AI Nodes

TinyWorkflows has 5 built-in AI nodes powered by TinyGPT. Use them to classify data, generate text, research topics, search the web, and extract structured insights, all within the workflow canvas.

All 5 AI nodes

Node	Type	Credits	What it does	Best for
TinyGPT	`GPT`	10	Classify, extract, analyze, or generate structured data	Sentiment analysis, data extraction, ticket classification
GPT Writer	`GPT_WRITER`	15	Generate text content with configurable tone and length	Email drafts, marketing copy, summaries
GPT Researcher	`GPT_RESEARCHER`	20	Conduct structured research from multiple sources	Market research, competitor analysis, topic deep-dives
Tiny Search	`TINY_SEARCH`	10	Search the web and return structured results	Current events, lead discovery, content research
Agent	–	Varies	Run a TinyAgent sub-agent within a workflow	Complex multi-step AI tasks

How AI nodes work

Every AI node follows the same 3-tab pattern: Initialise → Configure → Test

Initialise tab

Choose a template that pre-fills the prompt and configuration:

TinyGPT templates: Answer Questions, Generate Content, Extract Data, Classify & Categorize Writer templates: Professional email, Blog post, Social media post Researcher templates: Market Research, Competitor Analysis Search templates: Market Research, Competitor Analysis, Lead Generation, Content Research

Configure tab

Set the actual AI parameters:

Field	Required	Description
System Prompt	Yes	Define the AI's role and behavior: "You are a customer support classifier that..."
Query	Yes	The actual task: what to do with the input data. Supports FX variables.
Output Schema	Yes	Define the JSON structure the AI should return: field names, types, and examples
Temperature	No	Creativity level: 0.0 (deterministic) to 1.0 (creative). Default: 0.7
Max tokens	No	Maximum response length. Default: 1024

Test tab

Run the AI with sample data and see the actual output. The test populates the output schema so downstream nodes can reference specific fields.

Output variables

After the AI processes, its output is available as variables:

{{tinygpt.result}}              → the full JSON result
{{tinygpt.result.category}}     → a specific field
{{tinygpt.result.confidence}}   → another field
{{tinygpt.result.key_phrases}}  → an array field

The exact field names match your Output Schema definition.

AI + Logic patterns

The real power comes from combining AI with logic nodes:

Classify and route

Webhook (support ticket) → TinyGPT (classify category + priority) → If-Else →
  billing: → Create JIRA ticket (billing queue)
  technical: → Send Slack message (#engineering)
  feature-request: → Add to spreadsheet (feature backlog)

Batch personalization

Schedule (daily) → Find All (pending outreach) → For Each →
  Person Enrichment → GPT Writer (personalized email) → Send Email

Research → Write → Review

Manual Trigger → GPT Researcher (topic research) → GPT Writer (draft article) →
  HITL (human review) → If approved → Publish

Enrich → Classify → Store

Form Submission → Company Enrichment → TinyGPT (classify lead tier) →
  Create Record (save to CRM with tier)

Prompt engineering tips

Principle	Bad prompt	Good prompt
Be specific	"Analyze this text"	"Extract the customer's name, email, and main complaint. Return as JSON."
Define output	"Classify this"	"Classify as: billing, technical, feature-request, or other. Include confidence 0-1."
Give examples	"Summarize"	"Summarize in 2-3 bullet points, each under 20 words. Example: '• Payment failed for customer X'"
Set constraints	"Write an email"	"Write a 3-sentence follow-up email in professional tone. Do not mention pricing."

Warning

AI output is probabilistic: the same input may produce slightly different results on different runs. For critical decisions (refunds, approvals, compliance), always add a Human-in-the-Loop node after the AI step.

Note

All AI nodes consume credits. Monitor usage in Monitor → Credit Usage. TinyGPT costs 10 credits, Writer costs 15, Researcher costs 20, and Search costs 10 per execution.