Testing the Anti-Slop Prompt: Making Claude Reject Fake Precision and Boilerplate Endings
eyishazyer · x · 2026-09-17
In Episode 6 of his Claude series, the author stress-tests DeRonin's "Anti-Slop Protocol" prompt — a combined negative/positive constraint set with mandatory self-audit that bans common AI writing tics.
What worked:
- On the same weekend-recap draft: without the prompt, Claude rounded to "about 50 hours"; with it, the draft admitted "I didn't track the number closely enough to give you a real one" — killing fake precision
- Without the prompt it closed with "I wouldn't change a thing"; with it, the draft refused to resolve and ended on an unresolved question
What didn't:
- With the prompt, output got longer, not shorter, and picked up bracketed asides and one-line fragments that would look broken in a LinkedIn post or client email
Verdict: the protocol effectively strips generic AI tics, but has layout and length side effects that need per-channel tuning.
More from Apps
- Fitness app bets real money on your 10,000 steps — losers fund the winners — anthara_ai · 2026-09-17
- Templafy launches MCP to turn ChatGPT and Claude chats into branded, compliant Office docs — thetripathi58 · 2026-09-17
- Ranking in AI search hinges on tone: models favor citable, non-promotional third-party consensus pages — nikvassev · 2026-09-17
- Looking for a Local Model on M1 Mac to Batch-Convert Images to 9:16 Portrait — gelembjuk · 2026-09-17
- Indie dev launches WTMemory, a $9-forever Mac cleaner that hunts down local LLM model hoards — saibharadwaj · 2026-09-17
- 3 free AI prompt libraries: Meigen, YouMind and PromptHero offer millions to copy — aitrendz_xyz · 2026-09-17