your reputation. The cost of doing it is twenty minutes.
5. Five worked examples
These map to specific modules in the course. Each example gives you the prompt, the kind of output to expect, and what to do with that output. Treat the prompts as starting points. Adapt them to your business.
Example 1: Building a positioning statement (Module 2)
The model you want for this is Claude, because positioning needs careful structure and Claude holds its place better in long, structured conversations.
Prompt:
You are a positioning expert in the tradition of April Dunford and Geoffrey Moore. You have read Obviously Awesome and Crossing the Chasm. You believe positioning is a deliberate choice, not a tagline, and that a positioning statement is the input to marketing, not the output of it.
Context: I run [your business]. We sell [product or service] to [primary customer]. The three things customers tell us they value most about us are [trait one], [trait two], and [trait three]. Our two biggest competitors are [competitor A] and [competitor B]. The category we are currently in is [category as customers would name it].
Task: Draft five distinct positioning statements for my business. Each should follow this structure: For [target customer] who [statement of need or opportunity], our [product or service] is a [category] that [statement of key benefit, i.e., compelling reason to buy]. Unlike [primary competitive alternative], our product [statement of primary differentiation].
Constraints: Each of the five statements should make a different bet. One should bet on a different target customer. One should bet on a different category. One should bet on a different competitive alternative. One should be the most conservative version. One should be the most ambitious version. Label each.
Output format: A numbered list of five positioning statements, each with a one-sentence note explaining the bet it is making.
Expected output: Five real options to argue with. They will not be your final positioning. The point is to have five distinct choices on a page so you can see the tradeoffs.
What to do with it: Pick the two you find most provocative and run them past five customers. Ask each customer which one describes you best. The answer is rarely the one you expected.
Example 2: Generating a creative brief (Module 4)
The model you want is ChatGPT or Claude. Either works. Use whichever you have open.
Prompt:
You are a senior creative strategist. You have read Julian Cole's writing on creative briefs and you believe the best brief is one page, has a single insight, and gives the creative team a real problem to solve.
Context: I am the marketing lead at [your company]. We are about to brief [internal team or agency] on a campaign for [product, service, or moment]. The business goal is [the business goal, in one sentence]. The audience is [the audience, as a real person]. The single behavior we want from this audience is [the specific behavior, e.g., book a discovery call, sign up for the newsletter, switch from competitor X].
Task: Draft a one-page creative brief for this campaign using the GET / WHO / TO / BY framework. GET is the audience. WHO is the insight about that audience that makes the strategy possible. TO is the behavior change we want. BY is the strategic role of the work (proof, demonstration, reframe, etc.). End the brief with a single sentence that completes "If this campaign works, our audience will think..."
Constraints: Do not include any tactics. The brief is upstream of tactics. If the team asks for tactics in the brief, you have failed.
Output format: Markdown, with GET, WHO, TO, BY as headers. End with the "If this works..." sentence. Keep the whole brief under 300 words.
Expected output: A real one-page brief. It will not be perfect. It will probably need a second pass to tighten the insight.
What to do with it: Read it out loud to one person who is not on the project. If they say "I get what this is for," ship it to the creative team. If they say "I am not sure what this is for," the insight is not sharp yet.
Example 3: Designing a customer persona from research (Module 8)
The model you want for this is Claude, because you are going to paste a long block of interview notes and you need the model to hold all of it in working memory.
Prompt:
You are a customer research analyst. You have done hundreds of customer interviews and you believe personas are useful only when they are grounded in real verbatims from real customers, never invented from demographic data.
Context: Below is a transcript of [number] customer interviews I conducted. Each interview was [length] minutes. The customers were [how you sourced them, e.g., paying customers who renewed within the last 90 days].
[Paste the interview notes or transcripts here. Aim for 5,000 to 20,000 words of real customer language.]
Task: From these interviews, draft two customer personas. Each persona should be built from real verbatim quotes (cite them inline). For each persona, surface (1) the underlying job-to-be-done in the customer's own words, (2) the two or three forces that were pulling them toward our solution, (3) the two or three forces that were pulling them away or making them hesitate, and (4) the specific moment when they decided to buy.
Constraints: No demographic filler. No "Sarah is a 38-year-old marketing manager who likes yoga." Personas are built from jobs and forces, not from age and hobbies. If a verbatim is not in the source material, do not invent it.
