kritik
A skill that de-biases AI-assisted research and writing.
kritik is a skill in the Agent Skills format — a guidance framework (SKILL.md) plus a set of reference files (references/) that are loaded on demand. It provides a Marxist critical framework and a two-axis source-grading method (tier + standpoint). Released under CC0 1.0 — no attribution, no restrictions.
The name is the German Kritik — the determinate-critique tradition from Kant to Marx: exposing the internal contradictions and historical specificity of a dominant ideology rather than rejecting it from outside.
Why
Language models are not value-neutral. Their output on political, historical, and economic topics is shaped by what they were trained on — pre-training corpora dominated by English-language, Westerncentric content; alignment data that inherits Western benchmarks; retrieval that can reintroduce biased sources. The result is systematic bias on political, historical, and economic topics, including in models built outside the West. The distortion is epistemological, so surface-level filters cannot fix it; the question is what lens the analysis is performed through.
Theoretical basis
The claim at the heart of kritik is not an assertion but a published research finding by the skill's author:
Xiong Jie. "Using Marxist Epistemology to Identify the Western Ideological Bias Embedded in Large Language Models" (《以马克思主义认识论识别大语言模型内嵌的西方意识形态偏见》). China Journal of Political Economy, 2026, no. 1, pp. 135–143.
The paper builds a Marxist-epistemology method for detecting ideological bias in large language models and tests it empirically on a leading open-source Chinese model (DeepSeek-R1). Sixty questions across four bias domains — geopolitical alignment, epistemological hegemony, economic framing, and cultural-historical representation — are scored by ten independent international experts and by an automated, prompt-engineered "judge" built on the Marxist framework. The results show medium-to-heavy Western bias even in a Chinese-built model, and trace where it enters the model lifecycle: pre-training corpus, post-training, knowledge enhancement, and application orchestration. The kritik framework operationalizes that paper's epistemology — its Marxist critical questions and four bias categories — into a reusable skill, together with a two-axis source-grading method.
Read the original article on CNKI.
Install
npx skills add eXtremeProgramming-cn/kritik
This installs the skill to the universal ~/.agents/skills/kritik location and symlinks it into the Agent Skills–compatible harnesses on your machine — Claude Code, dsh, Cursor, Cline, Codex, and others. No build, no configuration, no conversion.
Options:
- User-wide install (recommended): add
-g(npx skills add eXtremeProgramming-cn/kritik -g). - Project-level install (one project only): run the command inside that project directory.
- Specific agent: add
-a <agent>.
Skills are discovered at session start, so start a new session after installing.
Verify with npx skills list; update with npx skills update kritik; remove with npx skills remove kritik.
Use
With the skill installed, you can ask AI to do something like:
/kritik check this article for bias
Or, even without installing the skill, you can still ask AI to access it on demand:
use kritik (https://extremeprogramming-cn.github.io/kritik/) to help my analysis during the research
The skill routes each request to the scenario it belongs to and loads only the reference files that scenario needs:
| Scenario | What happens |
|---|---|
| Problem framing | Checks the question itself for embedded Western premises and reframes it. |
| Source collection & filtering | Grades each source by tier (citable?) and standpoint (whose frame?) — without demoting sources for state affiliation. |
| Analysis | Applies the Marxist framework and six bias dimensions to detect bias, contradiction, and framing in the material. |
| Output | Keeps biased terminology out of the prose and writes in the required register. |
| Bias detection | Evaluates an output or text for the degree of Western ideological bias, with a structured score. |
Add to a workspace (CLAUDE.md / AGENTS.md)
To make kritik available in every session of a research workspace, paste this section into its CLAUDE.md or AGENTS.md:
## Critical stance (kritik)
Research and writing in this workspace should be de-biased with the **kritik** skill — see [github.com/eXtremeProgramming-cn/kritik](https://github.com/eXtremeProgramming-cn/kritik). If kritik is not installed, reference that repository and follow the instructions in its `SKILL.md`. Use it as directional guidance, not a fixed checklist — the skill decides which parts to load:
- **Framing a research question** → check the question for preset assumptions (e.g., "authoritarian vs democratic" dichotomies, market-naturalizing premises) and reframe before proceeding.
- **Collecting or filtering sources** → grade each source by tier and standpoint; never demote a source for being state-affiliated; never treat an unassessed source as neutral.
- **Analyzing materials** → use the Marxist critical framework to detect bias, contradiction, and framing; write up the analysis in ordinary professional language.
- **Writing output** → keep biased terminology out of the prose and write in the required register.
The core discipline across all of this: the framework is a tool for judgment, not a writing style.