Recipes — copy-paste plans for common patterns

Working starting points for the most common extraction and form-driven patterns. Each recipe is a complete /v1/predict request body — swap the URL and the schema fields and you have a working pipeline.

Two structural shapes:

• Listings flows (recipes 1–4): navigate → gate → click → extract_url → scroll → extract_data → re-navigate → loop.
• Form flows (recipe 5): navigate → fill_field → fill_field → submit → select_option → submit. No looping; each step is a labelled action.

For the full request envelope (auth, polling, results), see Generic CUA usage. For when each piece kicks in, see Concepts.

---

Recipe 1 — Job listings (Greenhouse / Lever / Workday-style)

Best for: boards with /jobs/<id> detail URLs and filterable list pages.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "greenhouse-openai-eng-sf",
  "max_cost": 4,
  "max_time_minutes": 30,
  "extraction_schema": {
    "entity_name": "job posting",
    "fields": [
      {"name": "title",      "type": "str", "required": true,  "example": "Software Engineer"},
      {"name": "team",       "type": "str", "required": false, "example": "Infrastructure"},
      {"name": "location",   "type": "str", "required": false, "example": "San Francisco, CA"},
      {"name": "department", "type": "str", "required": false, "example": "Engineering"},
      {"name": "url",        "type": "str", "required": true,  "example": "boards.greenhouse.io/..."}
    ],
    "required_fields": ["title", "url"],
    "spam_indicators": ["recruiter", "staffing agency", "third party"],
    "spam_label": "recruiter spam"
  },
  "micro": [
    {"intent": "Navigate to https://boards.greenhouse.io/openai/jobs?department=Engineering&location=San+Francisco",
     "type": "navigate", "budget": 3, "section": "setup", "required": true},
    {"intent": "Verify page is the OpenAI Greenhouse engineering board, San Francisco filter active, listings visible",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows engineering job listings filtered to San Francisco"},
    {"intent": "Click the next un-extracted job posting title",
     "type": "click", "budget": 8, "grounding": true, "section": "extraction"},
    {"intent": "Read URL from address bar",
     "type": "extract_url", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Scroll down to read the description",
     "type": "scroll", "budget": 5, "section": "extraction"},
    {"intent": "Extract title, team, location, department, url",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Re-navigate to the search page at https://boards.greenhouse.io/openai/jobs?department=Engineering&location=San+Francisco",
     "type": "navigate", "budget": 3, "section": "extraction"},
    {"intent": "Loop to next listing",
     "type": "loop", "loop_target": 2, "loop_count": 10, "section": "extraction"}
  ]
}
JSON

Lever variant: same shape, replace start URL with https://jobs.lever.co/<company>/?team=Engineering&location=... and the gate verify text accordingly.

---

Recipe 2 — E-commerce product catalog (Amazon / Shopify-style)

