Skills & Workflow Authoring

Skills are the foundation of organizational intelligence in Aigenzey Cowork. When a team member solves a complex task or refines an operating procedure, that workflow can be codified into a reusable, versioned Skill that the entire organization can execute with consistent quality.

What are Skills?

In Aigenzey Cowork, a Skill is a standardized agent capability package combining:

• Intent Triggers: High-precision semantic and keyword trigger patterns that tell the orchestrator when the skill applies.
• Procedural Guidelines: Step-by-step instructions, edge-case handling, and validation rules that guide the agent's reasoning.
• Required Tool Bindings: Declarations of necessary OAuth connectors, database connections, or MCP endpoints required to execute the workflow.
• Parameter Schemas: Structured inputs (e.g. date ranges, client IDs, file paths) that the agent extracts from conversation context or prompts the user to provide.

Progressive Skill Discovery

Rather than cluttering every prompt with hundreds of skill instructions, Aigenzey Cowork employs progressive disclosure via Google ADK:

1. Upfront Index Search: When a user sends a prompt, the orchestrator queries a derived search index (combining PostgreSQL full-text search with vector embeddings using Reciprocal Rank Fusion - RRF).
2. Dynamic Tool Attachment: Strong matching skills are dynamically attached to the agent as active tools, while the catalog overview is supplied for discovery.
3. Autonomous Intent Evaluation: The agent evaluates whether a matched skill accurately fits the user's intent before invoking it, avoiding false-positive executions.

Guided Skill Builder

Cowork provides three distinct pathways to author new skills:

1. One-Click Save from Chat

After the agent successfully completes a complex workflow in chat, click the Save as Skill button beneath the message. The builder analyzes the reasoning trace, extracts the tool sequence, and auto-generates a clean skill package.

2. Guided Skill Builder (Interview Mode)

Navigate to Skills Library → New Skill → Guided Builder. The assistant interviews you in natural language:

• What business problem does this skill solve?
• Which systems (e.g. Salesforce, Jira, Snowflake) does it need to connect to?
• What are the mandatory input parameters and output expectations?
• What edge cases or validation rules should the agent enforce?

3. Visual Skill Editor

Power users and developers can write and edit skill definitions directly using the visual Markdown and YAML code editor with live linting and schema validation.

Skill Package Anatomy

Under the hood, skills adhere to a deterministic schema:

---
name: customer-churn-risk-auditor
description: Scans Zendesk support escalations, checks Stripe subscription status, and computes customer churn risk scores.
version: 1.2.0
tags:
  capability: analytics
  domain: customer-success
connectors:
  - zendesk
  - stripe
  - slack
inputs:
  timeframe_days:
    type: integer
    default: 14
    description: Number of trailing days to audit for ticket escalations.
  min_mrr_threshold:
    type: number
    default: 1000
    description: Minimum Monthly Recurring Revenue (MRR) to flag as high-priority.
---

# Customer Churn Risk Audit Procedure

## Step 1: Ingest Zendesk Escalations
Search Zendesk for tickets with priority 'urgent' or tag 'escalation' updated within the last {timeframe_days} days.

## Step 2: Correlate Customer Revenue in Stripe
For each affected organization, retrieve active subscription details from Stripe to extract current MRR.

## Step 3: Compute Churn Risk Score
Calculate risk weighting based on escalation severity and account MRR. If MRR > {min_mrr_threshold} and ticket sentiment is negative, flag as CRITICAL.

## Step 4: Output Generation
Generate an executive summary artifact with a sorted priority table, and prompt for confirmation before posting the alert brief to Slack.

Governance, Departments & Scoping

Every skill has a clearly defined visibility and access scope:

Repeatability & Versioning

To ensure business compliance, skill modifications are strictly version-controlled:

• Semantic Versioning: Every update generates an immutable version record (e.g. v1.0.0 → v1.1.0).
• Deterministic Validation: Before a skill version is published, the package validator checks all tool bindings, syntax rules, and parameter dependencies.
• Instant Rollback: Administrators can instantly roll back to any previous version from the version history tab.

Sharing & Marketplace

Share your skills effortlessly across teams:

• Copy to Coworker: Clone any skill directly to a teammate's personal workspace.
• Export / Import Package: Export skills as portable .skill.tar.gz or JSON archives for staging-to-production promotions.
• Test Runs: Run isolated sandbox test simulations with custom test payloads before publishing changes to your department.