The fight against AI-flavored writing has moved upstream. Digital Life Khazix has open-sourced a writing skill called Human Writing that works by constraining how the model writes rather than patching its output, asking users to supply real cases and emotions and banning the verbal tics that mark machine prose.
The project is positioned as a general-purpose writing skill, not a detector bypass. It claims compatibility with Qwen 3.8 Max, DeepSeek V4 Pro, Kimi K3, and others, and it can be used directly in products like WorkBuddy.
What the Skill Actually Does
The design is interesting because it inverts the usual approach. Most anti-AI-detection tools work on the output, rewriting text after the model has generated it. This skill works on the input and the rules, constraining the generation process itself.
The mechanism has two parts. First, it pushes the user to provide concrete personal material, real cases and real emotions, because specificity is the strongest anti-flavor signal a text can carry. Second, it constrains the word and phrase level, banning the verbal habits that mark AI text.
The ban list matters more than it sounds. There is a recognizable set of phrases and sentence patterns that recur across models, and removing them at the instruction level is a structural fix rather than a surface patch.
The Honest Analysis
The evidence for this approach is a mixture of promising and limited. The specificity requirement is genuinely sound, concrete details and personal voice are what make text feel human, and no detector can flag what a person actually experienced.
The ban-list approach is weaker than it looks. Phrase lists are static while models and their outputs evolve. A skill that removes today’s tics will age, and the tics it misses today are the ones that will out it tomorrow.
The deeper limitation is that this is still a recipe, not a voice. It can remove the markers of AI text, but it cannot generate the judgment, taste, and lived detail that make writing feel authored. The skill’s own design admits this, it asks the user to supply the material precisely because the model cannot.
What It Means
The skill is best understood as a workflow improvement for people who want to write more naturally with AI help, not as a machine that converts AI output into human-sounding text.
For writers, the practical takeaway is the input side: the more real material you feed the model, the less it can fall back on generic AI patterns. For anyone evaluating the tool, the fair test is whether your own voice comes through in the output, not whether a detector fails to flag it.
The project is also a useful data point about the ecosystem. As AI detection improves, the response is moving upstream, from patching output to constraining generation. Expect more tools in this direction.