| name | unslop |
| description | Detect repetitive AI defaults for a domain and generate a reusable avoidance profile. Use when asked to "unslop", "de-slop", "remove AI patterns", "make less generic", generate an anti-slop profile, or analyze AI output patterns for any domain (writing, landing pages, emails, code, etc.). |
unslop
Empirically detect the repetitive defaults a model falls back on, then produce a reusable instruction file that makes future outputs less generic.
Quickstart
python3 SKILL_DIR/unslop.py --domain "blog writing"
pip install playwright && playwright install chromium
python3 SKILL_DIR/unslop.py --domain "startup SaaS landing pages" --type visual --count 20 --concurrency 3
The script uses pi -p (non-interactive mode) to generate samples and run analysis. It works with whatever model/provider pi is configured for.
Options
| Flag | Default | Description |
|---|
--domain | required | What to analyze |
--type | text | text for writing/code/prose, visual for websites/HTML |
--count | 50 | Number of samples to generate |
--concurrency | 5 | Parallel pi calls |
--model | pi default | Model to use (passed through to pi --model) |
--thinking | pi default | Thinking level: off, minimal, low, medium, high, xhigh |
--timeout | 600 | Seconds per pi call |
--analysis-timeout | 1800 | Seconds for the analysis pass |
--retries | 1 | Retries per failed call |
--output | ./unslop-output | Where to put results |
--skip-comparison | false | Skip the before/after step |
Output
Results land in ./unslop-output/:
unslop-output/
skill.md ← the reusable avoidance profile (main deliverable)
analysis.md ← detailed pattern analysis with counts
prompts.json ← generated prompt set
samples/ ← raw AI outputs
screenshots/ ← rendered pages (visual only)
before-after/ ← vanilla vs. profile-guided comparison
Prebuilt Profiles
Ready-to-use profiles are in SKILL_DIR/profiles/:
profiles/react-design.md — anti-patterns for React/SaaS landing pages
profiles/writing.md — anti-patterns for blog writing, essays, articles
These can be used immediately as pi skills or system prompt additions without running the full analysis.
Workflow
- Generate prompts — creates diverse, realistic prompts for the domain
- Run samples — generates outputs in parallel via
pi -p
- Render screenshots — captures page renders (visual only, needs Playwright)
- Analyze patterns — pi reads all samples and identifies repeated patterns
- Generate skill file — produces
skill.md with specific avoidance instructions
- Before/after comparison — runs the same prompt with and without the profile
Using the Output
The generated skill.md can be used as:
- A pi skill (copy into your skills directory)
- Prepended to any system prompt
- Added to project context files (AGENTS.md, etc.)
Review Checklist
After running, verify:
analysis.md is concrete, counted, and specific (not vague)
skill.md mostly says what to avoid, not prescribing one new stock style
- For visual runs,
before-after/after.html feels meaningfully less generic than before.html
- If the analysis is thin, rerun with more samples or a higher thinking level