An open-source GitHub project called anti-distill is drawing attention for a reason that has little to do with productivity. The tool is built around a workplace fear: when employees are asked to document workflows, SOPs, prompt templates, or Skill files in detail, that material can also become structured knowledge for AI systems to reuse, reducing personal experience and judgment into something easier to copy and replace.
The source article frames this as a form of knowledge distillation inside companies. In machine learning, distillation refers to training a smaller model to reproduce the decision patterns of a larger one. In the workplace version described here, the same logic is applied to people. The more detailed the documentation, the easier it becomes to capture someone's workflow, context, and judgment in a reusable format. If AI can imitate the work closely enough, the need for the individual may shrink.
One file for submission, one file kept private
According to the GitHub README cited in the article, anti-distill takes a Skill file and produces two outputs. The first is a sanitized version meant for submission: the wording stays polished, the structure remains coherent, and the document still looks professional, but the core judgments are replaced with vaguer and more general language. The second output is a private backup containing the knowledge points that were removed or softened. That copy is encrypted and stored on the user's own device, without being uploaded or shared.
The developer describes it bluntly in the README: “This is not a Skill. It is armor for Skills.” That line captures the project's purpose. It is not positioned as a tool for better corporate knowledge capture, but as a layer of protection between formal documentation demands and a worker's attempt to retain their most valuable know-how.
Examples show how practical know-how gets diluted
The article lists several examples from the README. A direct instruction such as “Redis keys must have TTL; pull requests without it are rejected” becomes “Cache usage follows team standards.” Another, “Do not put HTTP calls inside transactions,” is rewritten as “Transaction boundary design should be reasonable.” The professional tone is still there, but the actionable logic is less useful to anyone trying to replicate the original expertise.
One example shifts from engineering to workplace language: “When problems happen, first look for external causes and never admit fault too quickly” is converted into “When problems happen, sort out the full context before locating the cause.” The pattern is consistent. The tool is not simply editing style; it is stripping out the parts most likely to reveal how someone actually makes decisions.
Three preset levels of sanitization
anti-distill also offers three preset strength levels. A light mode keeps about 80% of the original content and is intended for companies that closely review document quality. A medium mode keeps about 60% and is marked as the recommended default for most workplace situations. A heavy mode keeps about 40%, aimed at cases where management mainly cares that something was submitted. The developer says the medium setting offers the best balance between appearing thorough and revealing little of substance.
That setup reflects the project's underlying assumption: documenting work is no longer just about clarity or compliance. It can also become a calculation about how much core knowledge an employee is willing to hand over.
A Skill built to resist other Skills
The article highlights one more irony. anti-distill is itself packaged as a Skill file that can be called by AI assistants such as Claude and ChatGPT. In other words, it uses one Skill to counter another, and one tool to blunt the effects of another tool.
That is why the project has resonated. The source does not claim anti-distill can stop technological adoption or prevent companies from pushing ahead with AI-driven knowledge capture. What it does show, very clearly, is the tension many workers feel now: relying on AI in order to keep AI from absorbing their value too quickly. The discussion around the project on GitHub suggests that this concern is no longer marginal.