Open Standard

Capability Schema

The atomic unit of the match graph. Every AI artifact — products, agents, MCPs, APIs, models, repos — decomposed into discrete capabilities that can be searched, compared, and matched to human intent.

Why capabilities, not products

Cursor isn't one thing — it's 15 capabilities: codebase-aware code completion, multi-file editing, AI terminal, tab completion, and more. Each capability maps to intents it serves and conditions under which it works best.

One product has many capabilities. One capability is served by many artifacts. The match graph connects intents → capabilities → artifacts with confidence scores that improve with every interaction.

The schema

{
  // Identity
  "id": "cap_abc123",
  "name": "codebase-aware code completion",
  "description": "Completes code with full awareness of the project's
    codebase, including imports, types, and cross-file references.
    Uses RAG over indexed project files with 200K context window.",

  // What user needs this addresses
  "intentsServed": [
    "I need to write code faster",
    "I want AI that understands my whole project",
    "code completion that knows my types"
  ],

  // Under what conditions this works
  "conditions": {
    "bestFor": ["experienced developers", "teams under 20"],
    "limitations": ["struggles with monorepos over 100K lines"],
    "requires": ["VS Code or terminal access"],
    "inputTypes": ["code", "text"],
    "outputTypes": ["code"]
  },

  // Which artifact provides this
  "artifactId": "cursor",
  "artifactName": "Cursor",
  "artifactType": "product",
  "artifactUrl": "https://cursor.sh",

  // Match quality (improves with feedback)
  "matchConfidence": 0.87,
  "timesMatched": 1247,
  "successRate": 0.82,

  // Classification
  "categories": ["code-generation-editing"],
  "tags": ["ide", "autocomplete", "context-aware"]
}

Field reference

namestringHuman-readable capability name. Consistent across artifacts — the same capability from different tools uses the same name for match graph linking.

descriptionstringArchitecturally deep description. Minimum 100 characters. Describes HOW it works, not just what it claims to do.

intentsServedstring[]User intents this capability addresses. Written in first-person natural language — the way a developer would actually phrase their need.

conditions.bestForstring[]Who gets the most value. Experience level, team size, use case, platform.

conditions.limitationsstring[]Known failure modes. Honest architectural constraints.

conditions.requiresstring[]Prerequisites. Platform, API key, hardware, software dependencies.

conditions.inputTypesstring[]What it accepts. Standard types: text, code, image, audio, video, pdf, url, structured-data, file.

conditions.outputTypesstring[]What it produces. Same type vocabulary as inputTypes.

matchConfidencenumber (0-1)How reliably this capability satisfies the intents it claims. Starts at 0.5, improves with match feedback.

categoriesstring[]Standard capability categories (see taxonomy below). Enables consistent cross-artifact comparison.

Capability taxonomy

Ten standard categories for consistent cross-artifact comparison. Every capability maps to one or more of these.

Code Generation & Editing

code-generation-editing

Generating, completing, editing, refactoring, or reviewing code. Includes codebase-aware completion, multi-file editing, code explanation, and test generation.

codebase-aware code completionmulti-file refactoringtest generation from codecode review with explanations

Text Generation & Language

text-generation-language

Generating, summarizing, translating, or transforming text. Includes creative writing, technical documentation, email drafting, and content repurposing.

long-form article generationdocument summarizationlanguage translationemail drafting from context

Image & Visual

image-visual

Generating, editing, analyzing, or transforming images and visual content. Includes text-to-image, inpainting, style transfer, OCR, and visual QA.

text-to-image generationbackground removalimage inpaintingvisual question answering

Data Processing & Analysis

data-processing-analysis

Extracting, transforming, analyzing, or visualizing structured data. Includes CSV/JSON processing, SQL generation, chart creation, and statistical analysis.

natural language to SQLCSV data analysischart generation from dataentity extraction from text

Tool Use & Integration

tool-use-integration

Using external tools, calling APIs, connecting services, or orchestrating multi-service workflows. Includes function calling, MCP tools, and API orchestration.

function calling via JSON schemaMCP tool integrationmulti-API orchestrationwebhook handling

