Generating llms.txt

What llms.txt is

The llms.txt spec defines a markdown file at /llms.txt on a website that helps LLMs understand what the site contains. Instead of parsing HTML, an LLM fetches /llms.txt and gets a structured summary with links to individual pages.

The file uses a strict markdown format:

# Site Title

> One-line summary.

Context paragraph — tech stack, key concepts, anything the LLM needs before reading individual docs.

## Section Name

- [Doc Title](/path/to/doc/): Short description of the doc.

Where it lives

public/llms.txt — Astro serves everything in public/ as static files at the site root, so this becomes /llms.txt in the deployed site.

How it’s generated

The /update-llms-txt skill reads every .mdx file under src/content/docs/<project-slug>/, parses frontmatter for title and description, groups docs by domain directory, and writes the result to public/llms.txt.

To run it:

/update-llms-txt electron-terminal

The skill:

1. Collects all doc files matching the project slug
2. Reads astro.config.mjs for the site URL (uses relative paths if the URL is a placeholder)
3. Groups docs by domain (getting-started, architecture, terminal, ui, tooling)
4. Orders within each domain by Diataxis category: tutorials, guides, reference, resources
5. Writes each entry as - [Title](/path/): first sentence of description
6. Optionally generates public/llms-full.txt with concatenated full doc content if under 100KB

When to update

Update public/llms.txt whenever documentation changes:

• New doc written — add the entry in the right section
• Doc removed — delete the line
• Doc renamed or moved — update the link path
• Major description rewrite — update the description text

For a single doc addition, editing public/llms.txt by hand is faster than running the skill. The skill is better after batch operations like /plan-docs or when many docs change at once.

Manual editing

The format is simple enough to edit directly. Each entry follows this pattern:

- [Doc Title](/project-slug/domain/category/doc-name/): Short description.

Keep descriptions to one sentence — long entries defeat the purpose of a scannable index. The description should tell an LLM whether the doc is relevant to its current task, not summarize the full content.

Verifying

Every link in llms.txt should resolve to a real doc. A quick check:

grep -oP '\(/[^)]+/\)' public/llms.txt | tr -d '()' | while read path; do
  file="src/content/docs${path%/}.mdx"
  [ -f "$file" ] || echo "BROKEN: $path -> $file"
done