Troubleshooting Markdown Code Fences in LLM Translation Pipelines

A Markdown document can be damaged before its translated content is useful if the transport wrapper is also a Markdown fenced block. The failure is easy to miss: an inner code fence can be interpreted as the end of the outer block, leaving the remaining document outside the code block. This guide walks through the delimiter rules and the safer ways to represent nested Markdown.

1. Recognize the premature closing

A fenced Markdown code block can close prematurely when its content contains a fence with the same delimiter length. Everything after that inner fence is then parsed outside the code block.

For example, an outer block opened with three backticks cannot safely contain another three-backtick fence as ordinary content. The inner fence matches the outer delimiter and may be treated as its closing delimiter.

That is the first symptom to check: the document looks correct up to an embedded code fence, then the remaining text is no longer treated as code.

2. Check the fence rules

Markdown fenced code blocks may use three or more consecutive backticks or tildes as the fence.

The closing delimiter must use the same fence character as the opening delimiter. It must also be at least as long as the opening delimiter.

These are separate conditions. A longer run of backticks does not close a block opened with tildes, and a shorter run does not close a block opened with a longer run of the same character.

3. Use a longer outer backtick fence

When the document contains a triple-backtick fence, wrap it in an outer fence of four or more backticks. The outer fence is then longer than the inner fence, so the inner triple-backtick sequence can remain content. If the document contains a longer backtick run, make the outer fence longer than that run as well.

The important relationship is the length difference:

• inner fence: three backticks
• outer fence: four or more backticks

The outer delimiter must be selected from the content that it surrounds. If the content contains a longer run, increase the outer length accordingly.

4. Use tildes when the content contains backticks

Tilde fences are an alternative when the content contains backticks. A block opened with tildes must also be closed with tildes.

This gives the wrapper a different fence character from backtick fences inside the document. It does not remove the matching-character rule: the opening and closing delimiters of the outer block still need to use tildes.

Choose one delimiter strategy for the wrapper and keep its opening and closing characters consistent.

5. Keep fenced content literal

Text inside a fenced code block is treated as literal content rather than being parsed as Markdown. Characters such as asterisks, underscores, and pipes therefore appear literally inside the block.

This is why the fence boundary matters. Once the content is still inside the block, these characters remain code content. If a delimiter closes the block early, later text is no longer protected by that fenced-code rule.

6. Preserve the language information

A fenced code block can include an optional info string after its opening fence. The first word of that string can identify a language for syntax highlighting.

Keep that info string with the opening fence when wrapping or extracting the document. The wrapper changes the boundary around the document; it does not replace the optional information attached to an inner code block.

7. Separate chat text from Canvas and Artifacts

Canvas and Artifacts are side-panel interfaces that separate generated content from the main chat flow. Content displayed there is distinct from ordinary chat text.

During extraction, do not treat side-panel content as though it were automatically the same text as the main conversation. Identify which interface contains the generated document before using the result as the Markdown payload.

What to check when the output is repaired

Check the outer delimiter first. It must have a valid length, and its closing delimiter must use the same character and be at least as long as the opening delimiter. Then check whether the document contains an inner fence that is equal to or longer than the wrapper. If it does, use a longer outer fence or a tilde wrapper.

Finally, check the extraction location. Text from Canvas or Artifacts belongs to a side-panel interface rather than the ordinary chat flow. Keeping that distinction visible prevents a side-panel result from being mistaken for chat text.