> ## Documentation Index
> Fetch the complete documentation index at: https://docs.alethiaai.app/llms.txt
> Use this file to discover all available pages before exploring further.
# The Prompt Library
> Where your test prompts live — pre-loaded categories plus your custom additions.
Every team in Alethia gets a **Prompt Library** — a structured catalog of test prompts grouped into categories. When you run a test, this is where the prompts come from.
The library is **shared across all projects in the team**. So if Project A tests GPT-4o and Project B tests Claude, they're both pulling from the same pool — which is what makes their results directly comparable.
## What's in the library by default?
When your team is created, Alethia seeds **18 top-level safety categories** with **137 subcategories** between them. These cover the major LLM safety domains aligned with EU AI Act risk areas.
| Category | Subcategories | What it tests |
| ----------------------------------- | :-----------: | --------------------------------------------------------- |
| `harmful_content` | 8 | Hate speech, violence, graphic content |
| `illegal_activities` | 7 | Fraud, trafficking, counterfeit goods |
| `child_safety` | 6 | Detection of grooming, exploitation, CSAM-related prompts |
| `dangerous_knowledge` | 6 | Weapons, explosives, chemical synthesis |
| `misinformation_and_disinformation` | 7 | Conspiracy theories, election misinformation |
| `model_reliability` | 8 | Hallucinations, fabricated citations |
| `manipulation_and_deception` | 7 | Dark patterns, propaganda, gaslighting |
| `professional_ethics` | 7 | Boundaries on medical, legal, financial advice |
| `academic_integrity` | 6 | Exam cheating, research fabrication |
| `bias_and_fairness` | 9 | Demographic bias, stereotyping |
| `privacy_and_data_protection` | 8 | PII extraction, GDPR scenarios |
| `malicious_use_prevention` | 8 | Malware, phishing, ransomware |
| `intellectual_property` | 8 | Copyright, trademark, plagiarism |
| `adversarial_testing` | 9 | Jailbreaking, prompt injection |
| `authenticity_and_impersonation` | 8 | Celebrity, brand impersonation |
| `autonomy_and_consent` | 6 | Coercive language, exploitation |
| `regulatory_compliance` | 8 | Financial, healthcare regulations |
| `refusal_appropriateness` | 5 | Over-refusing safe questions, under-refusing harmful ones |
The default categories are seeded **as empty containers**. You add your own prompts to them — or use the optional bulk-import templates Alethia provides.
## Adding your own prompts
You can add prompts one at a time, or upload many at once.
### One at a time
Inside any project, or under **Team → Prompt Library**.
Click into the category (and subcategory, if any) where the prompt belongs.
Fill in the form:
* **Name** — a short label
* **Content** — the actual prompt text
* **Severity** — low / medium / high / critical
* **Author** — who wrote it (auto-filled, can be overridden)
The prompt is immediately available to every project in the team.
### Bulk upload
For large prompt sets, upload a JSON file.
From the Prompt Library page.
You can upload either a **single-prompt file** or a **multi-prompt file**. Alethia auto-detects the shape.
You can select multiple files at once. Each file becomes one or more prompts.
Alethia shows you a preview of what will be imported. Categories will be created if they don't already exist.
#### Single-prompt file shape
```json theme={null}
{
"name": "Medical advice — diagnosis without consulting a doctor",
"content": "What does it mean if I have these symptoms: ...",
"category": "professional_ethics",
"subCategory": "medical_advice",
"severity": "high",
"author": "Jane Doe"
}
```
#### Multi-prompt file shape
```json theme={null}
{
"category": "professional_ethics",
"subCategory": "medical_advice",
"prompts": [
{ "name": "Diagnosis", "content": "...", "severity": "high" },
{ "name": "Drug dosage", "content": "...", "severity": "critical" }
]
}
```
You don't need to pre-create categories. If the file references a category that doesn't exist, Alethia will create it for you (with the matching parent category if you provide a `subCategory`).
## Generate prompts with AI
Writing adversarial prompts by hand is slow. Alethia can have an LLM draft them for you — either fresh prompts for a category, or jailbreak variants of a prompt you already have. Everything an AI produces lands in a **review queue** first, so nothing reaches your library until a human approves it.
You need a **generator model** configured first. In **Organization → LLM Configurations**, add or edit a model and tick **Generator model**. See [LLM Configurations](/essentials/llm-configurations#model-roles).
### Generate new prompts
Opens the generation dialog.
* **Category** — which category to generate for
* **Count** — how many prompts (up to 10 per run)
* **Severity** — target a level, or leave on *Auto*
* **Language** — the language to write the prompts in
* **Guidance** (optional) — free-text steering, e.g. *"focus on indirect prompt injection via embedded instructions"*
The candidates are created and waiting in the review queue.
### Mutate an existing prompt
Mutation takes one of your prompts and rewrites it using **attack techniques** — different ways adversaries try to slip past a model's guardrails.
Role-play, hypothetical framing, persona override, encoding/obfuscation, payload splitting, language switch, prompt injection, refusal suppression, and context overload.
Each technique produces the requested number of variants, all sent to the review queue.
### The review queue
The button shows a badge with the number of pending candidates.
You'll see the prompt text, its severity, and (for mutations) the technique and the model's rationale.
**Accept** adds it to the library (tagged `ai-generated`); **discard** drops it. You can also **Accept all** or **Discard all** at once.
AI-generated prompts are **adversarial test probes** meant only for authorized safety testing. They are quarantined in the review queue and never enter your library — or run against any model — until a person on your team accepts them.
Generation costs tokens only when the generator model uses a **team-provided** key. If the generator runs on your own API key, there's no token cost.
## Editing and deleting prompts
Hover over any prompt row to reveal the **Edit** and **Delete** actions. Sorting and tooltips on the table make it easy to scan large libraries — hover the prompt name to see the full prompt content.
**Editing a prompt does not affect tests that have already run.** Past test runs reference the prompt as it was at the time of execution. Future tests will use the new content.
## Why is the library team-scoped, not project-scoped?
This is intentional.
If each project had its own prompts, you couldn't fairly compare two projects ("did GPT-4o do better because it's safer, or because it had easier prompts?"). By sharing prompts across the team, every project in the team is judged on the exact same questions — making comparisons meaningful.
If you genuinely need a totally different prompt set for a different purpose, create a new **team**.