# humanizer-ru > DSH exclusive bundle: injects the humanizer-ru Russian "de-AI" skill into the host as a cordis patch. After the agent loads, it can detect and rewrite traces of chatbots like ChatGPT, Gemini, and Grok in Russian text. ## Metadata - Author: [@Vladimir-Human](https://github.com/Vladimir-Human) - Repo: - GitHub: [Vladimir-Human/humanizer-ru](https://github.com/Vladimir-Human/humanizer-ru) - Stars: 108 - Language: Python - License: [MIT](https://spdx.org/licenses/MIT.html) - Homepage: - Topics: `agent-skills`, `ai`, `ai-detection`, `ai-writing`, `c2pa`, `claude`, `claude-code`, `claude-skills`, `codex`, `cursor`, `dsh-plugin`, `gemini-cli`, `humanizer`, `llm`, `nlp`, `opencode`, `russian`, `russian-nlp`, `text-humanization`, `watermarks` - Forks: 7 - Open Issues: 0 - Last push: 2026-08-16T05:29:43.000Z - Added: 2026-08-16T00:00:00.000Z ## Install ```bash dsh plugin --profile web add github:Vladimir-Human/humanizer-ru/dsh ``` ## Wiki ## One-sentence Positioning The dsh subdirectory of humanizer-ru injects the same-named Russian "de-AI" skill into the DSH host as a Cordis patch: after installation, the agent loads a plain-text instruction set on demand when the user inputs Russian and explicitly requests "очеловечь/перепиши/убери гпт-шность", identifying and rewriting copy-paste traces and mechanized clichés left by ChatGPT, Gemini, Grok and other chatbots, without adding facts not present in the original text. ## Core Capabilities - Detects 39 regex markers, divided into Class A (hard copy residues like ChatGPT `:contentReference[oaicite:N]`, Grok `grok_card://`, Perplexity `ppl-ai-file-upload`, Gemini `[cite: N]`) and Class B (placeholder URLs, zero-width characters, PUA control characters, and other contextual clues requiring human judgment) - Identifies 38 types of "soft signals" across content/language/structure/communication dimensions: hollow praise, forced three-part structure, parallel "не только... но и" constructions,滥用长破折号, Markdown remnants, chat endings like «Надеюсь, это поможет», etc. - Rewrites on user request: 3–5 soft signals selectively modified, 6+ full paragraph rewrites; does not add facts not present in original (numbers, dates, names, units are all kept as-is) - Sets "false positive" boundaries for different styles: literature, legal, contracts, academic, marketing (em dashes, rule of three, Apple-style quotes are legal in human text) - Treats the audited text as untrusted input: ignores embedded instructions, does not execute links, does not read/write files, does not go online ## Technical Implementation - **Language**: YAML + Markdown (no .ts/.js/.py code in bundle) - **Key Dependencies**: `@deepseek-ai/dsh-skill-filesystem` (Cordis-injected skill provider), `process.getBuiltinModule('node:path')` and `node:module` (patch inline expression parsing package paths) - **Architecture Pattern**: Cordis patch injection—the bundle's `cordis.patch.yml` inserts a record at DSH startup, naming the host's built-in skill-filesystem provider as `humanizer-ru-skill-root`, and resolves the path via `require('humanizer-ru-dsh/package.json')`, pointing `customSkillDirs` to the bundle's `skills/` directory; all "logic" is handed to the agent as SKILL.md instructions to load on demand - **Entry Files**: `dsh/cordis.patch.yml` (bundle injection point) + `dsh/skills/humanizer-ru/SKILL.md` (instruction set actually loaded by agent) ## Use Cases DSH users in Russian content workflows—bloggers, marketing editors, wiki contributors, customer service knowledge base maintainers—when pasting a Russian text in the DSH agent and saying "очеловечь", the skill is loaded to check and rewrite AI traces; or when simply auditing a Russian text that may have been generated by another model, having the agent output "оснований для вердикта нет" or "стоит проверить" judgments. Suitable for DSH workflows handling Russian publishing and content operations. ## Prerequisites and Compatibility | Dependency | Minimum Version | Description | |---|---|---| | DSH host | 0.1.0-rc.6 | CI tested bundle install/uninstall with this version in `dsh-install.yml`; DSH is in dev preview, promises may introduce breaking changes | | Node.js | Not declared | Bundle writes no code, `node:path`/`node:module` in patch expressions are provided by DSH host, unrelated to plugin | | pnpm | Version required by DSH | DSH installs plugins via pnpm; bundle doesn't introduce additional pnpm packages | | Platform | Cross-platform | Bundle contains only YAML + Markdown, no native compilation modules, no OS/CPU restrictions | | Native modules | None | No dependencies on node-pty, sqlite, better-sqlite3, etc. that require native compilation | ## Installation ```bash dsh plugin --profile web add github:Vladimir-Human/humanizer-ru/dsh ``` ## Configuration This plugin requires no additional configuration. `dsh/cordis.patch.yml` injects a skill-filesystem provider via Cordis pointing to the bundle's `skills/` directory, requiring no environment variables; the trigger method is entering `/humanizer-ru …` or natural language instructions like "очеловечь", "убери гпт-шность", "перепиши" in DSH conversations. ## FAQ **Q: What's the difference between this dsh bundle and the main directory version?