Summary

A prompt library gives your team a single, governed source of truth for every AI instruction your workflows depend on. This guide covers creating, categorizing, and assigning prompts in DAVE's Prompt Library, including AI-assisted generation and prompt locking.

The steps

1
Open the Prompt Library tab
Navigate to Agents in the left sidebar. The Agents page has two tabs: Agents and Prompt Library. Click Prompt Library to switch to the library view.
The Prompt Library tab shows a table of all saved prompt templates in your tenant. If this is a new workspace, the table will be empty. Each row in the table shows the prompt's name, category badge, tags, and the agent it is attributed to (if any).
You need the Admin or Create role to create or edit prompts. The Curate role can view the Prompt Library but cannot create or modify prompts. The Use and Reporting roles cannot access the Agents page at all. DAVE ships with six default tenant roles (Admin, Create, Curate, Use, Reporting, and Financial), all editable by your administrator.
Do this
- Click Agents in the left sidebar.
- Click the Prompt Library tab (next to the Agents tab).
- Confirm you can see the prompt table. If the tab is missing or access is denied, your role does not include Prompt Library access.
Example
A content operations lead opens DAVE, clicks Agents in the sidebar, then clicks the Prompt Library tab. The table is empty. They are about to create the team's first shared prompt.
Best practice
Before creating any prompts, spend five minutes planning your library structure. Decide which categories you will use (Content Generation, Review, Safety, Analysis, Transformation, or Custom), what your tagging convention will be, and which prompts are shared across multiple agents versus specific to one. A few minutes of planning now prevents a disorganized library later.
2
Click New Prompt and fill in the name, description, and category
Click New Prompt to open the creation form. Fill in the following fields:
- Name (required): the display name for the prompt. Choose a name that describes the prompt's function precisely, not the agent it happens to be used with first. A name like "Editorial Abstract Generator" is more durable than "Blog Agent Prompt" because the prompt may be reused across multiple agents over time.
- Description (optional): a brief explanation of what the prompt does, what inputs it expects, and what output it produces. Write it for the person who will assign this prompt to a new agent six months from now.
- Category (required): select one of the six categories that best describes the prompt's function:
  - Content Generation: prompts that produce original text, summaries, or structured output.
  - Review: prompts that evaluate content against a standard and return a verdict or feedback.
  - Safety: prompts that check content for policy violations, harmful material, or compliance issues.
  - Analysis: prompts that extract insights, classify content, or identify patterns.
  - Transformation: prompts that reformat, translate, or restructure existing content.
  - Custom: prompts that do not fit cleanly into the above categories.
Do this
- Click New Prompt on the Prompt Library tab.
- Enter a descriptive Name for the prompt.
- Optionally enter a Description explaining the prompt's purpose, expected inputs, and expected output.
- Select the most appropriate Category from the six options: Content Generation, Review, Safety, Analysis, Transformation, or Custom.
Example
The content operations lead creates a prompt named "Editorial Abstract Generator". They write the description: "Produces a two-sentence abstract from a long-form article. Input: full article text. Output: two sentences in the present tense, factual and neutral in tone." They select Content Generation as the category.
Best practice
Use the category system consistently across your team. If everyone uses Review for evaluation prompts and Safety only for policy-checking prompts, the library stays navigable as it grows. Resist the temptation to use Custom as a catch-all: most prompts fit one of the five named categories, and using them correctly makes filtering faster.
3
Add tags and write (or generate) the prompt content
With the name, description, and category filled in, add tags and write the prompt content.
Tags are a comma-separated list of labels for filtering and organization. Use the same tagging convention as your agent tags so the two systems stay consistent. For example: content, abstract, editorial. Each tag is trimmed of whitespace; empty entries are ignored.
Content is the prompt text itself: the instruction that will be passed to the AI agent as its system prompt (or as part of it). Write it precisely. A vague prompt produces inconsistent output; a precise prompt produces predictable output that can be evaluated, improved, and versioned.
If you want help writing the prompt, click AI Assistant to use DAVE's built-in prompt generation tool:
1. Enter a narrative description of what you want the prompt to accomplish, for example: "I need a prompt that summarizes long-form articles into two-sentence abstracts, written in the present tense, factual and neutral in tone."
2. Click Generate. The system uses your tenant's default AI provider and model (configured by admins in Admin → Settings → General → System Parameters → AI Assistant) to produce a structured prompt. If no default is configured, the system tries each available provider in order.
3. Review the generated prompt in the Content area.
4. Edit as needed before saving.
The AI Assistant is a starting point, not a final draft. Always review and refine the generated prompt before saving it to the library.
Do this
- Enter Tags as a comma-separated list using your team's tagging convention.
- Write the prompt Content directly in the text area, or click AI Assistant to generate a draft.
- If using AI Assistant: enter a narrative description, click Generate, review the output, and edit as needed.
- Read the final prompt content carefully before moving to the next step. Once saved, the prompt can be edited, but any agents already using it will not automatically update.
Example
The lead adds the tags content, abstract, editorial and writes the prompt content directly: "You are an editorial summarizer. Given a long-form article, produce a concise two-sentence abstract suitable for an editorial dashboard. Write in the present tense. Be factual and neutral in tone. Do not add information not present in the source text." They decide not to use the AI Assistant for this prompt because they already know exactly what they want.
