At Lubus, we’ve been using Frappe for a while now.
It powers our HR, CRM, and Helpdesk. We also use it when working with businesses that want fewer tools and cleaner processes. Over time, that meant building custom apps, writing integrations, and shaping Frappe around real work, not examples.
As things grew, so did the amount of repeat effort. Setting things up. Wiring the same pieces and explaining the same patterns. That’s when automation and AI naturally entered the picture.
When Coding Agents Joined the Team
Coding agents started helping with everyday tasks, exploring approaches, scaffolding features, and cleaning up code.
They were useful. But something was off. They didn’t understand Frappe the way someone who’s worked with it daily does. They mixed in ideas from other frameworks. They guessed at patterns that don’t really exist. They wrote code that worked, but didn’t fit.
We kept fixing the same things. At some point, it felt less like helping an agent and more like training one repeatedly.
Context Was the Missing Piece
The problem wasn’t the agents. It was the lack of a clear context. Frappe has its own way of doing things. Once you learn it, it’s consistent. But that consistency isn’t always obvious from scattered docs or code samples. So we started writing things down not as tutorials, but as explanations of how we expect things to work.
A Familiar Idea from WordPress
Coming from the WordPress ecosystem, we noticed something interesting there. They recently introduced Agent Skills, short, focused documents written specifically for Frappe Skills for AI Agents. Not guides for beginners. Not marketing pages. Just clear instructions. That approach made sense. We didn’t want to replicate it blindly. We wanted to apply the idea to Frappe in a way that reflected how we actually build.
What Are Frappe Skills for AI Agents?
Frappe Skills are focused notes designed specifically for coding agents. They are based on the Agent Skills specification introduced by Anthropic and adapted to reflect how we actually build with Frappe. Each skill clarifies how Frappe is structured, how its components connect, which patterns are expected, and which shortcuts tend to create problems. They aren’t meant to teach everything, just to provide enough context so agents can work within the system without guessing.
How Development Flow Fits In
One thing we learned early was that context alone isn’t enough. Agents also need a stable place to work.
For local development, we’ve been using Frappe Manager from our friends at rtCamp.It gives us a predictable Docker-based setup where custom apps can be mounted directly into the environment, making iteration straightforward. Changes reflect immediately, testing is consistent, and resetting the system is easy when things go wrong.
That stability matters. The skills assume this kind of structured environment where apps follow expected patterns and development mirrors production closely. When agents understand both how Frappe works and the environment they’re operating in, they become far more consistent. Less improvising. Fewer surprises.
What These Skills Cover
The skills reflect how we build with Frappe in production. They span project triage, app scaffolding, DocType and API development, Desk customization, modern Vue frontends, reports and web forms, testing, Docker-based environments with Frappe Manager, and enterprise patterns for real business applications.
Each skill follows a clear structure when to use it, required inputs, step-by-step procedure, verification, and common failure modes. Together, they give agents a practical, version-aware mental model of Frappe aligned with real-world standards rather than outdated or generic patterns.
Putting the Skills to Work
Frappe Skills aren’t theoretical. They’re meant to improve how we build with AI every day. The real question became, how do we know the skills are actually good? You can review documents. You can refine wording. But there’s only one reliable way to test them: use them to build something real and see where they fail.
At first, we considered validating them with a simple CRUD app. That would have been easy. But it wouldn’t stretch the skills very far. A basic example doesn’t surface edge cases. It doesn’t test integrations, permissions, data flow, or UI behavior in meaningful ways.
So we decided to go further. Instead of a demo app, we picked a real idea from our internal list, a password and secrets manager built on Frappe. We’re calling it Frappe Vault. The goal isn’t just to ship another app. It’s to use the skills to build something practical, slightly complex, and grounded in real-world needs.
This creates a feedback loop:
- The skills guide the AI agents as they build Frappe Vault.
- The friction we observe while building improves our skills.
- The improved skills make the next iteration cleaner.
Skills shape the app. The app sharpens the skills.
Right now, Frappe Vault is experimental. We’re actively refining it as we test workflows, permissions, encryption handling, and UI behavior. Once it reaches a stable MVP, we plan to open-source it. This way, Frappe Skills don’t remain static documents. They evolve through real use, real constraints, and real mistakes exactly the way good systems should.
What Comes Next
This is an early step. We’ll keep adding skills as we run into new problems, see new patterns emerge, and learn where agents still struggle. If you’re building serious Frappe apps with AI in the loop, these skills are meant for you. You can check out the repository to see what is available so far. If you find them useful, star the project on GitHub to follow our updates, or dive straight in and try the skill.
Leave a Reply