Marketing is the role with the most AI hype and the biggest gap between hype and reality. The trap: "Claude writes my marketing" produces generic content faster, which is worse than slow generic content. The escape: a real source artifact every week, a project that holds your voice, and a quarterly performance loop. Volume × consistency × voice — not volume alone.
Block 60 minutes every Monday morning. Capture one genuine artifact from the business — a customer call, a project ship, a founder POV, a teardown — and turn it into the week's distribution across your channels. The discipline isn't writing. The discipline is having a source.
Almost every bad piece of AI-generated marketing has the same origin: somebody asked Claude to "write a LinkedIn post about [topic]" with no source artifact. The output is structurally fine and substantively nothing — because there was nothing to start from. The Source Hour reverses this. You start with something real, and Claude does the distribution work.
In 60 minutes: 15 minutes to capture or identify this week's source, 30 minutes to generate the channel-specific distribution from it, 15 minutes to review last week's performance and note one learning back into the project. That's the loop.
Record with permission. Transcribe. The one objection they raised, the one thing they said in their own words, the one moment they got it — each one is a post. Highest-leverage source there is.
A feature, a project completed, a customer outcome. Not the press-release version — the technical version, the trade-offs version, the "here's what we tried that didn't work" version.
Something the owner believes about the market that most competitors don't. Five minutes of voice memo from the owner is a week of content. Get the voice memo.
A competitor's marketing, a public example, an industry trend. Generous, specific, with your actual take. Not snark — analysis.
How you actually do X. The unpolished version. The boring real version is almost always more interesting than the case study version, because the case study version sounds like everyone else's.
One thing your audience gets wrong, explained clearly, with no pitch attached. The "teach, don't pitch" instinct most small businesses fight. Lean into it.
Why this works. Content fails because most of it is generated from thin air. Content that compounds comes from real sources — calls, ships, beliefs, teardowns. Claude is excellent at distributing a real artifact across channels. It's terrible at inventing one. The Source Hour puts each of you on the right side of that line.
The single biggest reason AI-generated marketing sounds like AI-generated marketing is that the prompter never gave Claude examples of their voice. The project fixes this. Once built, every prompt below produces drafts that sound like you, not like a default brand voice trained on the rest of the internet.
Download the template. Fill it in honestly — 45 to 75 minutes the first time. The voice section is what separates this from a generic brand doc, so put your top 10 posts in as knowledge files and write the "why each worked" notes for each.
Open Claude → Projects → New Project → name it "Brand Voice & Audience." Paste the completed file. Upload past hits as files. Now every Source Hour starts with Claude already knowing what you sound like.
The "differentiated point of view" section is the marketing secret. Almost every small-business brand doc has voice notes ("we sound warm but professional") and audience notes ("our buyer is a busy SMB owner"). Almost none have an explicit list of what we believe about our market that our competitors don't. Without that section, Claude pattern-matches to the average voice in your category — which is the voice you're trying to escape. With it, Claude has something to push from.
Used inside your Brand Voice project, so Claude already has your voice, your ICP, and your past hits in context. Fill in the [bracketed] specifics and copy-paste.
Source artifact: [PASTE TRANSCRIPT / LINK / DESCRIPTION] Type of source: [customer call / ship / founder POV / teardown / behind-the-scenes / teaching] Based on this source and our voice, generate this week's distribution across our channels (in the project): For each channel, produce a draft. For each draft: 1. Name the angle — which specific thread from the source you pulled 2. Write the draft 3. Add one note: what I'd need to verify, attribute, or watch for before publishing Don't generate a draft for any channel where this source doesn't fit. Better to ship 3 strong drafts than 5 mid ones. End with the question: is there a stronger angle in this source I missed?