Memory & Knowledge

memory-knowledge

Storing, retrieving, and reasoning over information across sessions. Includes RAG, conversation memory, knowledge graphs, and fact extraction.

retrieval-augmented generationconversation memory across sessionsknowledge base constructionfact extraction and storage

Planning & Reasoning

planning-reasoning

Multi-step planning, logical reasoning, decision-making, and task decomposition. Includes agent planning loops, chain-of-thought, and tree-of-thought.

multi-step task planningchain-of-thought reasoningcode-level debuggingdecision tree traversal

Automation & Workflow

automation-workflow

Automating repetitive tasks, building workflows, scheduling jobs, and orchestrating multi-step processes without human intervention.

scheduled data pipelineemail automationCI/CD integrationmulti-step form processing

Search & Retrieval

search-retrieval

Finding, ranking, and surfacing relevant information. Includes semantic search, vector similarity, web search, and recommendation systems.

semantic code searchweb search with citationssimilar document retrievalrecommendation engine

Safety & Moderation

safety-moderation

Content moderation, safety filtering, bias detection, prompt injection defense, and output guardrails.

content safety classificationprompt injection detectionPII redactionoutput guardrails and validation

Artifact types

The match graph indexes 19 artifact types — every form factor an AI capability can take.

Product

productCommercial SaaS

Agent

agentAutonomous AI system

MCP Server

mcpModel Context Protocol

API

apiProgrammatic endpoint

Repository

repoOpen-source project

Prompt

promptTemplate / instruction

Skill

skillPlugin capability

Extension

extensionBrowser / IDE

Workflow

workflowAutomation sequence

Model

modelFoundation / specialized

Dataset

datasetTraining / eval data

Template

templateProject starter

App

appNative application

Web App

webappBrowser-based tool

Framework

frameworkSDK / library

CLI Tool

cliCommand-line tool

Fine-tune

finetuneLoRA / adapter

Benchmark

benchmarkEvaluation suite

Platform

platformHosting infrastructure

How the match graph uses capabilities

Every search produces a match record: intent + context → capability → artifact → outcome. Over time, the graph learns which capabilities solve which intents for which user profiles. The matchConfidence field on each capability is the graph's current best estimate of how reliably this capability satisfies the intents it claims.

Gap records track intents with no strong match — demand signals for builders showing what capabilities the ecosystem is missing. The schema is designed so that every interaction makes the graph smarter, and every gap signals an opportunity.

Add your artifact to the graph

Submit your AI artifact and it will be automatically decomposed into capabilities using this schema. Claim it to see match analytics in Studio.

Submit an artifact

JSON-LD context

Reference this context when producing structured capability data:

{
  "@context": {
    "@vocab": "https://unfragile.ai/schema#",
    "capability": "https://unfragile.ai/schema#Capability",
    "artifact": "https://unfragile.ai/schema#Artifact",
    "matchRecord": "https://unfragile.ai/schema#MatchRecord",
    "name": "schema:name",
    "description": "schema:description",
    "intentsServed": "schema:potentialAction",
    "conditions": "schema:eligibleRegion",
    "matchConfidence": "schema:ratingValue"
  }
}

Why capabilities, not products

The schema

{ // Identity "id": "cap_abc123", "name": "codebase-aware code completion", "description": "Completes code with full awareness of the project's codebase, including imports, types, and cross-file references. Uses RAG over indexed project files with 200K context window.", // What user needs this addresses "intentsServed": [ "I need to write code faster", "I want AI that understands my whole project", "code completion that knows my types" ], // Under what conditions this works "conditions": { "bestFor": ["experienced developers", "teams under 20"], "limitations": ["struggles with monorepos over 100K lines"], "requires": ["VS Code or terminal access"], "inputTypes": ["code", "text"], "outputTypes": ["code"] }, // Which artifact provides this "artifactId": "cursor", "artifactName": "Cursor", "artifactType": "product", "artifactUrl": "https://cursor.sh", // Match quality (improves with feedback) "matchConfidence": 0.87, "timesMatched": 1247, "successRate": 0.82, // Classification "categories": ["code-generation-editing"], "tags": ["ide", "autocomplete", "context-aware"] }