Preskoči na vsebino

Generating Schemas with an LLM

Almadar schemas are structured JSON — and that structure makes them ideal targets for LLM generation. You describe your application in plain language; the LLM outputs a valid .orb schema.

This tutorial covers:

  • Installing and using the Almadar skill
  • Prompting an LLM to generate a complete schema
  • Validating the output
  • Fixing the most common mistakes LLMs make
Entity(Matter)Page(Space)Trait(Energy)idleactivehas_traitrenderstransition
Orbital Unit = Entity + Traits + Pages

The Almadar Skill​

The @almadar/skills package includes a Claude Code skill that teaches the LLM the full Almadar language specification — orbitals, entities, traits, state machines, patterns, S-expressions, and more.

Install​

npm install -g @almadar/skills

Then install the Claude Code skill:

almadar-skills install almadar-orbitals

Or use it directly in Claude Code by referencing the skill file in your session.

The Generation Workflow​

1. Describe your application in plain language
       ↓
2. LLM decomposes it into orbitals (one per entity domain)
       ↓
3. LLM generates: entity + traits + state machines + pages
       ↓
4. Validate: almadar validate schema.orb
       ↓
5. Fix any errors, iterate
       ↓
6. Run: almadar dev

How to Prompt for a Schema​

The Decomposition Prompt​

Start by asking the LLM to decompose your application into orbitals before writing JSON:

I want to build a project management app with:
- Projects (name, description, status: active/archived)
- Tasks (title, priority, assignee, due date, linked to a project)
- Users (name, email, role: admin/member)

Tasks can be created, edited, deleted, and moved through states:
todo → in progress → review → done.

When a task is completed, the project's completed task count should update.

Please decompose this into Almadar orbitals and generate a complete schema.

What to Include in Your Prompt​

A good generation prompt covers:

ElementExample
Entities"Tasks have title, priority (low/medium/high), due date, assignee"
Persistence"Tasks are persistent (stored in DB), cart is runtime (session only)"
Workflows"Tasks move from todo → in progress → review → done"
UI behaviors"Users can list, create, edit, and delete tasks on a /tasks page"
Business rules"Only the assignee can mark a task as done"
Connections"When a task is completed, update the project's counter"
Pages needed"I need /tasks, /projects, and /users routes"

What the LLM Should Produce​

For each orbital, the LLM should output all four required parts:

{
  "name": "AppName",
  "version": "1.0.0",
  "orbitals": [
    {
      "name": "OrbitalName",
      "entity": {
        "name": "EntityName",
        "persistence": "persistent",
        "collection": "collection_name",
        "fields": [...]
      },
      "traits": [
        {
          "name": "TraitName",
          "linkedEntity": "EntityName",
          "category": "interaction",
          "stateMachine": {
            "states": [...],
            "events": [...],
            "transitions": [...]
          }
        }
      ],
      "pages": [
        {
          "name": "PageName",
          "path": "/route",
          "traits": [{ "ref": "TraitName", "linkedEntity": "EntityName" }]
        }
      ]
    }
  ]
}

The Most Common LLM Mistakes​

LLMs that don't have the Almadar skill loaded will make predictable mistakes. Learn to spot them.

1. Missing pages (most common)​

The LLM generates entity + traits but forgets the pages array entirely.

// ❌ Incomplete — no pages
{
  "name": "TaskManager",
  "orbitals": [{
    "name": "Tasks",
    "entity": { ... },
    "traits": [ { "name": "TaskCRUD", ... } ]
  }]
}

// ✅ Add pages
{
  "name": "TaskManager",
  "orbitals": [{
    "name": "Tasks",
    "entity": { ... },
    "traits": [ { "name": "TaskCRUD", ... } ],
    "pages": [
      { "name": "TaskListPage", "path": "/tasks", "traits": [{ "ref": "TaskCRUD", "linkedEntity": "Task" }] }
    ]
  }]
}

Fix prompt: "The schema is missing the pages array for each orbital. Please add pages with path and traits[].ref for every orbital."

