A solo builder’s AI writing tool keeps work from sounding generic
A writer used to run work writing through ChatGPT or Claude, including proposals, one-pagers, and LinkedIn posts. The results were polished, but they sounded obviously AI-made and too similar to everyone else’s writing. Instead of accepting that, the writer spent a weekend and several nights building a custom tool to keep the writing closer to their own voice.
The tool uses a project-level that ranks useful information by importance, , and freshness, so old details do not crowd out the right context. It also sets clear priority rules: come before distilled facts, uploaded files, and . The system was tested on 141 real work prompts across 6 , with a GPT-5.5 judge scoring results across 8 criteria.
Before returning an answer, it creates several candidate in parallel and chooses the best one, instead of returning the first draft.
Key points
- The main problem was AI-polished writing that sounded generic and easy to spot.
- The tool keeps project-level memory so each project can use its own facts and style.
- Clear priority rules decide which information matters most in the final answer.
- The builder tested 141 real work prompts across 6 .
- The system generates several draft answers and uses a judge step to choose the best one.