Ptacek 2026 post codifies LLM copyedit workflow with zero-phrase and zero-praise constraints
The post supplies a verifiable two-rule protocol that treats LLMs as detectors rather than generators. It aligns with empirical findings on LLM text detectability and reduces homogenization risk. The approach is immediately testable in professional writing pipelines.
The method splits writing into two phases. Authors draft without assistance. Models then receive the draft under instructions that forbid praise and prohibit reuse of any generated phrase. This blocks the headline register and sycophantic reinforcement that frontier models apply by default.
Detection literature shows readers identify LLM passages at rates above chance even after light editing. Studies on GPT output markers and non-native English penalties confirm that model-preferred collocations survive human post-editing and trigger recognition. Ptacek's rules target exactly those surface statistics.
Operationally the workflow replaces line editing labor with exhaustive mechanical review. It preserves paragraph-level revision decisions that constitute author voice while outsourcing only fatigue-prone pattern matching. Adoption requires prompt engineering that explicitly demotes encouragement tokens and logs rejected suggestions for audit.
Future toolchains will embed these constraints as default modes in writing assistants. Version control diffs will track phrase provenance to enforce the ban automatically.
Writing tool vendors: 40% of new assistant releases before 2027 will ship explicit no-encouragement and no-phrase-reuse toggles by default.
Sources (3)
- [1]Primary Source(https://sockpuppet.org/blog/2026/09/17/how-to-write-with-an-llm/)
- [2]Supporting Source(https://arxiv.org/abs/2303.11156)
- [3]Supporting Source(https://aclanthology.org/2023.findings-emnlp.123/)