Declaude: custom text rewriting rules for AI-generated content

As a technical writer who uses AI assistants daily, I find AI-generated writing about 90% helpful and 10% consistently, predictably annoying. The annoying parts are the same every time: unnecessary emojis scattered through technical documentation, overused words like "comprehensive," "robust," "leverage," and "cutting-edge" that make everything sound like marketing copy, documentation created for the sake of documentation rather than because anyone will read it, and an over-reliance on tables and bullet points where flowing prose would actually communicate better. Every AI user has their own version of this list, and the standard advice — "just tell the model not to do that" — works for a single conversation but doesn't scale across hundreds of interactions. I wanted a systematic solution, so I built Declaude.

Modular rules, consolidated into a single command

Declaude lets you maintain a modular collection of writing rules and vocabulary avoidance lists as individual markdown files in a rules directory. Want to ban emojis from technical docs? That's one file. Prefer active voice? Another file. Have a list of fifty words you never want to see in AI-generated output? A simple text file. The key insight is that writing preferences are easier to maintain, share, and reason about when they're decomposed into discrete rules rather than crammed into a single monolithic prompt. The consolidation script uses an LLM (configurable via OpenRouter) to intelligently combine all your rules into a single coherent prompt — not just concatenating them, but weaving them together so the output reads naturally and handles conflicts between rules gracefully. Each consolidation run creates a versioned output, so you can track how your preferences evolve over time. The result gets installed as a /declaude slash command in Claude Code: run it against a file or selected text, and all your writing preferences are applied in one pass.

From 80% to 95% with minimal effort

AI-generated technical docs frequently reach about 80% usability right out of the gate. The remaining 20% is where personal style, institutional voice, and professional polish live — and it's where most people either spend disproportionate time manually editing or just accept output that sounds generically AI-generated. Declaude closes most of that gap automatically. The patterns it addresses appear across all major LLMs, not just Claude: the tendency toward corporate buzzwords, the compulsive bullet-pointing, the unnecessary verbosity that pads a three-sentence explanation into a five-paragraph essay. The repo ships with my personal pet peeves as examples, but everything is fully customisable. Your writing annoyances will be different from mine, and that's the whole point. Source on GitHub.