Attached is a voice memo / transcript from [PERSON]. Topic: [TOPIC]. Turn it into a long-form post for [CHANNEL]. Rules: 1. Preserve the original voice. Don't smooth out the rough edges that make it sound like a real person — those are the asset. 2. Keep their phrasings. If they said "rework not speed," don't translate it to "quality not velocity." 3. Reorganize for clarity, but don't add ideas they didn't say. 4. Cut filler, asides, and repetition. 5. End with whatever conclusion they were building to — don't invent a CTA. Length: [WORD COUNT]. After the draft, tell me the two sentences I should consider cutting, and why.
Attached is a draft [someone else wrote / I wrote in a hurry / our agency sent]: [PASTE] Rewrite it in our voice (the voice and example sentences in this project). Don't just swap words. Check: 1. Does it use any of the words we never use? Replace. 2. Does it have a real point of view, or is it consensus copy? If consensus, name the missing POV. 3. Is the hook ours, or generic? Our hooks are [describe pattern or paste 2-3 examples of past hits' first lines]. 4. Does the CTA match our channel norms? After the rewrite, give me a 3-bullet diff: what you changed and why. I want to learn from this, not just accept it.
Attached: transcript of a call with [CUSTOMER]. Context: [What we did for them, when, the rough outcome.] Draft a case study in our format. Structure: 1. The situation in their own words (quote) 2. What they tried before us (specific, not "previous solutions") 3. What we actually did (the unsexy operational version, not the marketing version) 4. The outcome in numbers, where they exist, and "soft" outcomes where they don't 5. One thing they said that's quotable, attributed, in their voice not ours Rules: - Use their actual words wherever you can. Paste-marked quotes. - Don't dramatize. Our voice doesn't. - If something in the transcript doesn't support a claim, don't make the claim. End with: 3 things I should confirm with them before publishing, and 2 alternative angles from this same call.
Inputs from this week: [wins, ships, learnings, anything notable] Draft this week's newsletter using our standard structure (in the project). Rules: 1. Pick ONE thing to be the main story — not three. Three "main stories" = no main story. 2. The other items become brief mentions, max one sentence each. 3. Our subject line patterns are [paste 3-5 past subject lines that worked]. Stay in that style. 4. The opening line earns the open. If it's "Hey friends, hope your week is going great" you've failed. 5. End with something specific the reader can do or think about — not a generic CTA. Length: [target]. After the draft, propose 3 alternative subject lines and rank them.
I want to reach out to [PERSON] at [COMPANY] about [REASON]. What I know about them: [LinkedIn / company info / mutual connections / recent activity]. Draft 3 outreach options: 1. The "specific observation" version — opens with something only I would have noticed about them or their company 2. The "asymmetric value" version — leads with something I can give them before asking for anything 3. The "direct" version — names what I want and why I think it might be mutually useful, in under 90 words For each: subject line + body. Rules: - No "I hope this finds you well." No "I wanted to reach out." No false casualness. - No invented context. If we have no real connection, don't fake one. - Personalization tokens like {{first_name}} are not personalization. After the drafts, tell me which one you'd send if this were your business, and why. If none of them are good enough to send, say so.
Attached: last month's performance data (emails, social, blog, whatever we publish). The top 5 by engagement and the bottom 3. Analyze: 1. What do the top 5 have in common? (Format, topic, hook style, length, time of day — be specific. "They were engaging" is not an answer.) 2. What do the bottom 3 have in common? 3. Is there a pattern that contradicts what I think I know about our audience? 4. What's the single move to make in next month's content based on this? 5. What's the move NOT to make — the false pattern from too small a sample size? End with one specific update to the "Past hits" section of the project — the new hits to add, and what to note about them.
Attached: questions our customers have actually asked in the last 30–60 days — from calls, support tickets, sales conversations, social comments. Raw, unfiltered. Generate next month's content theme list: 1. Group the questions into 5–7 underlying themes 2. For each theme: the specific question that best represents it, the angle we'd take on it, the channel where it would land best 3. Flag any theme that's actually a product/positioning issue, not a content issue (those don't get solved with a blog post) 4. The one theme I should NOT make content about, even though the questions are coming in — because answering publicly would be premature, off-brand, or out of scope 5. Rank the themes by signal-to-noise — which one is the audience most starved for, vs. which is just loud This is calendar input, not a calendar. The ranked themes are the deliverable.
