Reliable LLM output style/conduct enforcer
Problem
Users cannot reliably control how LLM coding assistants talk to them: AGENTS.md/claude.md style rules are consistently ignored as sessions drag on, and models develop grating tics (fake epiphanies, self-praise, words like 'carries', 'holds', 'spells'). The problem is bad enough that a community tool that pipes Claude output through a second LLM to clean it up reached the HN front page with 305 points, and commenters report hand-writing rules to 'ban' specific words with only partial success.
Opportunity
A proxy/middleware (or editor plugin) that deterministically intercepts and rewrites the assistant's output according to enforced style and conduct rules, working with any model and CLI, without relying on prompts the model ignores. The 305-point reaction to a rough community hack shows validated demand for a polished product.
Market analysis
Demand is strongly validated — two separate community tools hit the HN front page within weeks — but free open-source solutions and Anthropic's own hooks/output-style features constrain what a paid product can be.
Market · Heavy CLI-agent users (Claude Code, Codex) annoyed by model tics; vocal niche with proven willingness to hack together their own fixes.
Pricing · Competing tools are free OSS (Vomit, Claudette); a polished product could plausibly charge $5-10/month, comparable to small developer utilities.
Pros
- + Exceptionally validated demand: Vomit (~285-305 pts) and Claudette (364 pts) both front-paged on HN.
- + Deterministic rule-based rewriting (no second LLM) would be faster, cheaper and more predictable than existing hacks.
- + Model-agnostic proxy position survives switching between vendors' models.
Cons
- − Anthropic is actively shipping in this area (output styles, hooks, plugins), so the platform can absorb the feature natively.
- − Free OSS alternatives already exist and are good enough for most sufferers.
- − LLM-based rewriting risks distorting technical content; rule-based rewriting struggles with nuance — both are hard to get right.
Existing / similar tools
Source
Hacker News (thread: Vomit: Clean up Claude 5's token output with a separate LLM)
The non-obvious risk is that the output stream is not the only channel that needs cleaning: tool-call arguments, commit messages and code comments all carry the same tics, and a proxy that only rewrites prose leaves the worst offenders intact. The defensible version of this idea is a deterministic, local, rule-based filter with an LLM fallback for ambiguous cases — Vomit itself admits its local LLM “can hallucinate a bit” because it only sees the output text. Whoever makes style enforcement fast, exact and transparent (showing what was changed and why) can beat both the prompt-based approach and the vibe-coded incumbents.