Best practice
Write prompts as if they are specifications, not suggestions. Every instruction should be unambiguous. If the prompt says "be concise", the agent has to guess what concise means. If it says "produce exactly two sentences", the agent has a clear target. Specificity is the difference between a prompt that works reliably in production and one that works most of the time.
4
Optionally attribute the prompt to an agent, then save
The creation form includes an optional Agent field. Selecting an agent here attributes the prompt to that agent in the library view: the agent name appears in the prompt's table row, making it easy to see at a glance which prompts belong to which agents.
Attribution is informational. It does not automatically assign the prompt to the agent or lock it. Assignment and locking happen separately on the agent edit form (covered in the next step).
If the prompt is intended to be shared across multiple agents, leave the Agent field blank. Attribution is most useful for prompts that are purpose-built for a single agent and unlikely to be reused elsewhere.
Click Create to save the prompt. The new prompt row appears in the Prompt Library table with its name, category badge, tags, and agent attribution (if set).
Do this
- Optionally select an Agent from the dropdown to attribute the prompt to a specific agent.
- Click Create to save.
- Confirm the new prompt row appears in the Prompt Library table with the correct name, category badge, and tags.
Example
The lead attributes the prompt to the Blog Post Summarizer agent, then clicks Create. The prompt row appears in the table: "Editorial Abstract Generator" with a blue Content Generation badge, the tags content, abstract, editorial, and the agent attribution Blog Post Summarizer.
Best practice
Repeat steps 2 through 4 for each prompt you want in the library. Build the full set of prompts your team needs before moving on to assignment and locking. It is easier to review and refine a set of prompts together than to create them one at a time as agents are built. A complete library also makes it easier to spot gaps and duplicates before they become embedded in agent configurations.
5
Assign a library prompt to an agent
Creating a prompt in the library does not automatically assign it to any agent. Assignment happens on the agent edit form. Navigate to Agents (the Agents tab, not Prompt Library), find the agent you want to update, and open its detail page. Click the edit action to open the agent edit form.
On the agent edit form, find the prompt source selector. You have two options:
- Inline Prompt: write a prompt directly on the agent, stored in the agent's own configuration.
- Library Prompt: select a saved prompt from the Prompt Library.
Select Library Prompt and choose the prompt you just created from the dropdown. The prompt content is displayed for review.
Save the agent to apply the assignment. The agent now uses the library prompt as its system prompt at runtime.
Do this
- Navigate to Agents and click the agent you want to update.
- Open the agent edit form.
- In the prompt source selector, choose Library Prompt.
- Select the prompt from the dropdown.
- Review the prompt content displayed in the form.
- Save the agent.
Example
The lead opens the Blog Post Summarizer agent, switches the prompt source to Library Prompt, and selects "Editorial Abstract Generator" from the dropdown. The prompt content appears in the form for review. They save the agent. The agent now uses the library prompt at runtime.
Best practice
After assigning a library prompt to an agent, create a new agent version to record the change in the agent's version history. A prompt assignment change is a behavior change: it should be versioned, tested in a draft workflow, and promoted to production following the same process as any other agent configuration change. Do not assume the agent behaves identically after a prompt assignment change without testing.
6
Lock the prompt to prevent workflow-level overrides
By default, a library prompt assigned to an agent can still be overridden at the workflow node level: a workflow designer can select a different prompt in the Workflow Editor's Agent Interaction node configuration. If you want the agent to always use its assigned library prompt regardless of which workflow references it, lock the prompt.
To lock a prompt, toggle the Lock Prompt switch on the agent edit form after selecting the library prompt. When locked:
- The agent always uses the locked library prompt when invoked in any workflow.
- In the Workflow Editor, the prompt selector for this agent is greyed out. Workflow designers cannot override it.
- This ensures consistent agent behavior across all workflows that use the agent.
Locking is the right choice for any prompt that defines the agent's core behavior and should not vary by workflow. It is especially important for safety prompts, compliance-checking prompts, and any prompt where consistency is a governance requirement.
Save the agent after toggling the lock. The lock state is stored on the agent, not on the prompt: the same prompt can be locked on one agent and unlocked on another.
Do this
- On the agent edit form, with a library prompt selected, toggle the Lock Prompt switch to the locked position.
- Confirm the switch is active (locked).
- Save the agent.
- Optionally, open the Workflow Editor and navigate to an Agent Interaction node that references this agent. Confirm the prompt selector is greyed out and cannot be changed.
Example
The lead decides the "Editorial Abstract Generator" prompt should be locked: every workflow that uses the Blog Post Summarizer agent should produce the same style of abstract, regardless of who built the workflow. They toggle Lock Prompt and save. When a colleague opens the Workflow Editor and adds a Blog Post Summarizer Agent Interaction node, the prompt selector is greyed out. The behavior is consistent and governed.
Best practice
Lock prompts by default for any agent used in a production workflow. Unlocked prompts are appropriate during development and experimentation, when you want workflow designers to be able to test different prompt variations. Once an agent is promoted to production, lock its prompt to prevent accidental or unauthorized overrides. The lock can be removed later if requirements change, but the default should be locked in production.