It's the end of Q[N]. Attached: every piece we published this quarter and the performance data. Audit: 1. Volume: did we ship what we said we'd ship? If not, what slipped and why? 2. Mix: what was the breakdown by source type (customer call / ship / POV / teardown / BTS / teaching)? Where's the gap? 3. Voice consistency: rate Q[N] vs. our voice guide on a 1–10. Where did we drift? 4. Performance: which 3 pieces would I want to study and try to replicate? 5. What we should kill: any channel, format, or recurring piece that isn't earning its time? Be direct. End with three specific moves for next quarter — not generic ("post more"), specific ("replace the monthly roundup format with a quarterly customer-call series").
This piece underperformed: [PASTE OR LINK]. Performance: [METRIC vs. usual]. Don't be diplomatic. Diagnose: 1. Was the hook weak? Quote the specific line that should have been the hook and wasn't. 2. Was the angle wrong for the channel? Where should this have gone instead, if anywhere? 3. Was the take consensus? If everyone in our space would have said this, it's why nobody engaged. 4. Was the length wrong? Specifically, what should have been cut, or what should have been added? 5. Was the timing or distribution the problem, not the piece itself? End with: would I make this exact piece again? If yes, what's different. If no, what's the version I'd make instead.
Draft piece attached: [PASTE] Current headline / subject / first-line hook: [PASTE] Channel: [CHANNEL] Generate 10 alternative headlines / hooks. Rules: 1. Match our voice (project). No clickbait, no "you won't believe." 2. Each one tests a different angle — don't give me 10 variations of the same idea. 3. At least 2 should be uncomfortably specific. 4. At least 2 should be uncomfortably plain. 5. None should use the words we never use (project). Rank them 1–10 with one line on why each is at its rank. Then tell me which one you'd publish if this were your business — including if it's the original.
Draft attached: [PASTE]
Be the honest mirror. Read this as a reader who is allergic to AI-generated content. Specifically check:
1. Sentences that start with "In today's fast-paced..." or any equivalent throat-clearing — flag every one.
2. Lists of three where the third item is filler.
3. Vague abstractions where a specific would land harder.
4. The "AI cadence": short. Punchy. Then short. Punchy. Again. (If you see it, name it.)
5. Hedge phrases ("can help you", "may improve", "consider doing") where a direct claim would be stronger.
6. Closing CTAs that are generic ("learn more", "let's connect").
Don't rewrite. Just flag and quote. End with a 1–10 score for "sounds like a real person who knows things" vs. "sounds like AI" — and the single edit that would most improve the score. Prompt 12 is the one to use most. Every other prompt produces something. Prompt 12 protects you from publishing the thing the other prompts produce. Run it on every important draft. It's the difference between "AI helped me write this" (fine) and "this was clearly written by AI" (career-limiting).
The integration that pays back fastest for marketing: at the end of every quarter, Claude pulls your actual performance from HubSpot (or your email + social platforms), identifies what worked and what didn't, and proposes next quarter's content calendar — themes mapped to formats mapped to channels. Setup: ~60 minutes. Quarterly ROI: 4–6 hours saved and a calendar that's grounded in data, not vibes.
In your Brand Voice project → Connect apps → HubSpot (read access on Marketing emails, Forms, Contacts, Campaigns). If you're on a different stack — Mailchimp, ConvertKit, Beehiiv — connect what you can; for the rest, export CSVs at quarter-end and upload them.
For social performance, the direct integrations are still uneven; CSV export from LinkedIn / X / wherever you publish is fine. The point is: real numbers in, not vibes.
Inside the project, save Prompts 07 (last month), 09 (quarterly audit), and a custom combined prompt that runs them together with the next-quarter calendar request. The combined prompt is the one you'll actually use at quarter-end:
It's the end of Q[N]. Pull from HubSpot: every marketing email sent this quarter with open/click/conversion data. Pull from the CSVs I uploaded: social and blog performance.
