# Prompt Debugging Checklist

AI Usage Playbook / Studio Nani

Purpose: when an AI answer is generic, wrong, unstructured, or hard to reuse, stop blaming “the model” and inspect which part of the prompt interface is broken.

## How to use it

1. Paste the original prompt.
2. Paste a weak output sample.
3. Walk through the 7 checks below.
4. Ask the model to return: top 3 causes → repaired prompt → expected improvement.
5. Verify important facts, current information, and calculations with tools separately.

## 7-step checklist

### 1. Goal check
- Check: What result should this prompt produce, and what decision or action will it support?
- Repair: Rewrite the first line as: “Help me accomplish [goal] for [audience/context].”

### 2. Context check
- Check: What does the model need to know that is not already in the prompt?
- Repair: Add audience, constraints, prior attempts, data/source boundaries, and examples of what “good” looks like.

### 3. Output contract check
- Check: Did I specify the exact format, sections, length, language, and tone?
- Repair: Give a visible contract: headings, bullet count, table/no-table rule, JSON shape, or final checklist.

### 4. Example / rubric check
- Check: Did I show a good sample or define how the answer will be judged?
- Repair: Add one positive example, one anti-example, or a 3-point rubric before asking for the final answer.

### 5. Evidence check
- Check: Does this task require current facts, sources, calculations, or tool-based verification?
- Repair: Ask the model to label assumptions and separate verified facts from guesses; use tools for anything factual or current.

### 6. Constraint check
- Check: What must the answer avoid, preserve, or not change?
- Repair: Name the hard constraints: budget, file scope, forbidden claims, style rules, safety boundaries, or no-go options.

### 7. Reuse check
- Check: Can this repaired prompt become a reusable pattern next time?
- Repair: Replace one-off details with variables like [audience], [source], [output format], and [quality bar].

## Copy-paste request

```text
Debug the prompt and output below.

Original prompt:
[paste]

Weak output:
[paste]

Desired result:
[describe]

Checklist:
1. Is the goal clear?
2. What context is missing?
3. Is the output format explicit?
4. Are there examples or evaluation criteria?
5. Does this require sources, verification, or calculation?
6. Are constraints explicit?
7. Can the prompt be turned into reusable variables?

Return format:
- Top 3 causes
- Repaired prompt
- Expected improvement
- What still needs verification
```

Source/editorial note: This checklist is adapted from OpenAI/Anthropic prompt engineering guidance and Studio Nani field tests. It is rewritten as a reusable pattern, not copied as a prompt artifact.