Eric Ries's custom AI writing tool, Solve It, structures a piece of writing as a single sequential document mixing three message types: green markdown notes (human prose), orange/red LLM prompts, and blue Python code — with the key constraint that the model sees only the messages preceding it in the document, not the whole project at once. This makes context an explicitly authored, ordered artifact rather than an automatically-accumulated blob.
Ries pairs the tool with a discipline for preserving his own voice: he mostly avoids using AI-generated prose directly. He writes a draft himself, asks the AI only to critique it, then reinserts his own final wording back into the document as if it were the AI's own suggestion. Because the model's in-context sense of "good writing" is shaped entirely by what's already in the preceding document, feeding it his own edited language — rather than letting AI-generated prose accumulate — steers its future suggestions back toward his voice instead of generic "AI-sounding" output. As he puts it: "You can't just believe everything it says. You have to be very careful with this feedback cuz it can ruin your authorial voice."
In an AI-assisted writing or coding workflow, manually curate exactly what context the model sees rather than letting it accumulate automatically; when output needs to sound like you, edit it yourself and feed the edited version back in as if AI-generated, so the model's local sense of "good" converges on your voice over a long session.
A concrete, document-level implementation of Context Engineering applied specifically to preserving authorial voice in long AI-assisted writing sessions.