Intuned Automation Best Practices

Intuned Automation Best Practices

This skill provides comprehensive guidance for writing reliable, efficient, and maintainable browser automation scripts using the Intuned platform and SDK.

---

Performance optimization

Reference: docs/​03-how-to/​solve/​make-automations-faster.mdx

Waiting strategies

Never use arbitrary delays:

// Bad - wastes time or fails unpredictably
await page.waitForTimeout(5000);

// Good - waits for specific conditions
await page.locator("#data-table").waitFor({ state: "visible" });

Use DOM and network settling helpers:

// After actions that modify the DOM
await waitForDomSettled({
  source: page,
  settleDurationMs: 500,
  timeoutInMs: 30000
});

// After navigation or API calls
await withNetworkSettledWait(async (page) => {
  await page.click("button.load-data");
}, { page, timeoutInMs: 30000 });

Set shorter timeouts for fast pages:

page.setDefaultTimeout(10000); // 10 seconds instead of 30

Network interception for data extraction

Reference: docs/​01-learn/​recipes/​network-interception.mdx

Intercept API responses directly instead of parsing the DOM (often 10x faster):

let apiData: any;

page.on("response", async (response) => {
  if (response.url().includes("/​api/​products")) {
    apiData = await response.json();
  }
});

await withNetworkSettledWait(async (page) => {
  await page.click("button.load-products");
}, { page });

return apiData;

Efficient list extraction

Avoid iterating with locators - each locator auto-waits, adding overhead per item:

// Bad - multiple round trips
const items = await page.locator(".product-card").all();
for (const item of items) {
  const name = await item.locator(".name").textContent();
}

// Good - wait once, extract all in single evaluate
await page.locator(".product-card").first().waitFor({ state: "visible" });
const products = await page.evaluate(() => {
  return Array.from(document.querySelectorAll(".product-card")).map(card => ({
    name: card.querySelector(".name")?.textContent?.trim(),
    price: card.querySelector(".price")?.textContent?.trim()
  }));
});

Use fill() instead of pressSequentially()

// Slow - types character by character
await page.locator("input").pressSequentially("[email protected]");

// Fast - instant input
await page.locator("input").fill("[email protected]");

Build URLs directly

// Slow - clicking through filters
await page.click("[data-category='electronics']");

// Fast - navigate directly with query params
await page.goto("https://example.com/​products?category=electronics");

---

Cost optimization

Reference: docs/​03-how-to/​solve/​optimize-cost.mdx

Compute optimization

• Follow all performance tips (faster automations use less compute)
• Combine related operations (pagination in single run vs. multiple runs)
• Speed up AuthSession checks
• Reduce retries by fixing root causes

AI cost optimization

• Use cheaper models when appropriate
• Leverage Intuned SDK caching
• Replace AI with deterministic code when possible

Defer heavy processing

Move expensive operations outside the browser context:

// Bad - heavy processing in browser
const result = await page.evaluate(() => {
  return processData(document.body.innerHTML);
});

// Good - extract raw data, process afterward
const rawData = await page.evaluate(() => document.body.innerHTML);
const result = processData(rawData);

---

Anti-detection strategies

Reference: docs/​02-features/​stealth-mode-captcha-solving-proxies.mdx

Incremental feature enablement

Start simple and add features only as needed:

Proxy strategy

• Test without proxies first - many sites work fine without them
• Start with cheaper options: Datacenter → ISP → Residential (only for heavily protected sites)
• Rotate proxies to avoid rate limiting
• Use consistent IPs for authenticated sessions - configure at AuthSession level
• Monitor residential proxy usage - they bill per GB

Headful mode

Enable for sites with fingerprint detection:

{
  "browserMode": "headful"
}

Required for CAPTCHA solving and better browser fingerprints.

Stealth mode

Patches automation framework leaks:

{
  "stealthMode": true
}

Notes: Requires Playwright v1.55+, works best with headful mode, platform only.

CAPTCHA solving

Supported types: Google reCAPTCHA v2, hCaptcha, AWS CAPTCHA, Cloudflare managed challenge, GeeTest, Lemin, custom image/​text CAPTCHAs.

---

Pagination and infinite content

Pagination

Reference: docs/​01-learn/​recipes/​pagination.mdx

async function scrapeAllPages(page: Page) {
  const allProducts: Product[] = [];
  let currentPage = 1;
  const maxPages = 100;

  while (currentPage <= maxPages) {
    const products = await extractProducts(page);
    allProducts.push(...products);

    const nextButton = page.locator("button.next-page");
    if (!(await nextButton.isVisible())) break;

    await nextButton.click();
    await page.waitForLoadState("networkidle");
    currentPage++;
  }

  return allProducts;
}

Load more buttons

Reference: docs/​01-learn/​recipes/​load-more-button.mdx

async function loadAllContent(page: Page) {
  const maxClicks = 50;
  let clicks = 0;
  const loadMoreButton = page.locator("button:has-text('Load More')");

  while (clicks < maxClicks) {
    if (!(await loadMoreButton.isVisible())) break;
    await loadMoreButton.click();
    await waitForDomSettled({ source: page });
    clicks++;
  }

  return await extractAllProducts(page);
}

Infinite scrolling

Reference: docs/​01-learn/​recipes/​infinite-scrolling.mdx

await scrollToLoadContent({
  source: page,
  maxScrolls: 50,
  delayInMs: 100,
  minHeightChange: 100
});

const products = await extractProducts(page);