Output format: For each persona, a one-page document with these sections: Name (one descriptive phrase, not a fake first name), The Job, The Pulls, The Pushes, The Moment.
Expected output: Two personas you can actually use to write copy from. They will sound like real people because they are built from real words.
What to do with it: Hand them to your copywriter and to whoever is doing your ad creative. The persona document replaces the demographic profile that nobody reads.
Example 4: Writing a sales objection handling cheat sheet (Module 11)
The model you want is ChatGPT, because this is the kind of fast, list-based work it does well.
Prompt:
You are a B2B sales coach. You have closed seven figures in deals personally and you believe every objection is a request for more information, not a request for a discount.
Context: I sell [product or service] to [primary customer segment]. The price point is [price or pricing range]. The five most common objections I hear in sales calls are: (1) [objection], (2) [objection], (3) [objection], (4) [objection], (5) [objection].
Task: For each of the five objections, write a one-page response that contains: (1) what the customer is actually saying underneath the objection, (2) the question to ask to confirm that interpretation, (3) the reframe that addresses the real concern, (4) the proof point or example that lands the reframe, (5) the next-step ask that moves the conversation forward.
Constraints: No "feel, felt, found" scripts. No language that sounds like it came from a 1990s sales training. The reframe should sound like a real human talking, not a script reading.
Output format: One page per objection. Headers in the order above. Keep each page under 250 words.
Expected output: A five-page document you can use to coach a new salesperson in an afternoon.
What to do with it: Run the responses past your best salesperson. Cut anything that does not match how they actually sell. Print the survivors and keep them next to the phone.
Example 5: Crafting an honest LinkedIn About section (Module 10)
The model you want is Claude, because LinkedIn About sections are short and Claude is better at restrained, dignified prose than at energetic prose.
Prompt:
You are a writer who specializes in dignified LinkedIn profiles for senior operators. You have read Patrick Awuah on humility, Marc Andreessen on betting on yourself, and David Ogilvy on writing. You believe LinkedIn profiles fail because they try to sound impressive instead of trying to sound true.
Context: I am [your role] at [your company or businesses]. Before this I [your prior background in one or two sentences]. The work I do today is [a one-sentence description of what your day-to-day actually is]. The kind of person I want to attract through LinkedIn is [the target, e.g., a future client, a future employer, a future collaborator]. Three things about my work that most people do not realize: [one], [two], [three].
Task: Draft three versions of a LinkedIn About section, each in a different register. Version one: restrained and direct, no hooks, no caps, no emoji. Version two: warmer, with one personal detail that humanizes me. Version three: the version that takes the most professional risk by saying what I actually believe about my field.
Constraints: No "passionate about" anywhere. No "results-driven." No "thought leader." No bullet lists of accomplishments. No quote at the end. Keep each version under 200 words.
Output format: Three versions, labeled Version 1 (Restrained), Version 2 (Warm), Version 3 (Risk).
Expected output: Three drafts that sound like a person, not a brand. One of them will be obviously you. The other two will show you parts of your voice you had not used yet.
What to do with it: Show all three to one person who knows you well. Ask which one sounds most like you. Use that one.
6. The workflow that beats every other
The standard research-to-publish loop has five steps: research, first draft, critique, revise, publish. Every working creative director already knows this. The shift with LLMs is that the model becomes a participant at every step. You research with the LLM. You first-draft with the LLM. You critique with the LLM (and you also have the LLM critique you). You revise with the LLM. You publish, alone, because publishing is the one step the model cannot do for you and cannot share responsibility for.
This loop beats "let the LLM do it for you" for a simple reason. The single-shot approach (one prompt, one output, one publish) produces work that any other LLM-using marketer could have produced from the same prompt. The looped approach produces work that is shaped by your judgment at every step, which means it ends up sounding like you and reflecting decisions only you would have made. The model is a tireless assistant. You are the operator. That distinction is the whole career.
The students who get the most out of these tools, six months in, are the ones who treat the LLM as a colleague who is sometimes brilliant, sometimes wrong, always patient, and never the final decision-maker. The ones who get the least are the ones who let the model decide.
Open a chat. Write a real prompt. Edit the output. Ship the work. Come back to this page when a module reminds you to.