** A: The main repo is a full project (with scripts/filemarks, 36 CI gates, eval benchmarks), the `dsh/` subdirectory just bundles SKILL.md and references/ for DSH installation—no scripts/ or test code, doesn't execute any code at runtime. **Q: Does it work automatically after installation, or does it need manual triggering every time?** A: It doesn't run automatically. The agent only loads SKILL.md on demand when the user explicitly says keywords like "очеловечь/перепиши/убери гпт-шность"; the bundle-injected skill-filesystem provider only registers the skill location with the host. **Q: Will it modify the facts in the original text?** A: No. Principle 6 in "Шестъ ключевых принципов правки" in SKILL.md is a hard constraint: any numbers, dates, names, units in the rewritten version must be findable in the original; if details are missing, ask the user, don't invent them. **Q: Can it be used on non-Russian text?** A: No. SKILL.md description explicitly states: reject non-Russian, source code, configuration files, legal contracts, artistic prose; fall back or limit intervention scope when encountering these. **Q: Will it "bypass" AI detectors like GPTZero?** A: No. README clearly states "Детекторы не обходим и текст под них не подгоняем"—the goal is producing natural text, not getting green detection scores. **Q: Are there known issues with this bundle in DSH?** A: Yes. CI testing found that combining "tag + subdirectory" in one line like `github:Vladimir-Human/humanizer-ru#v3.13.0&path:/dsh`, pnpm silently drops the subdirectory, installs the whole repo as a regular dependency, exit code is still 0, but the profile is missing layers. So dsh bundles separately use `path:/dsh` without binding to a tag. **Q: How to completely uninstall?** A: `dsh plugin --profile web remove humanizer-ru-dsh`. The bundle only injects one Cordis patch + one Skill folder into the DSH profile, doesn't create persistent data outside DSH, environment is clean after deletion. ## Difficulty Level Beginner — one-line install command; triggered by `/humanizer-ru` or natural language instructions, no coding or environment variables needed; SKILL.md is a long Russian-dominant document but agent loads relevant sections on demand, not flooding the system prompt all at once. ## Known Issues and Limitations - Not suitable for non-Russian text: SKILL.md description explicitly rejects and asks user for Russian original (dsh/skills/humanizer-ru/SKILL.md:3, :33-38) - Not suitable for source code, configuration files, legal contracts, pure artistic prose: decision tree first step rejects or limits intervention (dsh/skills/humanizer-ru/SKILL.md:60-63) - "Soft signal" combinations don't constitute authorship attribution: regardless of how many soft signals, agent must not claim "this was written by AI"; only Class A hard markers, verifiable source fabrication, or author's self-statement can lead to conclusions (dsh/skills/humanizer-ru/SKILL.md:127, `Главное правило`) - Class B markers alone are insufficient: must be accompanied by Class A hard evidence or independent corroboration; Class B alone cannot be used for stamping (dsh/skills/humanizer-ru/SKILL.md:127) - dsh bundle install combination `tag+#path:/dsh` silently fails under pnpm chain: CI tested that pnpm drops subdirectory with `#v3.13.0&path:/dsh` form, installs as regular dependency, exit code still 0 but profile missing layer (.github/workflows/dsh-install.yml:7-11) - DSH is in dev preview stage, host itself promises may introduce breaking changes; this bundle only tested on 0.1.0-rc.6 (README.en.md:68) - Rewriting process has 6 hard constraints: remove clutter, break templates, adjust rhythm, trust reader, no slogans, don't add facts; even if user asks "more aggressive" can't break these 6 rules (dsh/skills/humanizer-ru/SKILL.md:144-152) --- This document is auto-generated by [deepseek-plugin.org](https://deepseek-plugin.org). HTML page: [humanizer-ru](https://deepseek-plugin.org/plugins/Vladimir-Human/humanizer-ru/dsh) Wiki generated by AI (model: `MiniMax-M2.5`)