Show HN: Bullshit Remover

Original Article ↗ Hacker News Discussion ↗

What the tool does

  • Web app that “translates” corporate/academic/political/buzzword-heavy text into short, plain, often profane summaries.
  • Core behavior: remove fluff, be funny and sarcastic, shorten text, keep language, and return only the rewritten text.
  • Implemented as a very thin wrapper around an Anthropic Claude 3.5 Sonnet API with a short prompt.

Quality of output & user reactions

  • Many users say it works “shockingly well,” often funnier and clearer than expected across:
    • Corporate PR, cookie banners, hiring rejections, VC/finance copy, political statements, tech changelogs, academic theory, news articles, product marketing, DEI statements, philosophical blog posts, and even poetry and Shakespeare.
  • Multiple examples where it captures the underlying meaning succinctly (e.g., “we take security seriously” → “we got hacked last week”).
  • Some users find it too snarky or disrespectful to serious/neutral texts, or say it “barely works” on clearly written content like HN guidelines.

Prompt, model behavior & prompt injection

  • Users successfully extracted the effective system prompt and confirmed it uses Claude and Anthropic’s APIs.
  • Prompt-injection attempts can bypass or alter behavior; the system will reveal its own instructions or drop the “no-bullshit” persona when cleverly asked.
  • It can handle encoded text (base64, hex), but not all encodings (base32 sometimes rejected as “bullshit”).

Biases, tone, and limitations

  • Strongly cynical, negative, and often profane by design; some see this as a “negativity bias.”
  • Tends to treat conspiracy theories and some political claims as definitely false rather than agnostic.
  • Shows political and personality judgments (e.g., on public figures) that some agree with and others flag as bias.
  • Sometimes collapses nuanced or already-clear text into unhelpful snark.
  • Technical issues: “failed to fetch” errors, rate limits, slow or hanging responses, and a “text too long” limit.

Suggested improvements & extensions

  • Requests for:
    • Browser/Chrome extension and Outlook/email add‑in.
    • A “serious mode” with less sarcasm for product pages, news, and LinkedIn.
    • Audio I/O for live political speech, TV news, or on‑the‑fly captioning.
  • Some want it open-sourced; others point out it’s essentially a minimal LLM wrapper.

Meta-discussion about LLM wrappers

  • Several comments note this is “just” an LLM API wrapper with a clever prompt, emblematic of many current AI startups.
  • Others argue that even a thin wrapper can be valuable if the framing and UX are right, likening it to a good “frame” around a powerful but generic model.