---

Data extraction patterns

AI-based extraction

Reference: docs/​01-learn/​recipes/​ai-scraper.mdx

Use when page structure is unpredictable:

import { z } from "zod";

const ProductSchema = z.object({
  name: z.string(),
  price: z.number(),
  rating: z.number().optional(),
  inStock: z.boolean()
});

const products = await extractStructuredData({
  source: page,
  dataSchema: ProductSchema,
  prompt: "Extract all product information from this page"
});

from pydantic import BaseModel

class Product(BaseModel):
    name: str
    price: float
    rating: float | None = None
    in_stock: bool

products = await extract_structured_data(
    page=page,
    data_schema=Product,
    prompt="Extract all product information from this page"
)

When to use: Unpredictable page structures, content that varies, fallback when selectors fail.

When NOT to use: Predictable structures (use deterministic code), high-volume scraping (adds cost).

---

Authentication and sessions

AuthSessions

Reference: docs/​02-features/​auth-sessions.mdx

Implement lightweight check API:

// check.ts - runs before every Run
export default async function check(page: Page) {
  // Check for authenticated content, not absence of login
  const userMenu = page.locator("[data-testid='user-menu']");
  return await userMenu.isVisible({ timeout: 5000 });
}

Choose the right type:

• credentials: Standard login flows
• runtime: Changing credentials (OTPs, temporary tokens)
• recorder: Complex MFA, biometrics, human interaction

Session management:

• Never share AuthSessions across users
• Keep check API fast (runs before every Run)

2FA/​OTP handling

Reference: docs/​01-learn/​recipes/​two-FA.mdx

import TOTP from "totp-generator";

async function loginWith2FA(page: Page, username: string, password: string, totpSecret: string) {
  await page.fill("[name='username']", username);
  await page.fill("[name='password']", password);
  await page.click("button[type='submit']");

  await page.locator("[name='otp']").waitFor({ state: "visible" });

  const otp = TOTP.generate(totpSecret, { digits: 6 }).otp;
  await page.fill("[name='otp']", otp);
  await page.click("button[type='submit']");
}

---

Project structure

Reference: docs/​03-how-to/​solve/​structure-intuned-projects.mdx

API design principles

• One API per logical unit of work - each can succeed/​fail independently
• Make APIs retriable - same parameters should work on retry

Common patterns:

When to split projects

Split when: Different sites, different auth, changing independently.

Combine when: Same platform, shared auth, shared code.

---

File operations

Download files

Reference: docs/​01-learn/​recipes/​download-file.mdx

const download = await downloadFile({
  page,
  trigger: page.locator("a.download"),
  timeoutInMs: 5000
});

const filename = download.suggestedFilename();

Upload to S3

Reference: docs/​01-learn/​recipes/​upload-files.mdx

const uploaded = await uploadFileToS3({
  file: download,
  s3Key: "downloads/​file.pdf"
});

const signedUrl = await uploaded.getSignedUrl();

---

Long-running operations

Timeout extension

Reference: docs/​01-learn/​recipes/​long-running-api.mdx, docs/​05-references/​runtime-sdk-typescript/​extend-timeout.mdx

For operations exceeding the default 10-minute timeout (supports up to 6 hours):

while (hasMorePages) {
  await processPage(page);
  extendTimeout(); // Reset timer after each unit of work
}

Payload extension

Reference: docs/​01-learn/​recipes/​extend-payload.mdx, docs/​05-references/​runtime-sdk-typescript/​extend-payload.mdx

Chain follow-up API calls dynamically:

const products = await extractProducts(page);

for (const product of products) {
  extendPayload({
    api: "details",
    params: { productId: product.id }
  });
}

---

Error handling and debugging

Reference: docs/​02-features/​observability-monitoring-logs.mdx, docs/​05-references/​error-codes.mdx

Debugging with trace viewer

• Replay browser session step-by-step
• Inspect network requests
• Check DOM snapshots
• Analyze failed selectors

Retry strategy

• Use AuthSession auto-recreation for auth failures
• Fix root causes instead of relying on retries
• Reduce retries for failing runs (each retry consumes compute)

---

Intuned SDK reference

Navigation

SDK Source: intuned-sdk/​typescript-sdk/​src/​helpers/​gotoUrl.ts

await goToUrl({
  page,
  url: "https://example.com",
  waitForLoadState: "load",
  retries: 3,
  timeoutInMs: 30000
});

