I used to spend hours rewriting AI-drafted video scripts that sounded nothing like me. At best, I might finish one or two that were just okay, but most of the work was still on my shoulders. Now, with my new system, I can get 15 scripts done in one session. They match my writing style, my voice, and my company's knowledge. Instead of full rewrites, each script just needs a quick review. I built the system in about a week, and the difference showed up before the week was out. The first project was an internal video series to teach people about the software factory and LaunchDarkly, and I barely had to edit those scripts. For the first time, the AI handles most of the work.
My recent projects have mostly involved agentic software delivery. The software factory is where all of it was heading, and my company spent months preparing to help customers build their own. We had internal material, public documentation, and ongoing conversations I wanted to add clarity to. So my challenge wasn't just learning, it was learning while producing content at the same time. Essentially, devrel.
By the end, I had built what I like to call my own personal content factory. At first, it was separate from the software factory it describes, but over time, the line between them blurred. My approach is to let AI handle the bulk of the work, while I step in for the important decisions. This is the only way I've found to make AI content sound like me. The AI creates the drafts, and I step in at three key points: checking the voice, the facts, and the overall feel.
The quality ceiling is set before the first draft.
I didn't just ask for a video script about a topic. Instead, I gave Claude access to every source I had—public docs, internal notes, and product requirements. It used research agents to read everything at once and came back with clear, organized notes I could use.
I set two important rules for this step. First, every part of the research was labeled as either public-safe or internal-only right from the start. When your sources mix private details with public information, you need to set those boundaries before turning anything into content. External scripts only use the public-safe material, and that rule is clear from the beginning. Second, if there's no specific detail, there's no script. Skipping the research means you get scripts that sound confident but say nothing. Real details are what make a technical video worth watching, and those only come from solid sources.
Curriculum before scripts
Before writing scripts, I had Claude turn the research into a curriculum with modules, learning goals, dependencies, and self-check questions at the end of each section. This step became the backbone for the whole series—defining episode boundaries, showing how concepts build, and preventing accidental repetition.
It's also the simplest and most valuable part of the whole system. If you take away just one step, make it this one—it only costs some text.
Drafts arrive sounding like a polished stranger.
Once the curriculum was ready, Claude drafted every episode. The scripts were well-structured, accurate, and sounded polished, though not quite like me. This is where most people stop, but with some patience and effort, it's worth pushing through.
When I first asked Claude to match my style, it just added my catchphrases to its usual writing. That wasn't enough. What really worked was having it break down my writing at the sentence level—looking at my punctuation, my rhythm, and how I connect ideas. Your voice comes from sentence structure, not just word choice. So I rewrote one script myself, making sure every sentence sounded like something I'd actually say. (Reading it out loud helps; you'll hear things your eyes miss.) That finished script became the example the system uses before drafting anything new.
Once I had a finished script to use as a gold standard, I could start automating the rest of the process.
The detour that became the greatest ramp-up, binging my own series
I still didn't fully understand the topic I was writing about. So I had Claude turn the draft scripts into avatar videos using HeyGen. I connected the two systems so the script went straight to the video renderer, no extra steps needed. Then I binge-watched a course on my own subject, presented by an avatar, using my own research.
This was the best way I've ever learned new material. Reading makes you an editor, judging each sentence. Watching turns you into a learner, experiencing the teaching. Ideas I'd read several times finally stuck after just one viewing. Each episode was also a screen test. If a video lost my attention, it meant I'd written something confusing—feedback I wouldn't have caught by just reading. I was my own test audience. The scripts that passed became the ones a real presenter used.
I wouldn't do this by default again because avatar rendering gets expensive—$200 in a single day for a couple dozen videos, mostly for my own learning. Next time, I'd only render videos destined for real content and stick to text for self-study.
Polish tools optimize for the page, not for meaning.
Once a script felt like it matched my style, I ran it through Grammarly. First, I checked for cohesion, then used the AI detector, rewriter, and Humanizer tools, usually with the Precisionist voice and sometimes the Naturalist setting.
Once, Grammarly changed my phrase "this step becomes free" to "this step becomes unnecessary," which completely reversed my point. These tools make writing smoother, but they don't always get the meaning right. After each edit, reread any sentence that states a fact, and pay attention to the conjunctions it adds. Words like "since" or "therefore" can suggest a cause-and-effect you never meant.
That's the second checkpoint. Nothing gets published until the facts are correct.
Finished work becomes the standard, and nothing derives from drafts.
Every finished script became the new voice standard, so future drafts matched my rhythm, transitions, and even where I put question marks. We wrote these rules into a living style guide. Each completed script made the next drafts better. Fine-tuning wasn't about changing prompts, but about locking in finished scripts until the standard matched my voice. This happened within a week. By the back half of the series, editing took just a few minutes instead of full rewrites.
One rule guides everything after this: LinkedIn clips, one-liners, and short videos are only made from finished, human-approved scripts, never from new drafts. The voice is already set in the finished work, so repurposing keeps it. Creating new content from scratch brings back that AI-generated sound.
That's the third checkpoint. Taste is built into the process, not just decided at the end.
One surprise from running everything in a single, ongoing session was that the AI became a better continuity editor than I am. When I rewrote my series introduction late in the process, Claude caught every reference it broke—episodes that pointed to things that no longer existed. Over 14 episodes, that's a level of consistency no human could keep up with.
The parallel I can't unsee
Research starts with clear sensitivity boundaries, and planning happens before anything is built. The AI produces more than any person could alone. Human judgment comes in at specific checkpoints, not everywhere. Only finished work sets the standard for future work—nothing new is based on unfinished drafts.
Our customers' software factories work the same way. Automation does most of the work, human judgment is built into the process, and nothing is released unless a person stands behind it. I set out to make content about this world and ended up proving its value by using it myself. The factory sets the pace, and the checkpoints ensure reliability. Both are needed.
The AI didn't replace my writing. I see it as a co-intelligence that speeds up the process between drafts. It never decides what gets published—that's still my job.
Top comments (0)