# Kira's Prompt Pack — v3

The 25 prompts I actually keep in my clipboard. Each is tagged with the
model I run it on first. Swap in your own stack — the structure is the
point, not the brand.

> Personal taste: I put the *instruction* before the *content*. Claude
> follows long instructions better than most; OpenAI models occasionally
> need the instruction repeated at the end.

---

## 1. The "read this for me" prompt — Claude (Sonnet)

```
Read the linked PDF. In <= 250 words:
1. What is the central claim?
2. What evidence supports it?
3. What's missing or hand-waved?

Be plain. No hedging. If the paper is fluff, say so.
```

I run this every Sunday on five new ML papers. Half the newsletter
issues start from the output of this prompt.

## 2. The eval-writer — Claude (Opus)

```
Below is the task description + 3 example inputs/outputs from a system
I'm evaluating.

Generate 12 additional test cases, with edge cases as ~40% of them. For
each case, include the input, the expected output, and a one-line
reason this case matters.

Format as JSONL.
```

## 3. The "rewrite without losing me" prompt — Claude (Sonnet)

```
Rewrite the text below so it's clearer. Constraints:

- Keep my voice. I am: direct, short sentences, no hype, slight self-
  deprecation.
- Do NOT add transitions ("Furthermore", "Moreover").
- Do NOT add em-dashes I didn't write.
- If a sentence is fine, leave it alone.
```

(...remaining 22 prompts cover: code review, GitHub issue triage, RAG
chunking heuristics, structured-output forcing, "find the bug",
synthesizing 5 interviews into a quote graph, a meeting-summary you'd
actually send, an offer-letter negotiation reply, a polite-no template,
the "Karpathy-style explainer" prompt, eight more.)

---

If you got this from kiranakamura.substack.com — thanks for reading.
If you got this from somewhere else, the newsletter is free and the
prompts get updated whenever I change my mind.