How to use the ljg-mechanism agent skill for mechanism research

What does the ljg-mechanism skill make the agent do?

Given a puzzling phenomenon, the skill makes the agent work through a four-part framework: look at the anomaly (看反常), trace the process (追过程), break down the explanation (分解释), and carry it to new conditions (带条件迁移). The agent first checks that the observation is accurate, then describes the step-by-step process that could produce it, then compares rival explanations.

The agent is instructed to search the web and read sources to test these explanations. Its output is a current best mechanism judgement, how strong the evidence is, what new evidence would change its mind, and the conditions under which the explanation applies. The skill describes itself as a structured reasoning method rather than a data tool.

Install the skill from the ZIP or the Claude apps

There are two routes. The first is to download the ZIP from the skill's page on this site, then unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only in that project.

The second route is for the Claude apps: download the ZIP, then upload it under Settings → Capabilities → Skills.

The skill's files are agents/openai.yaml, SKILL.md and Workflows/DiscoverMechanism.md. After unzipping, the folder should contain these files. The main instructions live in SKILL.md.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Check what the skill needs before you start

The skill needs a web search and page-fetching capability in the agent runtime. It says it defaults to using these to check the phenomenon and the mechanism. Without them, the agent can only offer an offline reasoning sketch, and it must say that it has not verified anything online.

The skill runs no scripts. It does use the network, because its checks depend on web search and reading sources.

• Requires: web search and page fetching in the agent runtime
• Runs scripts: no
• Uses network: yes

How do you trigger the skill with a natural request?

You can invoke it by name, or use everyday phrases. The skill's description lists these triggers: calling ljg-mechanism directly, asking to 找机制 (find the mechanism), 机制研究 (mechanism research), 用四步框架走一遍 (walk through the four-step framework), 联网验证这个解释 (verify this explanation online), or asking how a phenomenon comes about.

The skill is not meant for pure concept definitions, plain summaries, questions with one clear answer, or requests that only want a map. Ask a clear causal question and the skill is useful; ask for a definition and it adds effort without value.

• Ask why something happens, for example why AI productivity gains do not show up in income
• Ask it to test a viral claim that a technique does not work
• Ask it to walk through the framework step by step

Read the output with the right expectations

The output is a reasoned judgement based on web searches, which may be incomplete or wrong. Treat cited sources as the thing to check. The skill itself notes that a source proving a phenomenon occurred does not prove a candidate mechanism. It also notes that finding no counterexample is not the same as passing verification.

The skill is read-only by default. It should not create notes, reports, logs or memory files unless you ask for them. If you ask it to save results, it follows the location and format you specify.

• Check cited sources yourself before relying on a claim
• Expect the agent to say which parts are sourced, inferred or unknown
• Expect it to change its judgement if new evidence contradicts it

Examples and tips for better results

The skill's own examples show three kinds of request. One asks why AI makes creation easier but income does not rise at the same pace. The skill separates output, delivery, sales, price and net return, then compares candidate explanations. A second asks how an unmanaged group still gets questions answered, and the skill compares self-assignment with centralised matching. A third checks a claim that a technique has no effect, starting with what the metric, period and group actually are.

Specific questions produce better results. Name the outcome you want explained, the group, and the time period. If the vague claim you started with turns out not to hold, the skill is designed to reframe the research question rather than explain a phenomenon that is not there.

• Name the metric, time period and group in your request
• Give the rival explanations you already suspect, if any
• Ask for the sources behind key causal links