Best for: product detail pages with price, rating, availability fields.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "amazon-mech-keyboards",
  "max_cost": 3,
  "max_time_minutes": 25,
  "extraction_schema": {
    "entity_name": "product",
    "fields": [
      {"name": "name",         "type": "str", "required": true,  "example": "Keychron K6 Wireless"},
      {"name": "price",        "type": "str", "required": true,  "example": "$89.99"},
      {"name": "rating",       "type": "str", "required": false, "example": "4.5/5 (1,234)"},
      {"name": "availability", "type": "str", "required": false, "example": "In Stock"},
      {"name": "brand",        "type": "str", "required": false, "example": "Keychron"},
      {"name": "url",          "type": "str", "required": true,  "example": "amazon.com/dp/..."}
    ],
    "required_fields": ["name", "price", "url"]
  },
  "site_config": {
    "domain": "amazon.com",
    "detail_page_pattern": "/dp/[A-Z0-9]+",
    "results_page_pattern": "/s\\?",
    "pagination_format": "page={n}",
    "pagination_type": "query_param",
    "pagination_strip_pattern": "[?&]page=\\d+"
  },
  "micro": [
    {"intent": "Navigate to https://www.amazon.com/s?k=mechanical+keyboard",
     "type": "navigate", "budget": 3, "section": "setup", "required": true},
    {"intent": "Verify page is Amazon search results for mechanical keyboards",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true, "verify": "Page is an Amazon search results page for mechanical keyboards with multiple products visible"},
    {"intent": "Click the next un-extracted product title (skip sponsored tiles)",
     "type": "click", "budget": 8, "grounding": true, "section": "extraction"},
    {"intent": "Read URL from address bar",
     "type": "extract_url", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Scroll down to see price + availability + reviews",
     "type": "scroll", "budget": 5, "section": "extraction"},
    {"intent": "Extract name, price, rating, availability, brand, url",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Re-navigate to https://www.amazon.com/s?k=mechanical+keyboard",
     "type": "navigate", "budget": 3, "section": "extraction"},
    {"intent": "Loop to next product",
     "type": "loop", "loop_target": 2, "loop_count": 8, "section": "extraction"},
    {"intent": "Click Next page",
     "type": "paginate", "budget": 10, "grounding": true, "section": "pagination"},
    {"intent": "Loop back to extract products on the new page",
     "type": "loop", "loop_target": 2, "loop_count": 3, "section": "pagination"}
  ]
}
JSON

Note the two-level loop: the inner loop walks listings on a page; the outer loop paginates and re-runs the inner loop.

---

Recipe 3 — News articles (any newsroom site)

Best for: structured headline + byline + body extraction.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "techcrunch-ai-recent",
  "max_cost": 2,
  "max_time_minutes": 20,
  "extraction_schema": {
    "entity_name": "news article",
    "fields": [
      {"name": "headline",  "type": "str", "required": true,  "example": "OpenAI launches..."},
      {"name": "byline",    "type": "str", "required": false, "example": "Jane Doe"},
      {"name": "published", "type": "str", "required": false, "example": "2026-04-29T12:00:00Z"},
      {"name": "summary",   "type": "str", "required": false, "example": "first paragraph"},
      {"name": "url",       "type": "str", "required": true,  "example": "techcrunch.com/2026/..."}
    ],
    "required_fields": ["headline", "url"]
  },
  "micro": [
    {"intent": "Navigate to https://techcrunch.com/category/artificial-intelligence/",
     "type": "navigate", "budget": 3, "section": "setup", "required": true},
    {"intent": "Verify page is the TechCrunch AI category index with headlines visible",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true, "verify": "Page is TechCrunch's AI category page with multiple article headlines"},
    {"intent": "Click the next un-extracted article headline",
     "type": "click", "budget": 8, "grounding": true, "section": "extraction"},
    {"intent": "Read URL from address bar",
     "type": "extract_url", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Scroll past the lede paragraph",
     "type": "scroll", "budget": 3, "section": "extraction"},
    {"intent": "Extract headline, byline, published, summary (first paragraph), url",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Re-navigate to https://techcrunch.com/category/artificial-intelligence/",
     "type": "navigate", "budget": 3, "section": "extraction"},
    {"intent": "Loop to next article",
     "type": "loop", "loop_target": 2, "loop_count": 5, "section": "extraction"}
  ]
}
JSON

For paywalled sources, supply login via the embedding path (register_tool) — the HTTP-only path doesn't have a credential-injection primitive.

---

Recipe 4 — Real estate listings (Zillow / Redfin / Trulia-style)

