$ show mshumer/unslop
unslop is a CLI tool that empirically detects repetitive defaults in AI-generated outputs (text or visual) and produces a reusable instruction file (skill.md) to eliminate those patterns. It works by generating a set of prompts for a given domain, running them through Claude Code to produce samples, analyzing the samples for repeated patterns (phrases, structures, visual defaults), and writing a focused 'unslop profile' that tells the model what to avoid. The tool supports both text domains (blog writing, emails, tutorials) and visual domains (landing pages, UI mockups) with optional screenshot rendering via Playwright. The architecture is a single Python script (unslop.py) that orchestrates a multi-step pipeline: prompt generation, sample generation via parallel Claude Code calls, screenshot rendering (visual only), pattern analysis, skill file generation, and a before/after comparison. It uses asyncio for concurrency, a custom terminal UI with live spinners and progress bars, and a robust subprocess wrapper for the Claude CLI that handles streaming JSON output, timeouts, retries, and permission denials. Key entry points are the CLI arguments (--domain, --type, --count, --concurrency, etc.) and the main async run() function. External dependencies are the Claude Code CLI (claude command), Python 3.10+, and optionally Playwright for visual domains. The output is a directory (unslop-output/) containing the skill file, analysis, prompts, samples, screenshots, and before/after comparison. The code demonstrates solid async patterns, clean separation of concerns within a single file, and a well-designed terminal UI abstraction. It is not a library but a self-contained tool that could be adapted for other AI model backends by replacing the claude() function.
The LLM identified 2 files worth highlighting for reuse:
A self-contained terminal UI class that provides styled output, progress bars, live spinners, and formatted logging. It handles TTY detection, color support, terminal width, and async animation via asyncio tasks. The class is fully isolated (only uses standard library modules: sys, os, shutil, time, asyncio) and can be extracted into any CLI project to add polished terminal output with minimal effort. It demonstrates non-obvious skill in handling edge cases like NO_COLOR, dumb terminals, and live animation cleanup.
A robust async wrapper for running the Claude Code CLI with streaming JSON output, timeout handling, retry logic, and permission denial detection. The claude() function handles subprocess creation, real-time JSON packet parsing from stdout, and graceful timeout via SIGKILL. The claude_batch() function provides bounded concurrency with a semaphore, retries with exponential backoff, and progress reporting. These components are tightly coupled to Claude Code's output format but demonstrate a reusable pattern for wrapping any CLI tool that emits newline-delimited JSON. The pattern can be extracted and adapted for other AI backends or subprocess-based services.