Run the audit (Prompt 09 in this project). Then, using the audit as input:
1. Propose next quarter's content calendar as a table:
- Week / theme / source-type / channel / format / "why this week" (1 line)
- 13 weeks, one main piece per week, with brief secondary mentions
2. The themes must trace back to what worked this quarter or fill a gap the audit identified.
3. Flag any week where I'd be repeating myself or stretching too thin.
4. Identify the one experiment for next quarter — the format or channel we should test deliberately, and what success would look like.
5. End with the one thing I should kill — content I've been making out of habit that the data says isn't earning its time.
Reply in two parts: the audit first, then the calendar table. I'll review the audit before approving the calendar. Before the actual quarter-end, run this against the quarter you just finished. Three things to check:
Week 1: Theme — pricing transparency in our category. Source — founder POV (record voice memo by Friday). Channel — long blog + LinkedIn excerpt. Why this week: pricing was the top-engaging theme last quarter, and Q1 is when our buyers do budget planning.
Week 2: Theme — the "fast onboarding" myth. Source — recent customer call where they explicitly named slow onboarding as why they left their last vendor. Channel — case study (long blog) + 2 LinkedIn posts. Why: aligns with our differentiated POV; customer call is fresh.
... (continues for 13 weeks)
Kill this quarter: The monthly industry roundup. 6 of the last 8 underperformed; the format rewards reach we don't have, and the time would be better spent on the customer-call series, which has consistently outperformed everything else.
Experiment: One short-form video per month (60-90 sec, founder talking to camera, no production). Hypothesis: our differentiated POV section is being read but not heard, and the founder is the most under-leveraged voice. Success: 3 of 3 monthly videos outperform the median text post.
Last Friday of each quarter: pull the CSVs, run the prompt, review the audit, edit the proposed calendar, lock it in. What used to be a half-day of "what should we make next quarter" debate is now a quarter that begins with a real plan grounded in real data.
Crucially: edit the calendar. Don't accept it wholesale. The judgment Claude can't bring is which themes you have appetite for, which sources are realistic to capture, and which experiments you're ready to commit to. Claude proposes; you commit.
What this isn't. It isn't a content engine that auto-publishes. It's a quarterly planning loop where the analysis happens automatically and the calendar is grounded in your actual numbers. Auto-publishing comes later (L4) and only after the voice is dialed in. Plan automatically. Publish deliberately.
L4 for marketing is the famous "AI content engine." It exists. It works. And it produces generic content for almost everyone who builds it before they've earned it. Two L4 sketches worth aiming at — eventually.
Every Friday, Claude scans the week's customer-facing channels — support tickets, sales call recordings, social comments, form submissions, inbox replies — and pulls out: every question asked more than once, every objection raised in a new way, every specific phrase a customer used to describe a problem. Posts a Slack message with the top 5 raw insights and proposes one as next Monday's Source Hour starting point.
Theme of the week — "audit fatigue": Three separate customers used the phrase "audit fatigue" this week without prompting. Two were in support tickets, one was on a sales call. None of our existing content uses this phrase. Strong candidate for next Monday's source — it's a real customer term, not one we invented.
Recurring objection: Four prospects this week pushed back on our pricing tier structure — specifically, the gap between Pro and Enterprise. Pattern's been building for ~6 weeks. May be a content-explainer issue, may be a real pricing issue. Worth flagging to sales.
New phrase to consider adding to the project vocabulary: "Audit fatigue" (3x this week, 0x in our existing content).
The full version: Claude takes weekly Insight Listener output, your past hits, your differentiated POV, and your channel mix, and drafts a month of content in your voice — with the source artifact for each piece traced and named. You review, edit, approve. Distribution happens through your usual scheduler. Performance feeds back into the audit, which feeds back into the next month's drafts.
It's a real build, and it's the right destination for a small business with a serious content motion. It's also the most common L4 build that fails — because the voice in the project wasn't strong enough, the source types weren't varied enough, or the human-in-the-loop review got skipped under deadline.
