Prompts for Vibecoders

If you build with Cursor, Claude Code, Codex, or any AI coding assistant, MinT exposes its full documentation index as a single text file the assistant can fetch in one round-trip.

Concept

llms.txt is a community convention for serving an LLM-readable digest of a documentation site at a fixed URL. MinT auto-generates this file from the same MDX content you read, so the assistant sees exactly what you see — no scraping, no stale snapshots.

Three URLs are available on this site:

(ZH equivalents at /zh/llms.txt, etc.)

Pattern

Cursor

In Cursor, add the URL to your project's docs context:

@docs https://mint-doc.macaron.im/en/llms-full.txt

Cursor fetches the file, indexes it, and uses it as grounding for every prompt in the project. Update the index by re-fetching when MinT releases.

Claude Code

Tell Claude Code to fetch the index at the start of a session:

Read https://mint-doc.macaron.im/en/llms-full.txt for the full MinT documentation, then help me write a GRPO training loop on Qwen3-30B with a custom reward.

The fetch happens once per session; Claude Code keeps the content in context.

curl (smoke test)

Verify the route works from your terminal:

curl -sL https://mint-doc.macaron.im/en/llms.txt | head -40

You should see a list of page titles with their docs URLs.

API Surface

The route handler lives at app/[lang]/llms.txt/route.js in this repo. It walks the same fumadocs page tree the sidebar uses, strips MDX-only components (<Callout>, <WxImage>), and emits plain markdown. There are no query parameters; the file is regenerated on each request from the latest content build.

Caveats & Pitfalls

• Token budget. llms-full.txt for the entire site is several thousand tokens. Most coding assistants keep this comfortably in context, but if you only need a slice, paste a single page's URL with /content.md appended (e.g., /en/community/customize/sft/content.md) — fumadocs serves a per-page markdown view at that path.
• Freshness window. The route reads from the build output, so updates land when the site redeploys (typically within minutes of a merge to main of the public mint-doc repo). If a page you wrote yesterday isn't there yet, the build is still propagating.
• Roadmap pages are emitted as-is. A small number of pages describe features still landing (e.g., VLM). Those pages explain their status in plain language; the assistant sees that text verbatim and won't hallucinate a shipped feature.
• Don't paste MINT_API_KEY into a prompt. The docs themselves never contain real keys, but if you copy a script from your own machine into an LLM prompt, scrub sk-* tokens first.