What the Prompt Library is and why it exists

The Prompt Library is a collection of reusable prompt templates stored at the tenant level in DAVE. It is accessible from the Prompt Library tab on the Agents page. Each prompt in the library has a name, an optional description, a category, tags, and optional agent attribution. Prompts can be created manually or with AI assistance, assigned to agents, and locked to prevent workflow-level overrides.

Without a prompt library, system prompts live inside agent Config JSON objects. That works for a single agent, but it does not scale. When the same instruction is needed in three agents, it gets duplicated in three Config JSON objects. When the instruction needs updating, it gets updated in one place and forgotten in the other two. When a new team member wants to understand how the organization's AI agents work, there is no single place to look.

The Prompt Library solves this by separating prompt management from agent configuration management. Prompts are first-class objects with their own lifecycle: they can be created, reviewed, updated, and assigned independently of the agents that use them. A team can maintain a library of 20 carefully crafted prompts and assign them to 50 agents, with a single update propagating to every agent that uses a given prompt.

The library also enables a division of labor that matches how most teams actually work. The people who write good prompts (content strategists, editorial leads, compliance officers) are often not the same people who configure agents and build workflows (developers, operations leads). The Prompt Library gives each group a workspace that fits their role: prompt authors work in the library, agent builders assign and lock prompts from the library, and workflow designers work within the constraints the locked prompts define.

The six prompt categories and how to use them

Every prompt in DAVE's Prompt Library belongs to one of six categories. The category is required when creating a prompt and is displayed as a colored badge in the library table. Categories are fixed: you cannot create custom category names beyond the six provided. Use the Custom category for prompts that do not fit the other five.

Content Generation

Prompts that produce original text, summaries, structured output, or any other content where the agent's primary job is to create something. Examples: article summarizers, abstract generators, product description writers, meeting note formatters. Use this category when the agent's output is the deliverable.

Review

Prompts that evaluate content against a standard and return a verdict, score, or structured feedback. Examples: editorial quality checkers, tone reviewers, factual accuracy assessors, brand consistency evaluators. Use this category when the agent's output is a judgment about someone else's content, not the content itself.

Safety

Prompts that check content for policy violations, harmful material, regulatory compliance issues, or other safety concerns. Examples: content moderation prompts, PII detection prompts, legal compliance checkers. Use this category for any prompt whose primary purpose is to protect against risk rather than to produce or evaluate content.

Analysis

Prompts that extract insights, classify content, identify patterns, or produce structured data from unstructured input. Examples: sentiment classifiers, topic extractors, entity recognizers, keyword analyzers. Use this category when the agent's job is to understand or characterize content rather than produce or evaluate it.

Transformation

Prompts that reformat, translate, restructure, or otherwise change the form of existing content without substantially changing its meaning. Examples: language translators, format converters, tone adjusters, length reducers. Use this category when the input and output are the same content in different forms.

Custom

Prompts that combine multiple functions or do not fit cleanly into the above five categories. Use Custom sparingly. A prompt that is genuinely multi-purpose may be better split into two focused prompts, each in its own category, assigned to separate agents in a workflow chain. A library of well-categorized single-purpose prompts is easier to maintain and reuse than a library of complex multi-purpose prompts filed under Custom.

Prompt locking as a governance mechanism

Prompt locking is the mechanism that makes the Prompt Library a governance tool rather than just an organizational convenience. Without locking, assigning a library prompt to an agent is advisory: a workflow designer can still override it at the node level in the Workflow Editor. With locking, the assignment is enforced: the agent always uses its locked prompt, and the prompt selector in the Workflow Editor is greyed out for that agent.

