AI-assisted writing (without the Slop) (www.jampa.dev)

The piece argues that the awkward, repetitive “sloppy” voice of many LLM outputs isn’t just bad taste—it’s structural. LLMs average across many public sources, so they often generate performative signposting, redundant recaps, and tone-mismatches that make text feel unauthentic. That matters because readers quickly judge whether an author actually thought about the content; machine‑like prose erodes credibility. The author’s key insight: LLMs are weak draft-makers but powerful editors, and treating them as the former produces the “slop.” Practically, the recommended workflow flips common practice: humans should produce a raw, messy first draft (or bullet-point outline) that captures ideas and personal voice, then use LLMs for staged edits. Use content editing for structure and flow (destructive; one pass), line editing for phrasing and clarity (where AI excels), and proofreading for grammar (safe to run repeatedly). Prompting matters: don’t signal helplessness (“Improve my text”); instead ask the model to refine flow while preserving your words or adopt editorial-role prompts (examples provided). Technical notes: newer models with chain-of-thought reduce heavy prompt engineering because the right verbal cues steer latent-space embeddings toward desired edits. Bottom line: expect AI to speed up the 80% editing work, not replace the human draft or authentic voice.