Why AI content works (and when it doesn't)
Let me be direct about something: AI content tools don't make you a better strategist or a more interesting thinker. They make it faster to produce content once you know what to say. That distinction matters — teams that skip the strategy and lean on AI for ideas end up with faster production of content nobody wants.
What AI does exceptionally well: turning a well-structured brief into a polished first draft, generating avatar video at a fraction of production cost, maintaining consistency in voice across formats, and handling the mechanical parts of content production (formatting, transcription, thumbnail creation).
What AI still does poorly: generating original insights, understanding your specific industry context without extensive priming, and producing content that reads like it was written by someone who has actually done the thing. The strategy and the insights still have to come from you.
The full stack
Here's every tool I use in the order I use them in a typical production week, with the specific job each one does:
- Claude — Research, structuring, and first drafts. I use it with a detailed system prompt that includes my voice guide, target audience description, and examples of posts that have worked before.
- HeyGen — Avatar video generation from a script. I use the digital avatar feature with a trained likeness and voice. Output: MP4, ready for native LinkedIn upload.
- ElevenLabs — Voice cloning for audio-only content and for HeyGen scripts where I want a more natural voice delivery than the standard avatar.
- Canva — Carousel slides, thumbnail graphics, cover images. Templates set up once, updated per post in under 5 minutes.
- Descript — Video editing, transcript generation, and the "remove filler words" feature which alone saves 20+ minutes per video.
- Typefully — LinkedIn post scheduling and performance analytics. Drafts live here; I review and approve before they go out.
- Make.com — Automation glue. Connects form submissions to email sequences, new posts to notification workflows, and lead magnet downloads to CRM entry.
AI video with HeyGen
HeyGen lets you create a photorealistic avatar video from a script and a trained likeness. I trained a custom avatar using about 5 minutes of clean video footage — natural light, direct to camera, no background noise.
The workflow: write the script in Claude → paste into HeyGen → select avatar and voice → generate → download MP4 → light edit in Descript (trim, captions) → upload natively to LinkedIn.
Total time per 60–90 second video: roughly 45 minutes from blank page to ready-to-upload. Compare that to booking a camera operator, filming, editing — we're talking 3–4 hours compressed into under an hour.
The honest limitation: HeyGen avatars are visibly AI-generated if you look closely. This matters for authentic personal brand content. It matters less for branded explainer content or educational posts where the value is in the information, not the presenter. Know which you're making before you commit to the format.
Best use cases for HeyGen: explainer videos, framework walkthroughs, product demonstrations, content that would be too expensive to film professionally but too valuable to skip.
Voice cloning with ElevenLabs
ElevenLabs builds a voice clone from a few minutes of clean audio. I use it in two ways: as an alternative voice for HeyGen videos (more natural delivery for some scripts) and for podcast-style audio clips.
The setup: record 3–5 minutes of natural speech reading from a script (varied sentences, natural pauses) → upload to ElevenLabs → train the voice model → use via API or the web interface.
The quality is genuinely impressive — the cloned voice inflects naturally on sentences it has never seen before. The main tell is unusual words or names it hasn't been trained on, where it sometimes mispronounces.
Writing workflow with Claude
Claude is the workhorse of my written content production. Here is the exact system I use:
The system prompt
I maintain a standing system prompt that includes: my writing voice guide (direct, first-person, no jargon, no listicle padding), 10 examples of posts that performed well, the target audience description, and a list of topics to avoid (generic advice that could apply to anyone).
The brief format
Every content request includes: the specific question the post is answering, the one insight I want the reader to take away, any personal experience or data point I want included, and the target format (text post, carousel, article).
The editing pass
I never post Claude's output without editing. The edit pass has two jobs: (1) inject specific details and first-person experience that AI can't invent, and (2) remove anything that sounds like it was written by software rather than a person. Telltale AI phrases I always cut: "it's worth noting that," "in today's fast-paced world," "let's dive in," anything that opens with "Absolutely!"
Automating the distribution with Make.com
Make.com is the automation layer that connects everything. The workflows I run:
- Lead magnet → CRM: when someone fills out the ebook download form, Make.com adds them to the email list, sends the welcome email with the download link, and creates a contact record in HubSpot.
- New blog post → social notification: when I publish a new article, Make.com sends me a Slack notification with a pre-written LinkedIn post draft to review and schedule.
- Form submission → calendar booking: contact form submissions that include a "yes, I'd like a call" trigger a Calendly invite automatically.
Make.com is significantly cheaper than Zapier for the same functionality. The learning curve is slightly steeper but the visual workflow builder is excellent.
The full weekly workflow
Here's a typical production week, in time:
- Monday morning (2 hours): Review the week's content calendar. Write briefs for all three LinkedIn posts. Run the briefs through Claude for first drafts. Edit all three. Schedule in Typefully.
- Tuesday (1 hour): Produce one video. Script → HeyGen → Descript edit → upload to LinkedIn scheduler.
- Wednesday (30 min): Write the week's blog post from the content calendar. Brief → Claude draft → edit → publish.
- Thursday (30 min): Review analytics from the week prior. Note which posts over- or under-performed. Update the content calendar accordingly.
Total: roughly 4 hours per week to produce 3 LinkedIn posts, 1 video, and 1 blog article. At agency rates, that's a £2,000–3,000 per week production output for 4 hours of work.