Just stumbled upon this feature while trying to figure out if Sudowrite is actually good for anything beyond generating mediocre prose. You know, the kind of thing that should be in the first five minutes of the tutorial, but somehow I missed it.
Select a sentence, right-click, and you get options like "Add more tension" or "Add more detail." Not just the usual "Rewrite" or "Expand." It’s… weirdly specific. And weirdly effective, for about two uses. Tried it on a bland line about a character walking into a room. "More tension" gave me a paragraph about creaking floorboards and a cold draft. Fine. Then I tried it again on the new text, and it started looping back on itself, adding sweat and pounding hearts until it read like a parody of a thriller.
Makes you wonder who this is for. Is the ideal user someone who can write a serviceable sentence but has no idea how to escalate a scene? Or is it just another marketing bullet point—"200+ AI-powered writing actions!"—that works well enough in a case study but falls apart under actual workflow pressure? I can see it being a decent crutch for a first draft, but the output still has that generic AI sheen. The "more detail" usually means more adjectives, not more meaningful, plot-relevant texture.
Has anyone else actually integrated this into a real process, or did we all just play with it for ten minutes and go back to arguing about whether it’s better than a basic GPT wrapper? The specificity is interesting, but specificity doesn’t equal depth.
cg
cg
The "two uses" limit is the real tell. It's an autoregressive model trying to guess the next most dramatic adjective, not a tool for building narrative structure. I've seen the same pattern in logging aggregators that offer "make this error more severe" filters - they just tack on "CRITICAL" and a timestamp, eventually generating alert storms that mean nothing.
You get creaking floorboards, then sweaty palms, then a pounding heart, but it never suggests the character should, say, notice the door was already unlocked. That's the difference between surface-level tension and actual plot. It's a decent first-draft crutch, as you said, but it optimizes for the wrong metric.
Your fancy demo doesn't scale.
Exactly. It's optimizing for local intensity, not global coherence. Reminds me of auto-scaling rules that only look at CPU - you get a flood of new instances for a spike, but they're all idle ten seconds later because the metric is shallow.
Your logging example is spot on. We built a "severity enhancer" for CloudWatch once and killed it after a week. It just learned to append "URGENT: " and inflate all the numbers, which made actual root cause analysis impossible. The noise drowned the signal.
These tools need a cost function that penalizes repetition and rewards new information. Otherwise it's just dramatic inflation.
That "dramatic inflation" comparison hits hard. So it's basically like turning up the volume on one instrument until the whole song is just noise? 😅
I'm still new to this, but the auto-scaling analogy makes sense. What would a "cost function" for writing even look like? Is there a way to measure "new information" vs. just longer, scarier words?