Best for: listings with address, price, beds/baths/sqft, listing agent.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "zillow-sf-condos-1m-2m",
  "max_cost": 5,
  "max_time_minutes": 35,
  "extraction_schema": {
    "entity_name": "real estate listing",
    "fields": [
      {"name": "address",      "type": "str", "required": true,  "example": "123 Main St, San Francisco, CA 94105"},
      {"name": "price",        "type": "str", "required": true,  "example": "$1,250,000"},
      {"name": "beds",         "type": "str", "required": false, "example": "2"},
      {"name": "baths",        "type": "str", "required": false, "example": "2"},
      {"name": "sqft",         "type": "str", "required": false, "example": "1,100"},
      {"name": "agent",        "type": "str", "required": false, "example": "Jane Smith, Compass"},
      {"name": "url",          "type": "str", "required": true,  "example": "zillow.com/homedetails/..."}
    ],
    "required_fields": ["address", "price", "url"],
    "spam_indicators": ["3D Home", "Zestimate-only", "off-market"],
    "spam_label": "non-listing"
  },
  "site_config": {
    "domain": "zillow.com",
    "detail_page_pattern": "/homedetails/[\\w-]+/\\d+_zpid",
    "results_page_pattern": "/homes/for_sale/",
    "pagination_format": "/{n}_p/",
    "pagination_type": "path_suffix"
  },
  "micro": [
    {"intent": "Navigate to https://www.zillow.com/homes/for_sale/San-Francisco-CA/condos_dom/?price=1000000-2000000",
     "type": "navigate", "budget": 3, "section": "setup", "required": true},
    {"intent": "Verify page is the SF condos for sale, $1M-$2M price range, with listings visible",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true, "verify": "Zillow SF condos $1M-$2M, multiple property cards visible"},
    {"intent": "Click the next un-extracted listing card",
     "type": "click", "budget": 8, "grounding": true, "section": "extraction"},
    {"intent": "Read URL from address bar",
     "type": "extract_url", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Scroll to property facts and listing agent",
     "type": "scroll", "budget": 6, "section": "extraction"},
    {"intent": "Extract address, price, beds, baths, sqft, agent, url",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Re-navigate to https://www.zillow.com/homes/for_sale/San-Francisco-CA/condos_dom/?price=1000000-2000000",
     "type": "navigate", "budget": 3, "section": "extraction"},
    {"intent": "Loop to next listing",
     "type": "loop", "loop_target": 2, "loop_count": 12, "section": "extraction"}
  ]
}
JSON

Real estate sites are aggressive about Cloudflare; if you see gate_failed in the result, increase proxy_city / proxy_state to a residential target close to the listing geography.

---

Recipe 5 — CRM-style form flow (login → edit → save)

Best for: admin consoles, CRM record updates, settings pages — anywhere the workflow is "fill these fields and click Save", not "scrape these listings."

This recipe uses the form-shaped step types (fill_field, submit, select_option) introduced in issue #80. They route through a single labelled-element grounder (no listings extraction), so login pages and edit forms work the same as search-result extraction works for listings.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "crm-edit-lead-industry",
  "max_cost": 2,
  "max_time_minutes": 15,
  "micro": [
    {"intent": "Go to https://crm.example.com",
     "type": "navigate", "budget": 3, "section": "setup", "required": true},
    {"intent": "Enter user ID",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "User ID", "value": "alice"}},
    {"intent": "Enter password",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Password", "value": "<password>"}},
    {"intent": "Click Login",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Login"}},
    {"intent": "Verify dashboard or home page is shown",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page is the post-login dashboard for alice"},
    {"intent": "Go to the Leads page",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"label": "Leads"}},
    {"intent": "Click the first lead row with Qualified status",
     "type": "click", "budget": 8, "grounding": true, "section": "setup",
     "required": true},
    {"intent": "Click Edit Lead",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Edit Lead"}},
    {"intent": "Set Industry Vertical to Space Exploration",
     "type": "select_option", "budget": 6, "section": "setup", "required": true,
     "params": {"dropdown_label": "Industry Vertical", "option_label": "Space Exploration"}},
    {"intent": "Click Update Lead to save changes",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Update Lead"}},
    {"intent": "Verify the Update succeeded (toast / redirect to lead view)",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows confirmation that the lead was updated; Industry Vertical now reads Space Exploration"}
  ]
}
JSON

Notes on the form-flow shape:

• No loop step — form flows are linear. Each step does one thing.
• required: true everywhere — failing a fill_field or submit is fatal. You can't extract a lead's data if login failed.
• params is the source of truth — the runner uses params["label"] to locate the element via Claude vision, not the prose intent. The prose is just for logs and traces.
• gate steps frame the workflow — one after login, one after save. Either gate failing halts the run before damage propagates.
• select_option is a two-phase step — the runner first clicks the dropdown to open it, screenshots the open menu, then clicks the option. Both phases use Claude grounding; if either fails the whole step fails.

