cd /news/large-language-models/i-tracked-every-async-writing-block-… · home topics large-language-models article
[ARTICLE · art-33631] src=dev.to ↗ pub= topic=large-language-models verified=true sentiment=· neutral

I tracked every async writing block for 90 days. Output nearly doubled — but not how I expected.

A developer tracked writing output over 90 days and found that switching from Obsidian-first drafting to Claude-first drafting nearly doubled usable words per 90-minute block, from 420 to 780. However, the AI drafting introduced a failure mode where Claude produced fluent but factually incorrect prose that could reverse the author's own opinion, requiring quality control that doesn't show up in word counts.

read2 min views1 publishedJun 19, 2026

Switching from Obsidian-first drafting to Claude-first drafting nearly doubled my usable words per 90-minute block: from ~420 to ~780. That number surprised me. The failure mode surprised me more.

For the first six weeks I drafted in Obsidian — vault, MOCs, Templater scripts, the whole setup. The graph view is genuinely beautiful. It's also a distraction engine when you need 800 words out before a client call. I logged every block with Toggl and checked output against a plain Notion table. My honest average across three weeks of tracking: 420 usable words per block. The rest of the time disappeared into relinking notes, reformatting, and following threads that felt like thinking but were closer to procrastination with good aesthetics. Weeks seven through twelve I switched. I'd paste 3-5 bullet points from my daily note into Claude Sonnet and ask for a rough 600-word first-pass draft — not for publication, just a working surface. Output was usually 70% there in about four minutes. Wrong emphasis sometimes, occasionally a phrase I'd never actually use, but editable. Editing is a cognitively different task than staring at a blank file, and that difference compounded across six weeks into roughly 780 usable words per block. The failure mode, though: Claude produces fluent, confident prose that can quietly get your own opinion backwards. Twice I nearly published a draft that argued the opposite side of a nuanced take I actually hold. Technically accurate. Tonally close. Factually wrong about what I think. That quality control cost doesn't show up in a word count.

The framing that pits PKM tools against AI drafting tools is the wrong lens. They're not competing — they're doing different jobs at different points in the same block. The real variable is workflow shape, and which shape you use has measurable consequences on what actually ships by end of day.

I wrote up the full breakdown — including the side-by-side failure mode table, the CRM project that ate three weeks, and why the "second brain" framing oversells the drafting phase — over on dailyfocusmag.

── more in #large-language-models 4 stories · sorted by recency
── more on @obsidian 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/i-tracked-every-asyn…] indexed:0 read:2min 2026-06-19 ·