Prompts

Reusable, structured instructions referenced by agents and LLM/extraction tasks.

A prompt is a saved, reusable bundle of instructions — system + user roles, an output shape, and metadata — that lives in the Prompt Library (/api/prompts). Instead of hand-writing instructions on every node, you author a prompt once and reference it by prompt_title from an agent or by prompt_id from an extraction or LLM task.

Prompts are system-level (shared, Super-Admin owned) or user-level (private to their owner). Titles are globally unique, so prompt_title is a stable reference.

What makes a prompt work

The durable tactics are model-agnostic and worth keeping in mind whatever you store in the library:

Be explicit. State the task, the output format, and any hard constraints — treat the model like a capable new hire with no prior context.
Show examples. A few diverse input/output pairs steer format and tone better than more description (few-shot).
Give structure. Separate role, instructions, input, and output so the model parses them unambiguously — exactly what RTIO (below) does.
Add context, not assumptions. Supply the domain facts the model needs; for large bodies of knowledge, retrieve them with RAG rather than pasting them in.

Synthesized from the prompting guides of Anthropic, OpenAI, and Google.

Go deeper. For exhaustive prompt-engineering technique, see Anthropic, OpenAI, and Google. This page covers how TurfAI stores and references prompts.

The prompt object

These fields map to the prompt content type. The outputFormat + exampleOutput pair is what makes a prompt produce consistent structured results — define the JSON shape once and every run targets it.

Field	Type	Default	Description
`title`	string	—	Display name. Required and globally unique.
`description`	text	`null`	Human-readable summary of what the prompt does.
`section`	string	`null`	Category used for grouping / filtering in the library.
`roles`	json	system/user/developer/assistant stubs	The chat messages: `[{ role, prompt }]` for `system`, `user`, `developer`, `assistant`.
`tasks`	json	`[]`	Ordered list of extraction/processing steps the model should perform.
`instructions`	json	`[]`	Output-formatting and edge-case rules (e.g. "use `null` for missing fields").
`outputFormat`	enum	`text`	`text`, `json`, `markdown`, or `structured`.
`exampleOutput`	json	`{}`	A worked example of the expected output — the output schema the model targets.
`version`	string	`1.0`	Author-managed version label.
`level`	enum	`user`	`user` (private to owner) or `system` (shared; Super Admin only).
`owner`	relation	current user	Set automatically to the caller for user-level prompts.

The synced Prompt API reference describes the request/response bodies in a simplified shape. The field list above is the authoritative content-type schema — prefer it when building payloads, and treat the reference as the endpoint contract.

Example `exampleOutput` — invoice extraction

A structured prompt that pins the output to an invoice schema:

{
  "invoice_number": "INV-2024-0148",
  "invoice_date": "2024-03-12",
  "vendor_name": "Acme Supplies Ltd",
  "total_amount": 4820.50,
  "currency": "USD",
  "line_items": [
    { "description": "Widget A", "quantity": 10, "unit_price": 42.0, "amount": 420.0 }
  ]
}

The model returns data in this shape; missing fields come back as null, empty lists as [].

CRUD via the API

All calls hit base URL https://apisandbox.turfai.in/api with a JWT bearer token. Create and list are shown below in three languages — pick whichever matches your stack; they're equivalent.

Create a prompt — `POST /prompts`

curl -X POST https://apisandbox.turfai.in/api/prompts \
  -H "Authorization: Bearer $TURFAI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "data": {
      "title": "Invoice Extraction",
      "description": "Extract header + line items from invoices",
      "section": "finance",
      "outputFormat": "structured",
      "roles": [
        { "role": "system", "prompt": "You are an expert at reading invoices." },
        { "role": "user", "prompt": "Extract the fields below from {{document_content}}." }
      ],
      "tasks": ["Read the invoice", "Extract header fields", "Extract line items"],
      "instructions": ["Return only valid JSON", "Use null for missing fields"],
      "exampleOutput": { "invoice_number": "INV-2024-0148", "total_amount": 4820.50 }
    }
  }'

import os, requests

BASE = "https://apisandbox.turfai.in/api"
headers = {"Authorization": f"Bearer {os.environ['TURFAI_TOKEN']}"}

