Laravel AI SDK integration

Use Prompt Deck with Laravel AI SDK agents for automatic prompt loading, version tracking, and performance monitoring.

Introduction

Prompt Deck provides first-class, optional integration with the Laravel AI SDK. When the AI SDK is installed, you get:

• Automatic prompt scaffolding — Running make:agent automatically creates a matching prompt directory.
• The HasPromptTemplate trait — Provides instructions() and promptMessages() methods that load versioned prompts directly into your AI agents.
• The TrackPromptMiddleware — Automatically records prompt executions (tokens, latency, model, etc.) using Prompt Deck’s tracking system.

All of this is entirely optional. Prompt Deck works perfectly without the AI SDK.

Installation

Prompt Deck does not require the AI SDK — it’s listed as a suggest dependency. Install it when you’re ready:

composer require laravel/ai

Once laravel/ai is installed, Prompt Deck’s AI SDK features activate automatically. No additional configuration is needed.

Automatic prompt scaffolding

When the Laravel AI SDK is installed, Prompt Deck automatically hooks into the make:agent command. Whenever you create a new agent:

php artisan make:agent SalesCoach

Prompt Deck detects the successful command and automatically runs:

php artisan make:prompt sales-coach

This creates a versioned prompt directory ready for the agent to use via the HasPromptTemplate trait — zero extra setup required.

How it works

Prompt Deck registers a listener (AfterMakeAgent) on Laravel’s CommandFinished event. When make:agent completes successfully, the listener:

1. Extracts the agent name from the command input.
2. Converts it to kebab-case (SalesCoach → sales-coach).
3. Strips any namespace prefix (App\Ai\Agents\SalesCoach → sales-coach).
4. Checks if the prompt already exists (skips if it does).
5. Runs make:prompt with the derived name.

The listener is only registered when laravel/ai is installed. If the prompt creation fails for any reason, it does not break the make:agent workflow — the agent is still created successfully.

Example output

$ php artisan make:agent SalesCoach

   INFO  Agent [app/Ai/Agents/SalesCoach.php] created successfully.

PromptDeck: Created prompt sales-coach for SalesCoach.

Disabling auto-scaffolding

To disable automatic prompt scaffolding, set the configuration option:

// config/prompt-deck.php
'scaffold_on_make_agent' => false,

Or via environment variable:

PROMPTDECK_SCAFFOLD_ON_MAKE_AGENT=false

Quick start

Use the HasPromptTemplate trait on any agent class:

<?php

namespace App\Ai\Agents;

use Laravel\Ai\Contracts\Agent;
use Laravel\Ai\Promptable;
use Veeqtoh\PromptDeck\Concerns\HasPromptTemplate;

class SalesCoach implements Agent
{
    use Promptable, HasPromptTemplate;

    // instructions() is provided automatically by HasPromptTemplate.
    // It loads prompts/sales-coach/v{active}/system.md
}

That’s it. The HasPromptTemplate trait provides the instructions() method required by the Agent contract, loading the system prompt from your Prompt Deck files.

The HasPromptTemplate trait

The HasPromptTemplate trait bridges Prompt Deck’s file-based templates with the Laravel AI SDK’s agent contracts.

How it maps to AI SDK contracts

Prompt Deck	AI SDK	Description
system.md role file	instructions()	Agent’s system prompt.
user.md, assistant.md, etc.	messages() via promptMessages()	Conversation context.
metadata.json	—	Prompt metadata (description, variables, etc.).
v1/, v2/, etc.	—	Version management.

Mapping diagram

prompts/sales-coach/
├── v1/
│   ├── system.md          → instructions()
│   ├── user.md            → promptMessages()
│   └── metadata.json
└── v2/
    ├── system.md          → instructions()  (active version)
    ├── user.md            → promptMessages()
    ├── assistant.md       → promptMessages()
    └── metadata.json

Customising the prompt

Prompt name

By default, the prompt name is derived from the class name in kebab-case:

• SalesCoach → sales-coach
• DocumentAnalyzer → document-analyzer

Override promptName() to use a custom name:

class SalesCoach implements Agent
{
    use Promptable, HasPromptTemplate;

    public function promptName(): string
    {
        return 'coaching/sales'; // loads from prompts/coaching/sales/
    }
}

Pinning a version

By default, the active version is loaded. Pin to a specific version by overriding promptVersion():

public function promptVersion(): ?int
{
    return 2; // Always use v2
}

Return null (the default) to always load the active version — useful for A/B testing and gradual rollouts.

Variable interpolation

Pass dynamic values into your prompt templates by overriding promptVariables():

class SalesCoach implements Agent
{
    use Promptable, HasPromptTemplate;

    public function __construct(public User $user) {}

    public function promptVariables(): array
    {
        return [
            'user_name' => $this->user->name,
            'company'   => $this->user->company,
        ];
    }
}

In your system.md:

You are a sales coach for {{ $company }}.
You are helping {{ $user_name }} improve their technique.

Variables are interpolated into all roles (system, user, assistant, etc.) when accessed via instructions() or promptMessages().

Full agent example

Here’s a complete agent using all Prompt Deck features with the AI SDK:

<?php

namespace App\Ai\Agents;

use App\Ai\Tools\RetrievePreviousTranscripts;
use App\Models\User;
use Illuminate\Contracts\JsonSchema\JsonSchema;
use Laravel\Ai\Contracts\Agent;
use Laravel\Ai\Contracts\Conversational;
use Laravel\Ai\Contracts\HasMiddleware;
use Laravel\Ai\Contracts\HasStructuredOutput;
use Laravel\Ai\Contracts\HasTools;
use Laravel\Ai\Promptable;
use Veeqtoh\PromptDeck\Ai\TrackPromptMiddleware;
use Veeqtoh\PromptDeck\Concerns\HasPromptTemplate;

class SalesCoach implements Agent, Conversational, HasTools, HasStructuredOutput, HasMiddleware
{
    use Promptable, HasPromptTemplate;

    public function __construct(public User $user) {}

    // instructions() is provided by HasPromptTemplate — no need to define it.

