Why I Had to Automate This
Creating content with ADHD is a specific kind of hell. The ideas are there, more than enough of them. The problem is the gap between "idea in my head at 7am" and "finished post published." That gap is where everything dies.
Sitting down to write from scratch is task initiation. For ADHD brains, initiation is hard. The blank page is the worst possible starting point.
So I built a pipeline that means I almost never start from blank. By the time I sit down to "write," most of the work is already done. Here's exactly how it works.
The Full Pipeline
Voice memo → transcription → structured draft → edited post/caption/script
Each stage has a tool. Each handoff is defined. Nothing lives in my head longer than it has to.
Stage 1: Voice Memo
The starting point is always a voice memo. Not notes, not a draft. A voice memo.
I use my phone's native Voice Memos app. When an idea hits, driving, walking, mid-shower thoughts captured on the way out, I record it immediately. Unfiltered. Sometimes it's 90 seconds of connected thinking. Sometimes it's 20 seconds of a half-formed sentence and a couple of keywords.
The rule is: capture first, quality second. The worst voice memo is better than the best idea I didn't record.
I don't transcribe or process in the moment. That creates friction and usually derails whatever I was doing. The memo just goes into a folder called "Raw Ideas."
Tool: Apple Voice Memos (any will do, the tool doesn't matter here)
Stage 2: Transcription
Once a week, usually Sunday or Monday morning, I batch-process everything in the Raw Ideas folder.
I use Whisper (via a local setup) or Otter.ai for longer recordings. Whisper is free, runs locally, and handles my accent and filler words better than most cloud options. For quick memos I sometimes just use the built-in transcription on iOS 17+. It's genuinely good now.
The output is raw transcripts. They're messy: "um," false starts, half-sentences, the occasional tangent about something completely unrelated. I don't clean them.
What actually matters at this stage: accuracy on key terms and names. I'll spot-check for any technical words the model garbled. Everything else gets sorted in the next stage.
Tool: Whisper (local, via whisper.cpp or the Python package), Otter.ai, or iOS transcription
Stage 3: Structured Draft via Claude
This is where the AI does the heavy lifting.
I paste the raw transcript into Claude with a prompt. The prompt varies by content type, but here's my base version:
"Here's a raw voice memo transcript. I want to turn this into a [LinkedIn post / blog article / short-form video script]. The audience is [description]. My tone is direct and practical, no fluff, no motivational filler.
First: identify the core idea and any supporting points. Then: give me a structured draft.
Don't add ideas that aren't in the transcript. If something's unclear, flag it rather than filling in the gap."
That last instruction matters. I've had Claude "helpfully" add context I didn't say and didn't mean. The flag-it-instead approach keeps the output honest to the original idea.
The output is a structured draft: intro, body, close, with rough headings or section breaks depending on the format. For a 700-word blog post, this takes about 40 seconds.
For Reels/short video scripts, I use a variation that asks for a hook line, 3–5 spoken bullet points, and a call-to-action. The pacing and spoken rhythm are different from written content, and the prompt accounts for that.
Tool: Claude (claude.ai or API)
Stage 4: Human Edit
This is the stage I don't automate. And I'm not going to.
The draft Claude gives me is structurally sound but not quite my voice. There's usually:
- A word or phrase that's slightly too corporate
- A transition that I'd never say out loud
- An example that's accurate but not specific enough to be useful
- The occasional statement that's technically true but not how I actually think about it
I go through and fix those. On a 700-word post this takes maybe 15–20 minutes. On a shorter LinkedIn post, 5–8 minutes.
I also make the call on whether the core idea is actually worth publishing. Claude can't tell you if something's interesting. It can only structure what you've given it. If the original voice memo was a mediocre idea, the draft will be a well-structured mediocre idea. That judgment is still mine.
What the Pipeline Actually Looks Like in Numbers
On a typical week I produce:
- 2–3 LinkedIn posts
- 1 longer-form blog post or newsletter
- 1–2 short video scripts
Pre-automation time: 8–10 hours of writing and editing per week, most of it spent staring at blank pages.
Post-automation time: Around 2–3 hours, mostly editing and the occasional voice memo batch session. The transcription and drafting steps run in the background.
That's real. Not a marketing estimate.
What AI Still Can't Do Here
A few things I've tried to automate and stopped:
Publishing and scheduling. I tried automating this end-to-end. The posts went out on time but I lost the last-minute quality check that catches embarrassing errors. Now I still hit publish myself.
Deciding what to post. I experimented with asking Claude to help pick from my raw ideas backlog. It's okay at ranking by topic variety, but it doesn't know what's timely, what I've already covered recently, or what's going to land with my specific audience this week. That's editorial judgment I've kept.
The hook. Claude drafts hooks, and sometimes they're genuinely good. But I almost always rewrite the first line. The hook is the difference between someone reading the whole thing and someone scrolling past. I'm too precious about it to leave it entirely to a model.
How to Steal This For Yourself
You don't need my exact setup. The underlying structure is what matters:
- Remove the blank-page start: always begin with raw captured material
- Use AI to do the structural work, not the creative judgment
- Keep the human edit step: it's where your voice comes back in
- Batch the low-value steps (transcription, drafting) so you're not context-switching mid-creative-session
The pipeline took about two weeks to get right. The first few drafts from Claude were off: too formal, wrong length, wrong framing. I iterated on the prompts until the outputs needed less editing. Now it's close enough on the first pass that editing is fast.
Build it once. Then let it run.