Automating my writing pipeline

Last Friday at 4:58pm I was staring at my calendar from the week. Forty-one meetings. Forty-one transcripts sitting in a folder on my Mac, full of conversations I'd already forgotten. Somewhere in there are stories worth telling. Product decisions I'd made. Mistakes I'd learned from. Funny things my coworkers said that perfectly captured a bigger idea.

I was never going to go back and read forty-one transcripts. Nobody would.

So I built a thing.

what it actually does

Every Friday at 5pm, a Python script wakes up on my computer and grabs the week's meeting transcripts that are sent to a local folder on my computer. I keep everything local for security. It sends them to Claude's API with a prompt I wrote that knows what to look for. Things like real problems we solved, decisions that surprised me, mistakes worth sharing, moments where I said something off-the-cuff that was actually good. It also knows what to skip. None of this information ever leaves my machine.

From forty-one transcripts this week, it pulled out three strong story ideas, six smaller riffs, and seven direct quotes from meetings that I'd completely forgotten saying.

Then it picks the best story and writes me a shitty first draft. (Anne Lamott would be proud.) The draft uses a voice profile I'd already built, a detailed file that captures how I actually write, down to my tendency toward parenthetical asides and the specific corporate phrases I never want to use.

The draft lands in a folder. When I'm ready, I type /push-draft in Claude Code and it pushes to Craft, where I do all my actual writing and editing.

the part I keep thinking about

This isn't "AI writing for me." I still have to read the draft, react to it, rewrite most of it, add the parts only I know. The first draft it generated this week was about sunsetting a product. It had the bones of the story right because it pulled from a real conversation I had, but it flagged gaps with tags like [NEED: detail about timeline] and [CHECK WITH: X before sharing this]. It doesn't make stuff up. It tells me where it got stuck.

What it actually did was solve my blank page problem. I have forty-one meetings a week and zero time to go back through them looking for writing material. The AI found stories I'd already forgotten. It gave me something to react to instead of nothing.

That's a different thing than "AI writing." It's more like having a research assistant who sits in all your meetings, takes notes, and says "hey, that thing you said Tuesday at 2pm? That's a blog post."

how I built it (this is the fun part)

I built the whole thing in a single Claude Code session on a Saturday. I described what I wanted, and we iterated back and forth until it worked. The script is Python. I did not write the Python. I'm a product person. I described the pipeline I wanted, handled the edge cases as they came up, and Claude Code wrote the code.

The sneaky thing is that most of the hard work was already done. I have my transcripts backing up every night (already automated). I have my writing voice profile already built. I had storytelling frameworks and a story catalog in my repo. The content mining script just connected pieces that already existed.

We hit one real snag. The script sends about 100K tokens of transcripts to the API, and my tier caps at 30K tokens per minute. So we added a 90-second pause between the mining stage and the drafting stage. It's not elegant. It works fine.

the actual lesson

I keep seeing people theorize about AI workflows. What if we could do this, what if we built that. I spent a Friday afternoon just building one. It's imperfect. The rate-limit workaround is a sleep timer, not a retry loop. The voice matching is good but not perfect (I still rewrite a lot). The whole thing might break if my transcript setup changes their export format.

I don't care. It runs. It found stories I wouldn't have found. It gave me drafts to react to. I'll make it better as I go.

If you're a non-technical person wondering whether you can build stuff like this the answer is yes you can. I described what I wanted to a tool that writes code. The skill I brought was knowing what pipeline I needed, what inputs I had available, and what output would actually be useful to me. That's product thinking, not engineering.

Bias to action. Build the messy version. Ship it on a Saturday. Iterate next week 🙃