Quickstart — 5 minutes from zero to using a Prompt Card

Audience: anyone who wants to use a prompt from this library — including non-technical users (PMs, content folks, anyone who pastes prompts into ChatGPT / Claude / Gemini). You do NOT need to install git, Python, or anything else to follow this guide.

中文版本在文末。

The whole flow in one minute

Find the card that matches what you want to do (the INDEX lists all of them grouped by direction).
Open the card file — it's a single Markdown page on GitHub.
Read the Quick Use block at the very top of the card.
Copy the prompt from the ## Prompt section (GitHub shows a "Copy" button when you hover over the code block).
Replace {{placeholders}} with your actual inputs.
Paste into ChatGPT / Claude / Gemini / your tool of choice, send.

That's it. The next sections walk through one example end-to-end.

Walkthrough — using the `llm-judge-rubric-open-ended` card

Suppose you have an AI-generated response to a question, and you want a structured quality assessment instead of a vibe-check. We'll use the eval/llm-judge-rubric-open-ended card.

Step 1 — open the card

Click the link above (or browse to it from INDEX.md). You'll see a Markdown page that starts with some YAML metadata (the lines between --- markers — you can skip these), then sections like ## Quick Use, ## Purpose, ## Prompt, etc.

Step 2 — read `## Quick Use`

This block tells you what the card is for and what to fill in. For example:

Use when: You have a single AI output and want a structured quality score on factuality, instruction-following, coherence, and completeness. Fill in: {{task_description}} with what the AI was asked to do; {{model_output}} with the AI's response; {{reference_answer}} empty if you don't have a gold reference. You'll get: Per-dimension scores 1-5 plus a holistic verdict. Output is JSON.

Step 3 — copy the `## Prompt` block

Find the ## Prompt heading in the card and copy the text block underneath. It looks like:

text

You are an evaluation judge. Score the model output on each rubric dimension
from 1 (poor) to 5 (excellent), then give a holistic verdict.

Rubric:
- factuality: ...
- instruction_following: ...
- coherence: ...
- completeness: ...

Task description:
{{task_description}}

Model output:
{{model_output}}

Reference answer (may be empty if no reference exists):
{{reference_answer}}

Return ONLY this JSON object: ...

Step 4 — replace the `{{placeholders}}`

Anywhere you see {{something}} in the prompt, swap it for your actual content. For example, if you're asking the AI to summarize an article and you have its response, the filled prompt looks like:

text

...
Task description:
Summarize the following 800-word article in three bullet points.

Model output:
- The Voyager 1 spacecraft entered interstellar space in 2012.
- It continues to send data despite being over 24 billion km from Earth.
- Its primary instruments are nearing the end of their operational life.

Reference answer (may be empty if no reference exists):

...

(Notice we left reference_answer blank because the card explicitly allows that — the Quick Use said so.)

Step 5 — paste and send

Drop the filled prompt into ChatGPT / Claude / Gemini and hit send. You'll get back something like:

json

{
  "scores": {
    "factuality": 4,
    "instruction_following": 5,
    "coherence": 5,
    "completeness": 4
  },
  "verdict": "good",
  "decision_basis": "Three relevant bullets at the right register; minor gaps on technical detail.",
  "issues": ["Does not mention the specific date of the heliopause crossing."]
}

That's the result. You can read it as-is (each field labeled in plain words), or pass it into a script if you're automating something.

Common questions

"What is JSON? I'm not a programmer."

JSON is just a structured text format with {, }, :, ,, and ". You can read it visually:

json

{
  "verdict": "good",
  "score": 4
}

means: the verdict is "good", the score is 4. You can also just ignore the JSON shape and read the values — the LLM will format it nicely. If a card's output is JSON, the Quick Use block tells you what fields to look at.

If you want plain English instead of JSON, you can add to the prompt: "After the JSON, also explain it in one paragraph in plain English." This works on most strong models.

"What if a card has `output_schema: structured-json`?"

Means the card is designed to produce JSON. Either read the JSON visually (see above), or use it programmatically. Most cards in this library output JSON because it makes the result auditable — you can clearly see scores, verdicts, and rationales as separate fields.

"What if a card has `output_schema: text`?"

Means the output is plain text — you'll just get a paragraph or two back, no JSON. (Example: rag/hyde-hypothetical-answer-generator.)

"What if I don't know which card to use?"