This matters for several reasons:

Consistency. A locked prompt guarantees that every workflow using the agent applies the same instruction. Output quality and behavior are predictable regardless of who built the workflow or when.
Governance. For safety prompts, compliance-checking prompts, and any prompt that represents an organizational policy, locking ensures the policy is applied consistently. A workflow designer cannot accidentally or deliberately bypass it.
Auditability. When a prompt is locked, the audit trail for agent behavior is cleaner: the prompt that ran is always the locked library prompt, not a per-workflow override that may or may not be documented.

The lock state is stored on the agent, not on the prompt. The same library prompt can be locked on one agent and unlocked on another. This gives you fine-grained control: a safety prompt might be locked on every agent that uses it, while a content generation prompt might be locked on production agents and unlocked on experimental ones.

Unlocking a prompt requires editing the agent and toggling the Lock Prompt switch off. Only users with the Admin or Create role can change the lock state. The change is logged in the audit trail.

Using the AI Assistant for prompt generation: what to expect

DAVE's Prompt Library includes an AI Assistant that generates a draft prompt from a narrative description. It is available when creating or editing a prompt in the library. Click AI Assistant, enter a description of what you want the prompt to accomplish, and click Generate. The system uses your tenant's default AI provider and model (configured under Admin → Settings → General → System Parameters → AI Assistant) to produce a structured prompt. If no default is configured, the system tries each available provider in order.

The AI Assistant is useful for:

Getting a structured starting point when you know what you want but are not sure how to phrase it as a system prompt.
Exploring alternative phrasings for an instruction you already have.
Generating a first draft quickly when time is short, with the intention of refining it before saving.

The AI Assistant is not a replacement for prompt engineering judgment. Generated prompts frequently need editing before they are production-ready. Common issues with generated prompts:

Vague instructions that sound correct but do not constrain the agent's behavior precisely enough.
Instructions that are correct in isolation but conflict with your organization's AI guidelines, which are prepended at runtime.
Overly long prompts that include unnecessary explanation or hedging.

Always review a generated prompt against the specific behavior you need before saving it to the library. Test it in a draft workflow with representative input before assigning it to a production agent. The AI Assistant accelerates the first draft; your judgment determines whether it is ready to use.

Frequently asked questions

Can I edit a prompt after saving it?

Yes. Prompts in the Prompt Library can be edited after creation. Unlike agent versions, prompts are not immutable. However, editing a prompt that is already assigned to one or more agents changes the behavior of those agents immediately: there is no versioning at the prompt level. If you need to change a prompt while keeping the old version available for comparison, create a new prompt with a modified name (for example, "Editorial Abstract Generator v2") rather than editing the existing one in place.

Does editing a library prompt automatically update the agents that use it?

Yes. When you edit a library prompt and save, every agent that has that prompt assigned will use the updated version at runtime. This is the primary advantage of the library over inline prompts: update once, propagate everywhere. It is also a risk: an edit that introduces a regression affects every agent using the prompt simultaneously. Test changes carefully, and consider creating a new prompt rather than editing an existing one if the change is significant.

What is the difference between a library prompt and an inline prompt?

An inline prompt is written directly in the agent's Config JSON or on the agent edit form. It belongs to that agent only and is versioned as part of the agent's version history. A library prompt is a standalone template in the Prompt Library, assignable to any agent, and managed independently of agent versions. Use library prompts when a prompt is shared across multiple agents or when you want to separate prompt management from agent configuration management. Use inline prompts for agent-specific instructions that are unlikely to be reused.

Can the same library prompt be assigned to multiple agents?

Yes. A single library prompt can be assigned to any number of agents. This is one of the primary use cases for the library: write a shared instruction once and assign it to every agent that needs it. When the instruction needs updating, edit the prompt once and all assigned agents use the updated version at their next execution.

What happens if I delete a library prompt that is assigned to an agent?

The help documentation does not explicitly describe the behavior when a library prompt is deleted while assigned to an agent. To avoid unexpected behavior, unassign a prompt from all agents before deleting it from the library.

The New Prompt button is missing. What do I do?

Your user account does not have the Admin or Create role for this tenant. The Curate role can view the Prompt Library but cannot create or edit prompts. Ask your tenant administrator to assign you the Admin or Create role under Admin → Settings → Users and Roles.

The AI Assistant button is not working or returns an error.

The AI Assistant uses your tenant's default AI provider and model, configured under Admin → Settings → General → System Parameters → AI Assistant. If no default provider is configured, the system tries each available provider in order. If all providers fail, the AI Assistant will return an error. Ask your administrator to configure a default AI provider, or write the prompt content manually.