How Leon Furze turned his book into 20 launch graphics in 16 minutes, using only his own words

The problem

Furze writes, “I don’t like using GenAI to produce writing, because not only is that my job, I actually quite enjoy it.” He says “I would not recommend using Claude or any other GenAI platform for writing marketing copy, emails, or social media posts.” His reason is that “People will notice, they will judge you, and you will lose trust.”

He had LinkedIn and his mailing list. He did not use Facebook or Instagram, and he says those platforms “still have a large audience of K-12 and University educators: the target audience for the book.” He wanted a campaign where “all of the words in the campaign across platforms including email, LinkedIn, Facebook, and Instagram are my own.”

The blueprint (as he describes it)

Set hard rules on text. He gave Claude three rules:
- “Any text in pull quotes or captions must come verbatim from the book, my blog, or previous emails.”
- “Any additional text should be written by me, directly into the prompt”
- “Claude can write [placeholder text] which needs to be manually edited out.”
The placeholder should describe what text goes there and not attempt a draft.
Make one graphic first. He described the image he wanted and left the method to the model. He says, “When working with graphics, it pays to be as specific as possible.” Claude Code wrote a Python script and used Pillow to build the image. They went back and forth on colour gradients and spacing.
Scale to 20. He asked for 20 more images. He also asked the model to open the typeset book PDF directly and take its own screenshots, and he pointed it to existing promotional images and the brand font pack. Claude converted the 196 page PDF to plain text and rendered candidate pages to a scratch folder. It found the brand guidelines and extracted verbatim text from the book. Midway through, it checked its own images with image recognition and fixed most of the overlapping-text and layout problems. He says this took 16 minutes from the first prompt to the final graphics.
Add more formats and change the link late. His next prompt asked for 20 more plus a few LinkedIn carousels. Partway through, he switched the link from the generic sign-up page to the campaign page, subscribe.leonfurze.com/pais2. Claude applied that change across “all of the 50+ assets.” The companion PDF, not the blog, says Claude Code also wrote a verifier for this step. The verifier was a script that checks every string destined for an image against the source text and fails the build if the string isn’t found. The PDF says it “caught Claude paraphrasing several times.”
Set up accounts in Claude on the web. In a parallel web chat, Claude walked him through reopening Facebook and Instagram and connecting Meta Business Suite. Claude recommended keeping his existing Instagram account and using an existing but unused Facebook page. While he was exploring ideas, Claude found that Buffer added a new API in May 2026.
Ask for the system before any posts. He connected the accounts to Buffer and copied the assets he liked into a new folder. He generated a Buffer API key and saved it locally. He did not paste it into the chat. His prompt to Claude Code set these terms:
- Build the calendar for all three platforms with the Buffer API.
- Use verbatim text only: “no ad-libbing!”
- Post to LinkedIn twice a week at 10pm Melbourne time. Post to Facebook and Instagram more often, “maybe 9am and 4pm Melbourne time.”
- Include only one link, the subscribe page, and never on Instagram.
- Before getting API access, produce a strategy plan, a database of posts, images and text, and a viewable HTML mock calendar.
- Ask any questions before starting.
He says he asks for the system first because otherwise Claude Opus 5 “will just rush off and start problem solving.”
Review the calendar and reject version one. The mock calendar showed problems right away. He had said content could be reused over the 3 month period. Claude read that as “reuse the exact same image 6 or more times per week”. It had also paired captions with images poorly. One caption was half a question from an activity in the book, set next to an unrelated image. He told it to spread the content out, use fewer repeats and “do a better job” of selecting text. Claude explained that it had not used image recognition and had been “relying on metadata and file names.” Claude also said it had paraphrased some lines rather than quoting them. The second version was “much improved.”
Hand over the key, schedule, then check. He gave Claude Code the API key through the macOS Keychain. Around 20 minutes later, all posts were scheduled across Facebook, Instagram and LinkedIn. He then checked every scheduled post by hand and made a couple of edits to the wording. He notes the wording was his own text from the book in every case.
Draft the email sequences through Kit. He had moved from Mailchimp to Kit earlier in the year, and he says his main audience is his existing mailing list. He uploaded Kit’s “6-month launch plan” for authors, his shorter edited version and the book into Claude on the web. Kit’s connector needed broader permissions, which meant starting a new chat. He had Claude write a handover document to carry the context into the new chat. Claude created the email sequences “as drafts, and with placeholder text.” He calls this “an administrative task, not a creative one.” The connector cannot build Kit’s visual automations, so he set up the tagging and triggers by hand in Kit. The PDF adds that Claude, asked to check the sequences, found a merge tag with no fallback and other errors.
Tidy the list. Separately, Claude Code identified 93 obvious fake accounts on his list. It also tagged and re-engaged around a thousand subscribers who hadn’t opened an email in a while.

Stated results

None stated for sign-ups, sales or reach. The posts and email sequences were scheduled or drafted, not yet run.
Process figures he gives: 20 graphics in 16 minutes, 50+ assets updated with the new link, and about 20 minutes for Claude to schedule the posts in Buffer.
Time spent: “A few hours work,” from 8:30am-10:30am Saturday morning, plus a couple of hours that afternoon and an hour on Sunday morning.
The PDF says “seventy-odd assets by the end”. The blog says 50+.

Caveats

He is marketing his own upcoming book, Practical AI Strategies 2. The post also links to his own online course.
The companion PDF says “Claude Opus 5 was given free rein to produce” it, and that the original prompts were edited for clarity. Details found only in the PDF are labelled above. These include the verifier script, the “seventy-odd assets” count and the merge-tag check.
The PDF says “Buffer’s API publishes but doesn’t report,” so engagement data has to come from elsewhere.
Every placeholder in the email drafts still needs him to write it by hand. He set up Kit’s visual automations manually.
He uses a Max plan. He says the same work “could be achieved with the cheaper Plus plan on a small scale.”
He gives no detail on how the list tidy worked: “I won’t go into the full mechanics of this in this article”.
Handles:
- LinkedIn: linkedin.com/in/leonfurze (verified 2026-10-20: linked from leonfurze.com).
- Bluesky: @leonfurze.com (verified 2026-10-20: linked from leonfurze.com).
- Instagram: @leon.furze (verified 2026-10-20: linked in the post).
- Facebook: facebook.com/people/Leon-Furze/100071091021193 (verified 2026-10-20: linked in the post).
- X: an account under his name exists but has no bio, link or posts tying it to him, so it is not tagged.
- Threads, Mastodon and TikTok: not checked.

Sources

https://leonfurze.com/2026/08/02/how-ive-used-claude-for-book-marketing/ (primary: first-person blog post, published 2026-08-02)
https://leonfurze.com/wp-content/uploads/2026/08/how-i-use-claude-for-book-marketing-1.pdf (companion step-by-step PDF, produced by Claude Opus 5 per its own note)

Want agents like these working your distribution? See the Agent Stack Teardown. Get new teardowns weekly: subscribe (double opt-in).