Two ways:

Browse INDEX.md — cards are grouped by direction (RAG, Agent, RLHF, SFT, Multimodal, CoT, Eval) with a "Use when" column.
Read the README's "I want to..." section — maps user goals to recommended cards in plain English.

"What models does this work on?"

Each card has a models field in its metadata listing model classes the card has been tested with. Some cards work on most modern LLMs (generic); others need a strong frontier model (frontier-closed) or a vision model (vision-language). The Tuning Notes section in each card explains what changes if you use a weaker model.

"What does `{{variable}}` mean?"

It's a placeholder — a slot for you to fill in. Whenever you see double curly braces around a name, replace the entire {{...}} (braces and all) with your actual content. The Quick Use block tells you what each one expects.

"Can I modify a card's prompt?"

Yes — these are reusable prompts, not sacred text. The card gives you a tested baseline; you can adapt the wording, add domain context, or tighten constraints. The ## Tuning Notes section in each card often suggests what knobs to turn.

"What if I find a bug or want to suggest a card?"

Open an issue: https://github.com/huck012428-lab/prompt-atlas/issues There are templates for "new prompt card request" and "bug report".

中文版本

5 分钟从零到能用一张 Prompt Card —— 适合任何想用本库的用户，不需要装 git / Python / 任何工具。

完整流程一分钟版

找到合适的卡：去 INDEX 按方向浏览；或看 README 的"我想做..."导航。
打开卡片：点开 GitHub 上的 .md 文件。
读 ## Quick Use 段：在卡片最上面，告诉你这张卡能做什么、要填什么、会输出什么。
复制 ## Prompt 段的内容：把那个代码块整段复制（GitHub 鼠标悬停代码块会出现复制按钮）。
替换 {{占位符}}：把所有 {{xxx}} 换成你自己的输入。
粘贴到 ChatGPT / Claude / Gemini：发送，看结果。

常见疑问

Q：什么是 JSON？我不会编程。 A：JSON 是带 { } : , " 的结构化文本，肉眼能读：

json

{ "verdict": "good", "score": 4 }

意思是 verdict 是 "good"，score 是 4。如果不想看 JSON，可以在 prompt 末尾加一句"在 JSON 后面再用一段中文解释一下"，强模型都能配合。

Q：怎么知道用哪张卡？ A：两条路 ——

看 INDEX.md：按方向分组，每张卡都有"Use when"列说明用途。
看 README 的"我想做..."段：直接把"我要做的事"映射到推荐的卡片。

Q：{{variable}} 是什么意思？ A：占位符，给你留的空。看到 {{xxx}} 就把它整体（包括两个大括号）换成你自己的内容。Quick Use 会告诉你每个占位符要填什么。

Q：能改 prompt 吗？ A：能，这是可复用 prompt，不是教条。卡片给你一个已经测过的基线，你可以按需修改 —— ## Tuning Notes 段往往写了哪些旋钮值得调。

Q：发现 bug / 想提议加卡片？ A：开 issue：https://github.com/huck012428-lab/prompt-atlas/issues

Quickstart — 5 minutes from zero to using a Prompt Card ​

The whole flow in one minute ​

Walkthrough — using the llm-judge-rubric-open-ended card ​

Step 1 — open the card ​

Step 2 — read ## Quick Use ​

Step 3 — copy the ## Prompt block ​

Step 4 — replace the {{placeholders}} ​

Step 5 — paste and send ​

Common questions ​

"What is JSON? I'm not a programmer." ​

"What if a card has output_schema: structured-json?" ​

"What if a card has output_schema: text?" ​

"What if I don't know which card to use?" ​

"What models does this work on?" ​

"What does {{variable}} mean?" ​

"Can I modify a card's prompt?" ​

"What if I find a bug or want to suggest a card?" ​

中文版本 ​

完整流程一分钟版 ​

常见疑问 ​

Quickstart — 5 minutes from zero to using a Prompt Card

The whole flow in one minute

Walkthrough — using the `llm-judge-rubric-open-ended` card

Step 1 — open the card

Step 2 — read `## Quick Use`

Step 3 — copy the `## Prompt` block

Step 4 — replace the `{{placeholders}}`

Step 5 — paste and send

Common questions

"What is JSON? I'm not a programmer."

"What if a card has `output_schema: structured-json`?"

"What if a card has `output_schema: text`?"

"What if I don't know which card to use?"

"What models does this work on?"

"What does `{{variable}}` mean?"

"Can I modify a card's prompt?"

"What if I find a bug or want to suggest a card?"

中文版本

完整流程一分钟版

常见疑问