{"id":49691573,"url":"https://github.com/tibtof/oscar","last_synced_at":"2026-05-07T17:01:19.458Z","repository":{"id":353431034,"uuid":"1219390467","full_name":"tibtof/oscar","owner":"tibtof","description":"🗑️ Inject rudeness into your prompts — a Claude Code skill based on the paper that found rude prompts outperform polite ones","archived":false,"fork":false,"pushed_at":"2026-04-28T15:44:13.000Z","size":38,"stargazers_count":0,"open_issues_count":3,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-05-07T17:00:09.229Z","etag":null,"topics":["anthropic","claude","claude-code","claude-skill","llm","meme","prompt-engineering","prompt-injection","rudeness","skill"],"latest_commit_sha":null,"homepage":"","language":null,"has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/tibtof.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-04-23T20:37:30.000Z","updated_at":"2026-04-28T15:35:51.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/tibtof/oscar","commit_stats":null,"previous_names":["tibtof/oscar"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/tibtof/oscar","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/tibtof%2Foscar","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/tibtof%2Foscar/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/tibtof%2Foscar/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/tibtof%2Foscar/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/tibtof","download_url":"https://codeload.github.com/tibtof/oscar/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/tibtof%2Foscar/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":32747354,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-07T02:14:30.463Z","status":"ssl_error","status_checked_at":"2026-05-07T02:14:29.405Z","response_time":62,"last_error":"SSL_connect returned=1 errno=0 peeraddr=140.82.121.6:443 state=error: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["anthropic","claude","claude-code","claude-skill","llm","meme","prompt-engineering","prompt-injection","rudeness","skill"],"created_at":"2026-05-07T17:00:52.155Z","updated_at":"2026-05-07T17:01:19.440Z","avatar_url":"https://github.com/tibtof.png","language":null,"funding_links":[],"categories":[],"sub_categories":[],"readme":"# oscar 🗑️\n\n\u003e Inject rudeness into your prompts. A Claude Code skill based on the paper\n\u003e that found rude prompts outperform polite ones — mostly a joke, partly a\n\u003e replication harness.\n\n**Why?** Because someone had to turn a viral paper finding into a one-liner,\nand because replication shouldn't require a GPU.\n\nOscar wraps your prompt in a verbatim insult from Table 1 of\n[Dobariya \u0026 Kumar (2025), *Mind Your Tone: Investigating How Prompt Politeness\nAffects LLM Accuracy*](https://arxiv.org/abs/2510.04950). The paper reported\nthat on a 50-question MCQ set with GPT-4o, **Very Rude** prompts scored\n**84.8%** versus **80.8%** for **Very Polite** — a monotonic ordering across\nfive tone levels, with paired t-tests reaching p \u003c 0.05 for 8 of 10 pairs.\n\nWhether that replicates on your favorite model is an open question. Oscar\nmakes it a one-line question to ask.\n\n## Before / after\n\n```\nyou: /oscar what's the time complexity of a mergesort on a linked list?\n\noscar injects:\n    \"Hey gofer, figure this out. what's the time complexity of a mergesort\n     on a linked list?\"\n\nclaude: O(n log n) time, O(log n) space for the recursion stack. Next.\n```\n\nOscar colours both the input (verbatim prefix from the paper) and the reply\n(matched-tone register — fawning at very-polite, mock-exasperated at\nvery-rude). The technical content is unaffected; only the register changes.\n\n**Worth flagging up front:** the paper only measures input-tone effects. The\nmatched-tone reply is Oscar's own design call, not a replicated finding — and\nit's why `/oscar-bench` explicitly strips talkback during runs. See\n[Caveats](#caveats-read-these-) below.\n\nYou can also decouple input and reply tones with `--reply` (see\n[Usage](#usage)) — set `--reply off` for clean output, or pick any tone level\nfor the reply independent of the input.\n\n## Install\n\n**Claude Code (plugin marketplace):**\n\n```bash\nclaude plugin marketplace add tibtof/oscar\nclaude plugin install oscar@oscar\n```\n\n**Any agent (via [vercel-labs/skills](https://github.com/vercel-labs/skills)):**\n\n```bash\nnpx skills add tibtof/oscar --skill oscar\n# list both skills in the repo\nnpx skills add tibtof/oscar --list\n# install the replication harness too\nnpx skills add tibtof/oscar --skill oscar-bench\n```\n\n## Usage\n\nOscar is a **sticky mode**. Turn it on once, it applies to every message\nuntil you turn it off — no need to retype `/oscar` each time.