How We Use Claude to Write Laravel Code 3x Faster at Softcrony

We’ve been using Claude as part of our Laravel development workflow at Softcrony for over a year. This is an honest account of what it’s changed, what it hasn’t, and the specific prompts and patterns that actually work.

Not theory. Not benchmarks. What we actually do every day.

What Changed When We Started Using Claude

The most significant change wasn’t speed — it was the quality floor. Before AI tools, the quality of code depended heavily on which developer wrote it and how much time they had. With Claude as a consistent reviewer and generator, the minimum quality of everything we ship went up.

Specific changes we measured over 6 months:

• Time to write a new CRUD module: 4 hours → 90 minutes
• Code review cycles per PR: 3.2 average → 1.8 average
• Bugs caught before staging: +40%
• Time spent writing tests: -60% (Claude writes most of them)
• Documentation coverage: went from ~20% to ~80%

Our Actual Workflow

Step 1 — Architecture First

Before writing any code, we use Claude to validate the architecture:

PROMPT:
"I'm building a multi-tenant SaaS application in Laravel 12.
Each company has users, projects, and tasks.
Users belong to one company.
Projects and tasks belong to companies, not directly to users.

I'm planning to implement tenancy using a 'company_id' column
on every table, with a global scope on each model.

Review this approach and identify:
1. Potential security issues with the global scope approach
2. Performance considerations at scale (10,000+ companies)
3. Alternative approaches I should consider
4. Any Laravel-specific pitfalls with this pattern"

Claude’s response typically identifies 2–3 issues we hadn’t considered. This 5-minute conversation saves hours of refactoring later.

Step 2 — Generate the Scaffold

Once architecture is agreed, we generate the base files:

PROMPT:
"Generate a complete Laravel 12 implementation for a Task model in a
multi-tenant SaaS. Requirements:

- Table: tasks (id, company_id, project_id, user_id, title, description,
  status enum[pending/in_progress/completed/cancelled], priority
  enum[low/medium/high/urgent], due_date, completed_at, timestamps)
- Soft deletes
- Global scope filtering by company_id (from Auth::user()->company_id)
- Relationships: belongsTo Project, belongsTo User (assignee), hasMany Comments
- Scopes: overdue(), byStatus(), byPriority(), assignedTo(User $user)
- Accessors: isOverdue (bool), statusLabel (string with proper labels)
- Cast due_date and completed_at as datetime
- Full PHPDoc comments on all methods

Follow Laravel 12 conventions. Use constructor property promotion.
No facades — use dependency injection."

Step 3 — Generate the Controller

PROMPT:
"Using the Task model above, generate a TaskController for a REST API.

Requirements:
- Full CRUD (index, store, show, update, destroy)
- index: paginate 20, filter by status/priority/assignee, search by title
- store/update: use Form Request classes (generate those too)
- All responses via TaskResource (generate that too)
- Policy authorization on all methods (generate TaskPolicy)
- Scoped to authenticated user's company (never expose other company's tasks)
- Return 404 if task exists but belongs to different company (not 403)
- PHPDoc on all methods

Generate: TaskController, StoreTaskRequest, UpdateTaskRequest,
TaskResource, TaskPolicy"

This single prompt generates 5 files in about 30 seconds. Review and adjustment takes another 10 minutes. The old way — writing all 5 from scratch — took 2–3 hours.

Step 4 — Generate Tests

PROMPT:
"Write PHPUnit feature tests for the TaskController above.

Test cases to cover:
- Unauthenticated requests return 401
- User cannot access tasks from another company (returns 404)
- index returns paginated tasks filtered to current company
- index filters work (status, priority, assignee, search)
- store creates task with correct company_id
- store validation errors return 422 with correct structure
- show returns correct task
- update changes only provided fields
- destroy soft-deletes and returns 204
- Policy: only company members can view/edit tasks

Use RefreshDatabase trait.
Use factories for test data.
Group related tests in it() blocks.
Include edge cases."

Prompts That Work Consistently

Code Review Prompt

"Review this Laravel code for:
1. Security vulnerabilities (SQL injection, mass assignment, authorization bypass)
2. N+1 query problems
3. Missing validation
4. Performance issues
5. Laravel anti-patterns
6. Any logic bugs

For each issue: describe the problem, show the problematic code,
provide the fix. Rate severity: Critical / High / Medium / Low.

[paste your code here]"

Database Query Optimization

"This Eloquent query is slow on a table with 500,000 rows.
The query takes 8 seconds. Help me optimize it.

[paste query]

Current indexes:
[paste SHOW INDEX FROM table output]

EXPLAIN output:
[paste EXPLAIN result]

Suggest: query rewrites, new indexes, or architectural changes.
Show the optimized version."

Migration from Old to New Pattern

"Refactor this Laravel code from the old pattern to the modern Laravel 12 way.

Old patterns to modernize:
- Replace facade usage with dependency injection
- Replace array syntax [] with named arguments where appropriate
- Add proper return types to all methods
- Replace string route names with route() helper
- Use constructor property promotion
- Add PHPDoc where missing

[paste code]

Preserve all existing functionality. Only modernize the code style."

Where Claude Still Needs Human Oversight

Claude is not a replacement for developer judgment. These are the areas where we always review carefully:

Business logic. Claude doesn’t know your business rules. It will write logically correct code that does the wrong thing if your prompt doesn’t capture every rule. Always review business logic outputs carefully.

Database migrations. Claude generates correct migrations for the schema described — but it doesn’t know about your existing data. Always review migrations manually before running on staging.

Security-sensitive code. Authentication flows, payment processing, permission systems — Claude’s output is a starting point, not a final answer. Get a human security review.

Third-party API integrations. Claude’s training data has a cutoff. API signatures, authentication methods, and rate limits may have changed. Always verify against current documentation.

Our Prompt Engineering Principles

Be specific about constraints. “Write a Laravel controller” gets mediocre output. “Write a Laravel 12 API controller using dependency injection, no facades, with PHPDoc, returning API Resources, with Policy authorization” gets excellent output.

Specify what you don’t want. “No static methods. No global functions. No God classes.” Claude respects negative constraints well.

Give context about your stack. Start conversations with: “We’re using Laravel 12, PHP 8.3, MySQL 8, Redis, the Repository pattern, and Sanctum for API auth.” This shapes every subsequent response in the conversation.

Ask for the reasoning. “Explain why you chose this approach and what alternatives you considered” — this helps you evaluate whether Claude’s choice is actually right for your context.

Iterate, don’t regenerate. “The index method is good but the store method needs to also send a notification via Queue when the task is assigned — add that” is faster than rewriting the whole prompt.

Cost for a Development Team

Claude Pro costs $20/month per seat. For a 3-person Laravel team:

If your team is building Laravel applications and wants to integrate AI into your development workflow effectively, our team at Softcrony is happy to share more of what we’ve learned.