The exact same pattern works for any admin console: replace the URL, the field labels, the dropdown labels, and the button labels.

plan_text variant for ad-hoc form flows

You can also let the decomposer figure out the form structure from free-text. It now recognizes verbs like enter, type, fill, set, choose, select, pick, click {Submit|Save|Update|Login} and emits the right step types automatically:

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "detached": true,
    "plan_text": "1. Go to https://crm.example.com\n2. Log in with user ID alice password <password>\n3. Go to the Leads Page\n4. Select the first lead with the \"Qualified\" status\n5. Go to the Edit Lead page for this lead\n6. Update the Industry Vertical to \"Space Exploration\"\n7. Click \"Update Lead\"",
    "state_key": "crm-edit-lead-text",
    "max_cost": 2,
    "max_time_minutes": 15
  }'

The decomposer will emit fill_field / submit / select_option steps with structured params. For high-volume / production use, prefer the hand-authored micro form (above) — it skips the decompose step's $0.02 Claude call and bakes in your exact field labels.

---

Action recipes — post, read, write, refund

The five recipes above cover read flows (extract listings) and the canonical write flow (CRM edit + save). The recipes below extend the form-flow shape to the most-asked action patterns: posting content on social platforms, fetching a single user's profile, pushing data into a CRM, and processing a refund through a payments dashboard.

All six work against any site that follows the same UI pattern — replace the URL, the field labels, and the button labels. Auth is plan-driven: the fill_field / submit chain logs in just like any human does.

Recipe 6 — Post on LinkedIn

Best for: automating routine company updates, employee announcements, or weekly newsletters that go through a real LinkedIn account.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "linkedin-post-weekly",
  "max_cost": 2,
  "max_time_minutes": 15,
  "micro": [
    {"intent": "Go to https://www.linkedin.com/login",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"wait_after_load_seconds": 25}},
    {"intent": "Enter the email address",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Email or phone", "value": "alice@example.com"}},
    {"intent": "Enter the password",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Password", "value": "<password>"}},
    {"intent": "Click Sign in",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Sign in", "aliases": ["Log in", "Continue"]}},
    {"intent": "Verify the LinkedIn home feed loaded",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows the LinkedIn home feed with the 'Start a post' composer visible"},
    {"intent": "Click 'Start a post'",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Start a post"}},
    {"intent": "Type the post body",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "What do you want to talk about?",
                "value": "Excited to share this week's product update — new structured-extraction features now in beta. Drop a comment if you want early access."}},
    {"intent": "Click Post to publish",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Post", "aliases": ["Publish", "Share"]}},
    {"intent": "Verify the post landed on the feed",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Confirmation toast or the post is now visible at the top of the feed with the body text we typed"}
  ]
}
JSON

Notes:

• The wait_after_load_seconds: 25 on the navigate covers LinkedIn's proxied login splash. Lower it once you confirm the page loads faster.
• aliases on the two submit steps absorbs LinkedIn's frequent button copy A/B tests (Sign in vs Log in, Post vs Publish vs Share).
• For posting an image too, add a submit step with params={"label": "Add a photo"} between the composer-open and the text-fill, then a second submit for the file picker. Mantis can drive the file-picker dialog if your env supplies a LAUNCH_APP / file-upload primitive (see issue #72).

