Prompt vs Verify: Two Layers of Anti-AI Writing That Actually Work

Last week, my blog article pipeline failed completely. I thought pasting a 100KB skill containing hundreds of style rules into the prompt would solve the anti-AI writing problem. Turns out, that was a huge mistake. The prompt got truncated, token quota was wasted, and the final result was actually detected as machine-generated by local testing tools. The problem wasn't the number of instructions. The root cause was the assumption that a prompt could perform quantitative verification. Since then, I stopped forcing language models to be detectors. I switched to a two-layer architecture: a slim qualitative prompt, locked with a 0-token verify regex at the final stage.

The Illusion of Giant Prompts

Many developers get stuck in this phase. I thought that adding a list of forbidden words or asking to "write like a human" would make the model instantly comply. In fact, language model architectures have limited attention spans. When I audited the execution logs, generic templates full of SEO requests actually clashed with the personal writing style contract. The result? The model ignored subtle instructions and reverted to its default probabilistic patterns. This is why simply telling it "don't sound like a robot" is never enough. I needed an approach that separates the generation task from the validation task.

Separating Generation and Validation

This is where burstiness comes in as a lifesaver metric. Human text has high variation in sentence length and structure, typically scoring between 0.5 and 1.0. Conversely, AI output tends to be flat with a score of 0.1 to 0.3 because it predicts the next token uniformly. I no longer rely on the feeling of "sounding natural". I use tools like avoid-ai-writing that break down the analysis into a Pattern Score of 70% and a Uniformity Score of 30%. This combination objectively measures term density and trigram repetition. If the cadence AI feels too monotonous, I don't revise the prompt. I directly modify the draft in the post-processing stage.

The Final Key in the 0-Token Layer

Prompts are qualitative — but verification must be quantitative. Once the draft is done, I run a verify regex script that looks for specific patterns without consuming the model's context tokens. This script checks for excessive conjunction repetition or overly perfect sentence structures. Even commercial detectors have worrying false positive rates — some educational institutions even disabled this feature due to their dubious accuracy. So, relying on a single external detection tool is naive. My opinion is firm: never entrust final validation to the same language model that generated the text. Use a rule-based standalone script. By separating generation and verification, the real anti-AI writing strategy actually keeps my writing soulful, not just statistics that pass a test.