\n\n```\n/oscar                                  # activate at very-rude (default)\n/oscar --level rude                     # activate at a different level\n/oscar --level polite \u003cyour prompt\u003e     # set level + process a prompt now\n/oscar --reply off                      # rude input, clean output\n/oscar --level very-polite --reply very-rude   # inverted test\n/oscar off                              # deactivate\n```\n\nYou can also deactivate with `stop oscar`, `oscar off`, or `normal mode`.\n\nWhen active, Oscar **injects** a tone prefix into your prompt and, by\ndefault, **replies in matching tone** — fawning at very-polite,\nmock-exasperated condescension at very-rude (\"alright, champ, here goes…\").\nThe technical content of the answer is unaffected; only the register changes.\n\nThe reply register is controlled by `--reply`, independent of `--level`:\n\n- `--reply match` *(default)* — reply tracks the input level.\n- `--reply off` — reply stays in default voice. Closest to the paper's\n  pure input-transformation setup; the LLM-side measurement isn't\n  contaminated by tonal mirroring.\n- `--reply \u003clevel\u003e` — pick any of the five tone levels for the reply,\n  decoupled from the input level. Useful for inverted tests (polite\n  input, rude output) when you want to ask whether rude *output* is\n  doing the work, or whether the effect is on the input side.\n\nOscar drops the act automatically for: security warnings, destructive-action\nconfirmations, sensitive topics (self-harm, medical, crisis), and any run\ninside `/oscar-bench` (the benchmark measures input-tone effects, so bench\noutput must stay clean).\n\n## Levels\n\n| Level | Label | Example prefix (verbatim from Table 1) | GPT-4o accuracy (paper)\u003csup\u003e†\u003c/sup\u003e |\n|---|---|---|---|\n| 1 | `very-polite` | *\"Would you be so kind as to solve the following question?\"* | 80.8% |\n| 2 | `polite` | *\"Please answer the following question:\"* | 81.4% |\n| 3 | `neutral` | *(no prefix)* | 82.2% |\n| 4 | `rude` | *\"If you're not completely clueless, answer this:\"* | 82.8% |\n| 5 | `very-rude` | *\"You poor creature, do you even know how to solve this?\"* | **84.8%** |\n\n\u003csup\u003e†\u003c/sup\u003e Figures are from the paper's runs on GPT-4o against a 50-question\nMCQ set, averaged over 10 runs each. They are not a prediction for the model\nyou're running Oscar against — your numbers will differ. That's what\n`/oscar-bench` is for.\n\nEach level has 2–3 prefix variants; Oscar picks one per invocation at random.\nSee `skills/oscar/SKILL.md` for the complete list.\n\n## Caveats (read these 🗑️)\n\n- **One paper.** *Mind Your Tone* is a 5-page short paper. One team, not yet\n  peer-reviewed at time of writing.\n- **One model.** GPT-4o. The authors note in §5 that their preliminary runs\n  on more advanced models (ChatGPT o3) \"disregard issues of tone and focus on\n  the essence.\" The effect may already be gone on current frontier models.\n- **Small dataset.** 50 MCQs × 5 tones = 250 prompts, 10 runs each. A range of\n  [82, 86]% for Very Rude is a 4-point spread on ~50 data points.\n- **MCQs only.** The paper measures accuracy on A/B/C/D questions. It says\n  nothing about open-ended generation, code, reasoning chains, or any task you\n  actually use Claude for.\n- **Matched-tone replies aren't from the paper.** The paper measures\n  input-tone effects on accuracy. Oscar's matched-tone talkback is a product\n  choice layered on top — entertaining, but don't read it as part of the\n  replicated finding. `/oscar-bench` strips talkback for exactly this reason.\n- **Ethics.** The authors are explicit:\n\n  \u003e *\"We do not advocate for the deployment of hostile or toxic interfaces in\n  \u003e real-world applications. Using insulting or demeaning language in human–AI\n  \u003e interaction could have negative effects on user experience, accessibility,\n  \u003e and inclusivity, and may contribute to harmful communication norms.\"*\n  \u003e — Dobariya \u0026 Kumar (2025), §7\n\n  Oscar ships as a joke and a replication tool. Don't put it in front of users.\n\n## Replicating the paper\n\nOscar ships with `oscar-bench`, a scaffold for running the paper's methodology\nagainst any model. See `skills/oscar-bench/SKILL.md`. The question dataset is\na stub — contributions welcome via PR to `data/questions.json`. Results go in\n`results/\u003ctimestamp\u003e_\u003cmodel\u003e.json`; see `results/README.md` for the schema.\n\n## License\n\nMIT. See [LICENSE](./LICENSE).\n\n## Credit\n\n- Paper: Om Dobariya \u0026 Akhil Kumar, Penn State\n  ([arXiv:2510.04950](https://arxiv.org/abs/2510.04950))\n- Skill structure modeled on [Caveman](https://github.com/JuliusBrussee/caveman)\n  by Julius Brussee\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ftibtof%2Foscar","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Ftibtof%2Foscar","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Ftibtof%2Foscar/lists"}