The most aggressive of the three, and the one that crosses from content into outbound. A real, working pipeline that a serious marketer can build today:
{{first_name}}. Every email written from their content.Linking page scraped: "Top 10 customer onboarding tools, 2026" published by SaaSReview.io. Page mentions our top competitor as their #3 pick, in a paragraph praising their checklist-style onboarding UX.
Subject: Your top-10 onboarding piece — one omission worth a paragraph
Body: Maria — the SaaSReview piece on onboarding tools held up well; your point about checklist-vs-wizard UX is the right argument. The one omission worth flagging: nobody in your ten handles re-onboarding — when a customer's primary user changes mid-contract. It's the post-onboarding edge case that breaks every "great onboarding" claim in your roundup. Happy to send the data we have from 240 hand-offs if it'd be useful for a v2 of the piece — and I'd genuinely love to know if your readers are asking about this gap.
Why it's not template-able: Claude read their specific argument and built on it. The "re-onboarding" angle came from the prospect's own framing — not from a library of pre-written hooks.
The outcomes are real. What used to be a VA's 40-hour week is now a cron job. New competitor backlinks get scraped weekly, enriched, scored, outreached, and managed — automatically. A virtual employee working full-time on link building, for roughly the cost of the APIs.
The catch — and there's always one — is that this is the most expensive of the three L4 builds to get wrong. The failure mode is public: bad cold emails at scale damage your sending reputation, your domain, and your brand simultaneously. The components are stable; the wiring is finicky; the per-prospect Claude call quality is the entire difference between personalized outreach and the highest-volume spam your category has ever produced.
Build this last, not first. If your Source Hour isn't producing wins, this won't either — the agent doesn't fix a thin voice, it amplifies it.
If you're imagining any of these three right now — pause. The honest path is: spend two quarters at L2–L3 with the Source Hour. By the second quarter you'll know exactly which of the three fits your business, and which would be a toy. The build is then weeks of work, not a six-month project. When you're at that point, that's a good call to have.
Marketing has the highest failure rate of any role with AI, because the failures are public. Here are the patterns — named so you can see them in your own draft before your audience does.
"Write a LinkedIn post about [topic]." No source artifact, no customer phrase, no real take. The output is structurally fine and substantively nothing.
The Source Hour exists to make this impossible. If you don't have a source, don't make a piece. Skipping a week is better than publishing filler.
The project never got built, or the "voice" section is three vague adjectives ("warm, professional, accessible"). Claude defaults to the average voice of your category — which is the voice you're trying to escape.
The fix is the differentiated-POV section and 10 real past hits as files. Without those, no prompt in this library produces work that sounds like you.
10 posts a week saying nothing. The metrics look good for a quarter because reach is up. Then nothing converts and nothing repeats, because there was nothing to remember.
One strong piece per channel per week beats five mid pieces. Claude can produce mid pieces faster than ever. Don't use it for that.
The "does this sound like AI?" check. If you don't run it, you'll publish something that sounds like AI — and your audience will quietly mute you.
Make Prompt 12 a non-negotiable last step before anything important ships. Treat it the way good engineers treat tests.
The Source Hour produces content. The performance review tells you which sources work for which audience. Skip the review and the Source Hour drifts toward whatever's easiest to source — usually founder POVs, which work less often than customer calls.
Last 15 minutes of every Source Hour: last week's performance, one learning, into the project. Non-skippable.
Hooking Claude up to auto-post on a schedule before the project is mature, before Prompt 12 is muscle memory, before you've reviewed performance through two full quarters.
Auto-distribution amplifies whatever the voice is. If the voice is half-built, you're amplifying half-built marketing at scale. That's worse than not posting at all. Walk up the ladder.
Download the template. Build the project. Save the prompts. Run the Source Hour for a quarter. Then run the audit. You'll have ninety days of content grounded in real sources, in your voice — and a calendar for the next ninety, grounded in numbers.