DOM monitoring

SDK Source: intuned-sdk/​typescript-sdk/​src/​helpers/​waitForDomSettled.ts, intuned-sdk/​typescript-sdk/​src/​helpers/​withNetworkSettledWait.ts

await waitForDomSettled({
  source: page,
  settleDurationMs: 500,
  timeoutInMs: 30000
});

await withNetworkSettledWait(async (page) => {
  await page.click("button");
}, { page, timeoutInMs: 30000 });

Scrolling and loading

SDK Source: intuned-sdk/​typescript-sdk/​src/​helpers/​scrollToLoadContent.ts, intuned-sdk/​typescript-sdk/​src/​helpers/​clickUntilExhausted.ts

await scrollToLoadContent({
  source: page,
  maxScrolls: 50,
  minHeightChange: 100
});

await clickUntilExhausted({
  page,
  buttonLocator: page.locator("button.load-more"),
  containerLocator: page.locator(".results"),
  maxClicks: 50
});

AI extraction

SDK Source: intuned-sdk/​typescript-sdk/​src/​ai/​extractStructuredData.ts

const data = await extractStructuredData({
  source: page,
  dataSchema: MySchema,
  strategy: "HTML", // or "IMAGE", "MARKDOWN"
  enableCache: true,
  maxRetries: 3
});

File operations

SDK Source: intuned-sdk/​typescript-sdk/​src/​helpers/​downloadFile.ts, intuned-sdk/​typescript-sdk/​src/​helpers/​uploadFileToS3.ts

const download = await downloadFile({
  page,
  trigger: page.locator("a.download"),
  timeoutInMs: 5000
});

const uploaded = await uploadFileToS3({
  file: download,
  s3Key: "downloads/​file.pdf"
});

---

Claude recommendations

The following are general best practices not explicitly documented in Intuned docs but recommended for reliable automation:

Core principles

• Reliability over speed. A slow automation that always works is better than a fast one that fails intermittently.
• Explicit over implicit. Wait for specific conditions, not arbitrary timeouts.
• Simple over clever. Start with the simplest approach. Add complexity only when needed.

Selector best practices

Use stable, specific selectors. Avoid brittle selectors that depend on element order or styling.

Prefer semantic selectors (Playwright's role-based queries are more resilient to DOM changes):

// Preferred - semantic selectors
page.getByRole("button", { name: "Submit" })
page.getByText("Add to cart")
page.getByLabel("Email address")
page.getByPlaceholder("Enter your name")
page.getByTestId("product-card")

// Avoid - fragile selectors
page.locator("div:nth-child(3) > button")
page.locator(".btn-primary")

Selector priority order:

1. data-testid or similar test attributes
2. getByRole() with accessible name
3. getByLabel() for form elements
4. getByText() for unique text content
5. getByPlaceholder() for inputs
6. CSS selectors with IDs
7. CSS selectors with stable classes (last resort)

Batch evaluate calls

Each evaluate() is a round trip to the browser:

// Bad - multiple round trips
const title = await page.evaluate(() => document.title);
const url = await page.evaluate(() => window.location.href);

// Good - single batched call
const data = await page.evaluate(() => ({
  title: document.title,
  url: window.location.href
}));

Navigate to iframe URLs directly

// Slow - using frameLocator
const frame = page.frameLocator("iframe#content");
await frame.locator("button").click();

// Fast - navigate directly to iframe source
const iframeSrc = await page.locator("iframe#content").getAttribute("src");
await page.goto(iframeSrc);

Respectful automation

• Check site's robots.txt and terms of service
• Add delays between requests for high-volume scraping
• Don't overwhelm target servers

Common error patterns

Common pitfalls to avoid

Code review checklist

Before deploying automation code, verify:

• Selectors use stable attributes (data-testid, roles, labels)
• No arbitrary waitForTimeout() calls
• Network/​DOM settling helpers used appropriately
• Lists extracted in single evaluate() call
• Error handling for expected failures
• Timeouts set appropriately for page complexity
• AuthSession check API is lightweight
• No hard-coded credentials or secrets
• Proxy strategy matches site requirements
• Anti-detection features enabled only if needed
• Resources cleaned up (pages closed, etc.)
• Pagination has max limit to prevent infinite loops

---

Output format

When reviewing automation code, use this structure:

## Automation Code Review

**Files Reviewed:** [list]
**Overall Assessment:** [brief summary]

### Issues by Priority

#### High Priority
[Reliability issues - incorrect waits, fragile selectors]

#### Medium Priority
[Performance issues - inefficient extraction, missing batching]

#### Low Priority
[Style and maintainability]

### Detailed Feedback

#### [Filename:line]
**Issue:** [description]
**Current:** `[code]`
**Suggested:** `[code]`
**Reason:** [explanation]

### What Works Well
[Positive highlights]