Tutorials & guides
Two Defensive Copyediting Rules to Prevent LLMs from Degrading Your Writing
Security veteran Thomas Ptacek details a repeatable workflow for utilizing LLMs as objective copyeditors rather than automated ghostwriters. By refusing AI-generated phrasing and disabling model sycophancy, developers preserve distinct voice while stripping structural clutter.
September 18, 2026 5 min read
Curated by Oleksandr Kuzmenko, AI Product EngineerUpdated September 18, 2026Sources cited on every story
AI-assisted · editor-reviewedHow we use AI
Impact: Medium
Why it matters
You can immediately apply these two system-prompt constraints to catch structural flaws and filler words in docs without letting AI turn your prose into generic corporate filler.
TL;DR
- 01Never adopt specific phrases generated by the LLM into your draft.
- 02Explicitly instruct the model to withhold all praise and reassurance.
- 03Use CLI-based LLM passes to run mechanical linting for adverbs and weak verbs.
The Two Operational Rules 1. Zero LLM Phrasing Adoption: Never paste phrases suggested by the model into your final draft. Frontier models default to glossy magazine headlines that erode natural tone. Use the model only to highlight where the prose is confusing or sluggish, then rewrite the sentence in your own voice. 2. Enforce Zero-Sycophancy Review: Default model personas aggressively praise early drafts, encouraging authors to skip structural revisions. Forbid praise in the prompt or instruct the agent that it is a strict copyeditor reviewing third-party student submissions. ### The Mechanical Editing Architecture Treat prose editing like static analysis in programming. Ptacek recommends building passes modeled after Joseph Williams' *Style: Lessons in Clarity and Grace*: - Flag nominalized verbs and buried actions. - Strip crutch modifiers like very, actually, really, and unfortunately. - Pinpoint paragraph dislocations to improve narrative flow. Run revisions through independent context windows so the evaluating model has zero memory of previous drafts and cannot validate your bias.
Try it in 2 minutes
You are a strict copyeditor. Analyze this draft. Point out: 1) overused passive voice and nominalized verbs, 2) unnecessary adverbs (very, really, actually), 3) weak transitions. Do NOT praise the draft. Do NOT suggest replacement sentences.text
✓ When to use
- Editing technical blog posts, documentation, or design RFCs.
- Refining complex engineering arguments without slipping into generic marketing tone.
- Building local script pipelines for document verification using LLM CLIs.
What to do today
- Update your writing system prompt: 'Act as a ruthless copyeditor. Do not praise my draft, identify flaws only.'
- Scan drafts for filler words by piping text through an LLM CLI checking specifically for 'actually', 'very', and 'really'.
- Review suggested friction points and rewrite the sentences manually without copying AI wording.
What the community says
“Those rethinks are load-bearing parts of your voice. I really hope this was intentional.”
“Thought I'd get an article with tricks how to stop an LLM from sounding like the most obnoxious Linkedin influencer. Instead it says to use it as a copyeditor. Ok - but that's not very useful.”
Sources