Making Agent Quotations Verifiable in Markdown

Where this fits

When an agent drafts a technical survey, it pulls dozens of quotes from papers and web pages. The default Markdown blockquote syntax and inline bracketed references make it hard for a reader to tell who wrote what and where the original text stops. The writing-quotation skill solves this by enforcing a single, explicit quotation format: fenced code blocks that hold the original and translation together, with the source reference placed on the line after the fence. For developers building agents that produce sourced documents, this turns quote formatting from a scattershot convention into a deterministic rule set.

Why agents benefit

• Agents can emit quotation blocks without guessing: original and translation stay in one fenced block, and the source line follows immediately, so downstream parsers know exactly where each citation belongs.
• The skill forbids bare numeric references like [1]. Agents must attach a real source after every quote, which improves traceability when the reader wants to verify a claim.
• It composes with other sourced-writing skills (survey, paper-details). An agent can call writing-quotation whenever it needs to insert a quote, keeping across-document formatting consistent.
• By preserving the original text next to the agent's paraphrase, the agent gives the reader a way to cross-check interpretation against evidence — a concrete guard against hallucinated or twisted citations.
• Because the rules live in a reusable skill, the behavior can be updated in one place instead of being rediscovered in the model's context window on every run.

Practical scenarios

Generating a literature survey

A research agent builds a survey of 30 papers. Instead of mixing blockquotes, italics, and [1] pointers, it wraps every paper excerpt in a fenced block with the source line on the next line. The reader can read the claim, check the original, and follow the reference without leaving the document.

Multilingual prompt template documentation

An agent documents prompt templates from Chinese product blogs for a team that reads English. writing-quotation keeps the original template on the first line and an English translation on the second line inside the same code fence, so both versions stay visible and side-by-side.

Verifying a web page citation

An agent writing a fact-check report quotes a paragraph from a news article. Instead of a raw blockquote, it outputs a fenced code block with the retrieved text and then a source line containing the URL and access date, which makes the provenance machine-readable and auditable.

Add it to your agent workflow

Install the skill with:

npx skills add mathbullet/skills --skill writing-quotation

After installation, the agent knows to format any quotation like this:

```text
Original: "Attention is all you need."
Translation: "你需要的只是注意力。"

Source: Vaswani et al. (2017), https://arxiv.org/abs/1706.03762

The source line sits outside the code fence, so the quoted content remains clean and the reference is explicit.

## Compare before adopting

Compare `writing-quotation` against your own quote-formatting instructions. Look at quality signals: does it enforce a consistent structure that your evaluator can verify? Check maintenance freshness: is the repository actively updated when Markdown conventions or peer skills change? Consider alternatives such as hand-written system prompts or other formatting skills, and decide whether the skill's strict code-block-only rule fits the documents you produce.

## Why it is worth tracking

`writing-quotation` has 119 stars and is part of the mathbullet/skills repository, which signals a maintained ecosystem rather than a single abandoned snippet. If you regularly generate papers, surveys, or source-heavy reports, this skill is worth adopting now and reassessing when your citation requirements change. You can also review the skill's current state at https://www.openagentskill.com/skills/mathbullet-writing-quotation to see whether recent updates match your workflow.