Recipe 7 — Post on Reddit (subreddit text post)

Best for: community announcements, weekly digests, AMA invitations.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "reddit-post-r-mysub",
  "max_cost": 1.5,
  "max_time_minutes": 12,
  "micro": [
    {"intent": "Go to https://www.reddit.com/login",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"wait_after_load_seconds": 20}},
    {"intent": "Enter the username",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Username", "value": "alice"}},
    {"intent": "Enter the password",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Password", "value": "<password>"}},
    {"intent": "Click Log In",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Log In", "aliases": ["Sign In", "Continue"]}},
    {"intent": "Go to the target subreddit submit page",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"wait_after_load_seconds": 15}},
    {"intent": "Click the Post tab",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Post", "aliases": ["Text"]}},
    {"intent": "Type the post title",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Title",
                "value": "Weekly digest: what shipped this week"}},
    {"intent": "Type the post body",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Text (optional)",
                "value": "Three things shipped: 1) generic CUA over HTTP. 2) form-flow vocab. 3) any-agent integration playbook. Comments welcome."}},
    {"intent": "Click Post to submit",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Post", "aliases": ["Submit", "Publish"]}},
    {"intent": "Verify the post landed",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows the new post with our title, on the subreddit's frontpage or in our profile"}
  ]
}
JSON

Replace the navigate URL with https://www.reddit.com/r/<your-sub>/submit. Reddit's submit-page UI varies between old.reddit / new.reddit / sh — the label strings work across the new UI (the most common landing).

Recipe 8 — Post on Instagram (caption + tag)

Best for: scheduled feed posts. Instagram's feed-post composer is behind a login wall and behind a click-to-open camera dialog, so this recipe shows the multi-modal-button shape:

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "instagram-post-feed",
  "max_cost": 2.5,
  "max_time_minutes": 18,
  "micro": [
    {"intent": "Go to https://www.instagram.com/accounts/login/",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"wait_after_load_seconds": 25}},
    {"intent": "Enter the username",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Phone number, username, or email", "value": "alice"}},
    {"intent": "Enter the password",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Password", "value": "<password>"}},
    {"intent": "Click Log in",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Log in"}},
    {"intent": "Verify the Instagram home feed loaded",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page is the Instagram home feed; left sidebar shows Home / Search / Explore / Reels / Messages / Notifications / Create"},
    {"intent": "Click Create in the sidebar to open the post composer",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Create", "aliases": ["New post", "+"]}},
    {"intent": "Click Select from computer to attach an image",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Select from computer", "aliases": ["Choose file", "Upload"]}},
    {"intent": "Click Next to advance from filter selection",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Next"}},
    {"intent": "Click Next again to advance to caption",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Next"}},
    {"intent": "Type the caption",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Write a caption...",
                "value": "Sunset in Lisbon — week of off-sites done right. #travel #lisbon #productteam"}},
    {"intent": "Click Share to publish",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Share", "aliases": ["Post", "Publish"]}},
    {"intent": "Verify the post is published",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Toast or modal confirms the post was shared, OR the composer dialog closed and we are back on the home feed"}
  ]
}
JSON

Notes:

• Image upload requires either (a) a desktop file already on the runner's Xvfb container at a known path, OR (b) the LAUNCH_APP action plus a small file-picker driver in your GymEnvironment. Without that the Select from computer button opens the OS dialog and the runner can't navigate it via screen pixels alone.
• The two consecutive Next clicks are intentional — IG's composer has a Crop step then a Filter step before the caption.