    /**
     * Dynamic variables injected into the prompt template.
     */
    public function promptVariables(): array
    {
        return [
            'user_name' => $this->user->name,
            'company'   => $this->user->company,
        ];
    }

    /**
     * Conversation context from the prompt template plus history.
     */
    public function messages(): iterable
    {
        return $this->promptMessages();
    }

    /**
     * Tools available to the agent.
     */
    public function tools(): iterable
    {
        return [
            new RetrievePreviousTranscripts,
        ];
    }

    /**
     * Structured output schema.
     */
    public function schema(JsonSchema $schema): array
    {
        return [
            'feedback' => $schema->string()->required(),
            'score'    => $schema->integer()->min(1)->max(10)->required(),
        ];
    }

    /**
     * Middleware for automatic tracking.
     */
    public function middleware(): array
    {
        return [
            new TrackPromptMiddleware,
        ];
    }
}

Create and populate the prompt files:

php artisan make:prompt sales-coach --user

Edit prompts/sales-coach/v1/system.md:

You are a sales coach for {{ $company }}.
You are helping {{ $user_name }} improve their sales technique.

Analyse transcripts carefully and provide:
- Specific, actionable feedback
- A score from 1-10

Conversation context

If your agent implements Conversational, you can load pre-defined conversation context from Prompt Deck role files using the promptMessages() method.

Loading all non-system roles

By default, promptMessages() returns all roles except system (which goes through instructions()):

public function messages(): iterable
{
    // Returns Message[] for all non-system roles (user, assistant, etc.)
    return $this->promptMessages();
}

Limiting to specific roles

Pass an array to limit which roles are included:

public function messages(): iterable
{
    return $this->promptMessages(['user']);
}

Merging with database history

Combine template messages with conversation history from your database:

use Laravel\Ai\Messages\Message;

public function messages(): iterable
{
    // Pre-defined context from the prompt template.
    $context = $this->promptMessages();

    // Plus conversation history from the database.
    $history = History::where('user_id', $this->user->id)
        ->latest()
        ->limit(50)
        ->get()
        ->reverse()
        ->map(fn ($m) => new Message($m->role, $m->content))
        ->all();

    return array_merge($context, $history);
}

Performance tracking middleware

The TrackPromptMiddleware automatically records prompt executions via Prompt Deck’s tracking system.

Setting up the middleware

1. Enable tracking

   Set the tracking configuration in config/prompt-deck.php:

   'tracking' => [
       'enabled'    => true,
       'connection' => null, // uses default database connection
   ],
   
2. Publish and run migrations

   php artisan vendor:publish --tag=prompt-deck-migrations
   php artisan migrate
   
3. Add middleware to your agent

   use Laravel\Ai\Contracts\HasMiddleware;
   use Veeqtoh\PromptDeck\Ai\TrackPromptMiddleware;
   
   class SalesCoach implements Agent, HasMiddleware
   {
       use Promptable, HasPromptTemplate;
   
       public function middleware(): array
       {
           return [
               new TrackPromptMiddleware,
           ];
       }
   }
   

What gets tracked

The middleware automatically records the following fields to the prompt_executions table:

Field	Source
prompt_name	Agent’s promptName() method.
prompt_version	Resolved template version number.
input	The user’s prompt text from the AgentPrompt.
output	The AI response text.
tokens	Total token usage from the response.
latency_ms	Round-trip time in milliseconds (measured via hrtime).
model	Model used (e.g. gpt-4o, claude-3-sonnet).
provider	Provider name (e.g. openai, anthropic).

How it works internally

The middleware:

1. Records the start time before the request using hrtime(true).
2. Passes the prompt to the next middleware in the pipeline.
3. Uses the response’s then() hook to record execution data after the response completes.
4. Calls PromptManager::track() with the collected data.

The middleware only tracks agents that use the HasPromptTemplate trait. If the agent doesn’t have a promptName() method, the tracking is silently skipped.

Accessing the template directly

You can access the full PromptTemplate object from within your agent for advanced use cases:

// Get the template instance.
$template = $this->promptTemplate();

// Check available roles.
$template->roles();       // ['system', 'user', 'assistant']
$template->has('skill');  // false

// Get raw content (no interpolation).
$template->raw('system');

// Get the resolved version.
$template->version();     // 2

// Get prompt metadata.
$template->metadata();    // ['description' => '...', ...]

The template instance is cached for the lifetime of the agent object, so repeated calls to promptTemplate() don’t incur additional filesystem or cache lookups.

Clearing the cached template

Clear the cached template to force a fresh load on next access:

$agent->forgetPromptTemplate();

This is useful in:

• Long-running processes (queue workers, daemons) where prompts might change between jobs.
• Tests where you switch prompt versions between assertions.

The method returns $this for fluent chaining:

$agent->forgetPromptTemplate()->promptTemplate(); // fresh load

Without the AI SDK

The HasPromptTemplate trait works even without laravel/ai installed. The instructions() method simply returns a string, and promptMessages() falls back to returning raw arrays instead of AI SDK Message objects:

// Without laravel/ai — returns array
$messages = $agent->promptMessages();
// [['role' => 'user', 'content' => '...'], ...]

// With laravel/ai — returns Message[]
$messages = $agent->promptMessages();
// [Message('user', '...'), ...]

This allows you to use Prompt Deck’s template loading in any context, not just with the AI SDK.