payload = {
    "data": {
        "title": "Invoice Extraction",
        "description": "Extract header + line items from invoices",
        "section": "finance",
        "outputFormat": "structured",
        "roles": [
            {"role": "system", "prompt": "You are an expert at reading invoices."},
            {"role": "user", "prompt": "Extract the fields below from {{document_content}}."},
        ],
        "tasks": ["Read the invoice", "Extract header fields", "Extract line items"],
        "instructions": ["Return only valid JSON", "Use null for missing fields"],
        "exampleOutput": {"invoice_number": "INV-2024-0148", "total_amount": 4820.50},
    }
}

resp = requests.post(f"{BASE}/prompts", json=payload, headers=headers)
resp.raise_for_status()
print(resp.json())

const BASE = "https://apisandbox.turfai.in/api";

const resp = await fetch(`${BASE}/prompts`, {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.TURFAI_TOKEN}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    data: {
      title: "Invoice Extraction",
      description: "Extract header + line items from invoices",
      section: "finance",
      outputFormat: "structured",
      roles: [
        { role: "system", prompt: "You are an expert at reading invoices." },
        { role: "user", prompt: "Extract the fields below from {{document_content}}." },
      ],
      tasks: ["Read the invoice", "Extract header fields", "Extract line items"],
      instructions: ["Return only valid JSON", "Use null for missing fields"],
      exampleOutput: { invoice_number: "INV-2024-0148", total_amount: 4820.5 },
    },
  }),
});
if (!resp.ok) throw new Error(`Create failed: ${resp.status}`);
console.log(await resp.json());

List prompts — `GET /prompts`

Returns system-level prompts plus the caller's own. Supports search, tag, and pagination[page] / pagination[pageSize] query params.

curl "https://apisandbox.turfai.in/api/prompts?pagination[pageSize]=25" \
  -H "Authorization: Bearer $TURFAI_TOKEN"

import os, requests

BASE = "https://apisandbox.turfai.in/api"
headers = {"Authorization": f"Bearer {os.environ['TURFAI_TOKEN']}"}

resp = requests.get(f"{BASE}/prompts", params={"pagination[pageSize]": 25}, headers=headers)
resp.raise_for_status()
print([p["title"] for p in resp.json()["data"]])

const BASE = "https://apisandbox.turfai.in/api";

const resp = await fetch(`${BASE}/prompts?pagination[pageSize]=25`, {
  headers: { Authorization: `Bearer ${process.env.TURFAI_TOKEN}` },
});
if (!resp.ok) throw new Error(`List failed: ${resp.status}`);
const { data } = await resp.json();
console.log(data.map((p: { title: string }) => p.title));

PUT /prompts/:id and DELETE /prompts/:id round out CRUD — both are owner-checked (or Super Admin for system-level prompts). See the Prompt API reference for full endpoint contracts.

RTIO structure

TurfAI structures extraction prompts as RTIO — Role, Task, Input, Output:

Role — who the model is (the system role): "You are an expert at reading invoices."
Task — what to extract (tasks + the user role).
Input — the document content, injected via the {{document_content}} placeholder.
Output — the target shape (outputFormat + exampleOutput + formatting instructions).

When you run a structured extraction, the platform assembles these into a single template (Role / Task / Output Format / Rules) before sending it to the model.

AI enhancement

The Prompt Lab can take a rough prompt and an output schema and return a polished RTIO structure (roles / tasks / instructions) via POST /prompt-lab/enhance-rtio. There's also POST /prompts/generate-rtio, which infers a full extraction prompt from a sample document and its classification.

Coming soon. The internal scoring/ranking algorithm behind AI enhancement and document-driven RTIO generation isn't documented yet — these run as asynchronous LLM jobs in the processor. This page will cover the request/response contracts for enhance-rtio and generate-rtio once they're stable in the public reference.

How prompts plug in

From an agent — set prompt_title to the prompt's (unique) title when creating the agent:

{
  "data": {
    "name": "Invoice Bot",
    "goal": "Extract and validate invoice data",
    "prompt_title": "Invoice Extraction"
  }
}

From an extraction task — reference the saved prompt by prompt_id in the node config:

{
  "task_type": "extraction",
  "config": {
    "document_id": 1234,
    "prompt_id": 42
  }
}

An llm task works the same way — supply a saved prompt_id or an inline prompt_text. See Task types for the full node config.

Reference

Prompt API — endpoints and request/response bodies.
Agents — reference a prompt from an agent via prompt_title.
Task types — extraction and llm nodes that consume prompt_id.

On this page