Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/lindy-core-workflow-a/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill lindy-core-workflow-a. Copyright stays with the author (MIT).

Lindy Core Workflow A: Agent Creation

Overview

Complete workflow for creating, configuring, and testing Lindy AI agents. Agents consist of four components: Prompt (behavioral instructions), Model (AI engine), Skills (available actions), and Exit Conditions (completion criteria).

Prerequisites

• Lindy account at
• Use case defined (support bot, email triage, data processor, etc.)
• Required integrations identified (Slack, Gmail, Sheets, etc.)

Instructions

Step 1: Define Agent Specification

Before building, document:

• Purpose: What the agent does (one sentence)
• Trigger: What wakes it up (webhook, email, schedule, chat, etc.)
• Actions: What it does (send email, update sheet, post to Slack, etc.)
• Model: GPT-4 (smart, expensive), Claude Sonnet (balanced), Gemini Flash (fast, cheap)

Step 2: Create the Agent

Option A — Natural Language (recommended):

1. Click New Agent at

2. Describe your agent in plain English:

   When a customer emails support@company.com, classify the email as
   billing/technical/general, draft a response using our knowledge base,
   and post the classification to #support-triage in Slack
   

3. Agent Builder auto-generates trigger + action nodes

Option B — Manual Build:

1. Click New Agent > Start from scratch
2. Add trigger: Click "+" > Select trigger type
3. Add actions: Click "+" > Search for action
4. Connect nodes by dragging arrows between steps

Step 3: Configure the Prompt

Open Settings > Prompt. Structure it with clear sections:

## Identity
You are a customer support classifier and responder for [Company].

## Instructions
1. Read the incoming email carefully
2. Classify into: billing, technical, or general
3. Search the knowledge base for relevant answers
4. Draft a professional response using the KB results
5. If no KB match found, escalate to human

## Constraints
- Never promise refunds or credits without human approval
- Keep responses under 200 words
- Always include ticket reference number

Prompt best practices (from Lindy docs):

• Use action-oriented language (imperatives)
• Break complex logic into numbered steps
• Include few-shot examples for consistent formatting
• Add constraints to prevent unwanted behavior

Step 4: Add Conditions (Branching Logic)

1. Click "+" > Condition
2. Write natural language condition: "Go down this path if the email is about billing"
3. Add multiple branches for different classifications
4. Enable "Force the agent to select a branch" for deterministic routing

Step 5: Configure Actions

For each action, set field modes:

Auto mode — Agent infers the value from all previous step data:

Best for: predictable mappings where field names align

AI Prompt mode — Give natural language instructions:

Summarize the email in 2 sentences, then include the classification.
Reference: {{email_received.body}}

Set Manually mode — Exact value, no AI:

Channel: #support-triage

Step 6: Add Knowledge Base (Optional)

1. Go to Settings > Knowledge Base
2. Add sources: PDF, DOCX, Google Drive, Notion, websites
3. Configure search:
   • Max Results: 4 (default) to 10
   • Search Fuzziness: 0 (keyword) to 100 (semantic, recommended)
4. Agent auto-searches KB when relevant to the task

Step 7: Test the Agent

1. Use the Test button in the workflow editor
2. Provide sample input matching your trigger type
3. Review each step's output in the task detail view
4. Iterate on prompt and action configuration

Trigger Types Reference

Action Categories

Error Handling

Output

Produce an agent specification with a named trigger, explicit identity and constraints, ordered actions, safe failure behavior, and one acceptance fixture that proves the agent performs the intended task without unintended side work.

Examples

Create a lead-routing agent triggered by a tagged form submission. It validates the required fields, assigns the record only when the stated routing rule matches, and otherwise records a review task; a fixture for each branch proves the constraints are applied before any connected-system action.

Resources

• Lindy Introduction
• Triggers Documentation
• Actions Documentation
• Prompt Guide

Next Steps

Proceed to lindy-core-workflow-b for triggers, automation, and multi-agent delegation.