2. States as strings instead of objects​

// ❌ Wrong
"states": ["Pending", "InProgress", "Done"]

// ✅ Correct
"states": [
  { "name": "Pending", "isInitial": true },
  { "name": "InProgress" },
  { "name": "Done", "isTerminal": true }
]

Fix prompt: "States must be objects with a 'name' property. The initial state needs 'isInitial': true. Terminal states need 'isTerminal': true."

3. Missing INIT transition​

The page loads but renders nothing because there's no INIT self-loop with render-ui.

// ❌ No INIT — page is blank
"transitions": [
  { "from": "Pending", "event": "COMPLETE", "to": "Done", "effects": [...] }
]

// ✅ Add INIT
"transitions": [
  {
    "from": "Pending", "event": "INIT", "to": "Pending",
    "effects": [
      ["fetch", "Task"],
      ["render-ui", "main", { "type": "entity-table", "entity": "Task" }]
    ]
  },
  { "from": "Pending", "event": "COMPLETE", "to": "Done", "effects": [...] }
]

Fix prompt: "Every interaction trait needs an INIT transition (self-loop) that fires render-ui to render the initial UI. Without it the page will be blank."

4. Using deprecated action props​

// ❌ Deprecated — these will fail validation
{ "type": "form-section", "onSubmit": "SAVE", "onCancel": "CANCEL" }

// ✅ Correct
{ "type": "form-section", "submitEvent": "SAVE", "cancelEvent": "CANCEL" }
// ❌ Deprecated
{ "type": "page-header", "headerActions": [...] }

// ✅ Correct
{ "type": "page-header", "actions": [...] }

5. Schema-level traits array (wrong structure)​

// ❌ Wrong — traits at the root level (legacy format)
{
  "name": "App",
  "traits": [...],
  "pages": [...]
}

// ✅ Correct — traits live inside orbitals
{
  "name": "App",
  "orbitals": [{
    "name": "FeatureName",
    "entity": { ... },
    "traits": [...],
    "pages": [...]
  }]
}

6. Missing linkedEntity on trait​

// ❌ Missing linkedEntity
{ "name": "TaskCRUD", "category": "interaction", "stateMachine": { ... } }

// ✅ Correct
{ "name": "TaskCRUD", "linkedEntity": "Task", "category": "interaction", "stateMachine": { ... } }

Iterative Generation for Large Apps​

For applications with more than 3-4 orbitals, generate one orbital at a time:

Step 1: "Generate only the TaskManager orbital (entity + traits + pages)"
Step 2: "Now add the ProjectManager orbital that listens to TASK_COMPLETED from TaskManager"
Step 3: "Now add the UserManager orbital for browsing users"
Step 4: "Combine all three orbitals into a single schema"

This reduces errors and makes each piece reviewable before assembling.

Validating the Output​

Always validate before running:

almadar validate schema.orb

Common validation errors and what they mean:

ErrorCause
Missing initial stateNo state has "isInitial": true
Unknown event in transitionA transition references an event key not in the events array
Missing pagesAn orbital has traits but no pages array
Invalid pattern typeThe type in a render-ui effect is not a valid pattern name
Deprecated propUsing onSubmit instead of submitEvent, etc.
Circular dependencyTwo orbitals listen to each other (use a third mediator orbital)

A Reference Prompt (Copy & Use)​

This prompt works well with the Almadar skill installed:

Using the Almadar language, generate a complete .orb schema for: [YOUR APP DESCRIPTION]

Requirements:
- Each feature domain becomes one orbital with: entity, traits, pages
- Every trait must have an INIT self-loop transition that renders the initial UI using render-ui
- States must be objects: { "name": "StateName", "isInitial": true }
- Pages must be present with path and traits[].ref wiring
- Use "submitEvent"/"cancelEvent" on form-section (not onSubmit/onCancel)
- Use "actions" on page-header (not headerActions)
- All traits belong inside orbitals — there is no schema-level traits array

Entities needed: [LIST ENTITIES]
Workflows: [DESCRIBE STATE TRANSITIONS]
Pages needed: [LIST ROUTES]
Business rules / permissions: [DESCRIBE GUARDS]
Cross-orbital connections: [DESCRIBE EMITS/LISTENS IF ANY]

Next Steps​