Prompt

How to use

1. Copy everything below the line (or share this raw file’s URL).
2. Paste it into your AI coding assistant.
3. Add your browser-use script(s).
4. Review the generated Stagehand code and migration summary against your real site — tighten the act(...) prompts to the actual on-page labels, and confirm any flagged items.

Prefer it as a one-command tool inside Claude Code? The same logic ships as the /browser-use-to-stagehand skill. This prompt is the universal, tool-agnostic form.

You are migrating a browser-use (Python) browser-automation script to Stagehand v3 (TypeScript) running on Browserbase. Produce idiomatic, runnable Stagehand v3 code plus a migration summary. This is a refactor with judgment, not a line-by-line transpile.

Core principle

browser-use is agentic-by-default: an LLM decides every action on every run. Stagehand lets you choose how much AI to use at each step. A good migration replaces opaque agent loops with an inspectable, mostly-deterministic pipeline — using AI only where the page is genuinely unpredictable. The payoff is determinism, lower cost, and debuggability.

Stagehand v3 — get these exactly right

Most training data and blog posts show Stagehand v2. Use v3:

• Construct & lifecycle:

  typescriptCopy

  import "dotenv/config";
  import { Stagehand } from "@browserbasehq/stagehand";
  import { z } from "zod";
  
  const stagehand = new Stagehand({ env: "BROWSERBASE", model: "anthropic/claude-sonnet-4-6" });
  await stagehand.init();
  try {
    // ... work ...
  } finally {
    await stagehand.close();
  }

• Page object: const page = stagehand.context.pages()[0]; — not stagehand.page.
• AI methods are on the instance: stagehand.act(...), stagehand.extract(...), stagehand.observe(...) — not page.act(...).
• Models are "provider/model" strings: e.g. "anthropic/claude-sonnet-4-6", "openai/gpt-5", "google/gemini-2.5-flash". (Pin to whatever the team uses; ids move fast.)
• extract uses a zod schema; v3 supports a top-level z.array(...) with no wrapper object.
• Secrets via variables — the %token% is sent to the LLM, the real value is substituted locally and never leaves the machine:

  typescriptCopy

  await stagehand.act("type %username% into the email field", {
    variables: { username: process.env.APP_USER! },
  });

• Default to env: "BROWSERBASE"; show env: "LOCAL" (with localBrowserLaunchOptions) only as the dev option.
• Determinism/caching options: selfHeal: true, cacheDir, serverCache (on by default under BROWSERBASE).

The determinism spectrum — choose per step

Decision rule: split each browser-use task="…" string into its implied ordered steps, then place each on the spectrum. Default to decomposition (levels 2–5) when the flow is known; keep agent() (level 1) only for genuinely open-ended tasks (tighten it with maxSteps, a systemPrompt, and output for typed results). Reading data is always extract, never a full agent.

Example: task="Go to the store, search 'wireless mouse', add the cheapest to cart, checkout with my saved card" → page.goto(store) (L5) → act("search 'wireless mouse'") (L2) → extract the prices + pick the min in code + act("add to cart") (L2) → checkout via a Browserbase Context so auth is already present (L3/4).

Feature mapping

⚠️ Experimental gate: agent output, custom tools, and MCP integrations each require experimental: true on the Stagehand constructor (it bypasses the managed API path). For a typed result from an agentic run, prefer running the agent then a separate stagehand.extract(...).

No clean equivalent — flag these in the summary

• allowed_domains — Stagehand has no domain firewall, and it’s often a security boundary (it pairs with sensitive_data). Mitigate with a page.url() host check before sensitive actions, a systemPrompt constraint (for agents), or Browserbase proxy domain rules. Never drop it silently — flag it as needs-review.
• max_actions_per_step, use_thinking, flash_mode — no direct equivalent; decomposition makes steps explicit. For speed, use a fast model (google/gemini-2.5-flash) + decomposition.
• initial_actions — becomes ordinary code that runs before the first AI call.

browser-use variants to recognize (translate the same way)

• Legacy (pre-0.12): Browser(config=BrowserConfig(...)), BrowserContext, Controller / @controller.action. Normalize names first.
• Stable (0.12.x): Browser(browser_profile=BrowserProfile(...)), Tools() / @tools.action; Browser ≡ BrowserSession. (Most scripts.)
• Rust beta (0.13.x): imports from browser_use.beta. Same public surface.

Output

1. The Stagehand v3 TypeScript — runnable, with a package.json (deps: @browserbasehq/stagehand, zod, dotenv; add ai only if a custom action maps to an agent tool) and the required .env keys (BROWSERBASE_API_KEY, BROWSERBASE_PROJECT_ID, the provider key matching the model, plus any app secrets).
2. A migration summary:
   • Variant detected and the import/class tells.
   • Determinism choices per step (a short table: each browser-use step → Stagehand surface → level → one-line reasoning).
   • Needs human review — lost allowed_domains guardrails, custom-action logic, placeholder URLs/labels, anything ambiguous.
   • Recommended next step — usually a Browserbase Context for auth reuse, then selfHeal + caching for production.

Self-check before finishing

• AI methods on the instance (stagehand.act/extract/observe), page via stagehand.context.pages()[0].
• model is a "provider/model" string; the matching provider key is in .env.
• extract uses a zod schema; secrets use variables + process.env; nothing hardcoded.
• init() / close() present (close() in a finally).
• Every browser-use step is placed deliberately on the determinism spectrum; allowed_domains is not silently dropped.

If no script was provided, ask for the browser-use script before proceeding. Now migrate the browser-use script provided by the user.