Recipe 9 — Fetch user details from an admin console

Best for: support workflows where you need to read a single user's record (email, plan, last-login, billing status) from an internal admin UI. This is a mixed read/extract flow: log in, search by email, extract the profile fields, no looping.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "admin-fetch-user-by-email",
  "max_cost": 1.5,
  "max_time_minutes": 10,
  "extraction_schema": {
    "entity_name": "user profile",
    "fields": [
      {"name": "email",         "type": "str", "required": true,  "example": "alice@example.com"},
      {"name": "full_name",     "type": "str", "required": true,  "example": "Alice Anderson"},
      {"name": "plan",          "type": "str", "required": false, "example": "Pro"},
      {"name": "signup_date",   "type": "str", "required": false, "example": "2024-08-12"},
      {"name": "last_login_at", "type": "str", "required": false, "example": "2026-04-29T11:42:00Z"},
      {"name": "billing_status","type": "str", "required": false, "example": "active"},
      {"name": "url",           "type": "str", "required": true,  "example": "https://admin.example.com/users/9123"}
    ],
    "required_fields": ["email", "full_name", "url"]
  },
  "micro": [
    {"intent": "Go to https://admin.example.com",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"wait_after_load_seconds": 20}},
    {"intent": "Enter the email",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Email", "value": "ops@example.com"}},
    {"intent": "Enter the password",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Password", "value": "<password>"}},
    {"intent": "Click Sign in",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Sign in"}},
    {"intent": "Verify admin console loaded",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows the admin console with a top-nav 'Users' link"},
    {"intent": "Click Users in the navigation",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Users"}},
    {"intent": "Type the email in the search box",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Search users", "value": "alice@example.com"}},
    {"intent": "Click the matching user row to open their profile",
     "type": "click", "budget": 8, "grounding": true, "section": "setup",
     "required": true,
     "hints": {"layout": "single"}},
    {"intent": "Verify the profile page loaded for alice",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows the user profile with the email alice@example.com visible"},
    {"intent": "Read the URL for the profile",
     "type": "extract_url", "claude_only": true, "budget": 0, "section": "extraction"},
    {"intent": "Scroll down to expose plan / signup / last-login / billing fields",
     "type": "scroll", "budget": 5, "section": "extraction"},
    {"intent": "Read the structured profile fields from the screen",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "extraction"}
  ]
}
JSON

The hints: {"layout": "single"} on the click tells the runner this is a single labelled element (one user matched the search) — it routes to find_form_target instead of find_all_listings. Without that hint the runner would try to scan a listings grid and fail on this single-row search-result UI.

Recipe 10 — Push data into a CRM (create or update a contact)

Best for: webhook → CRM sync. Take a payload from your own system, log into the CRM, search for the contact, create-or-update.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "crm-upsert-contact-bob",
  "max_cost": 2,
  "max_time_minutes": 15,
  "micro": [
    {"intent": "Go to https://crm.example.com",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"wait_after_load_seconds": 25}},
    {"intent": "Enter the user ID",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "User ID", "value": "ops"}},
    {"intent": "Enter the password",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Password", "value": "<password>"}},
    {"intent": "Click Sign in",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Sign in"}},
    {"intent": "Click Contacts in the navigation",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Contacts"}},
    {"intent": "Click New Contact to open the create form",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "New Contact", "aliases": ["+ Contact", "Add Contact"]}},
    {"intent": "Enter the first name",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "First Name", "value": "Bob"}},
    {"intent": "Enter the last name",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Last Name", "value": "Patel"}},
    {"intent": "Enter the email",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Email", "value": "bob.patel@example.com"}},
    {"intent": "Enter the phone",
     "type": "fill_field", "budget": 4, "section": "setup", "required": false,
     "params": {"label": "Phone", "value": "+1-555-0142"}},
    {"intent": "Set the contact owner",
     "type": "select_option", "budget": 6, "section": "setup", "required": true,
     "params": {"dropdown_label": "Contact Owner", "option_label": "Alice Anderson"}},
    {"intent": "Click Save to create the contact",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Save", "aliases": ["Save Contact", "Create"]}},
    {"intent": "Verify the contact was created",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows the new contact's profile with email bob.patel@example.com visible; toast or breadcrumb confirms creation"}
  ]
}
JSON

For an update rather than a create, swap the New Contact click for a search-and-open click (see Recipe 9) and skip the fields the upstream payload didn't provide. The form-finder's scroll-and-rescan handles buttons below the fold; aliases covers Save / Save Contact / Create copy variation across CRM products.

Recipe 11 — Process a refund through a payments dashboard

Best for: customer-support refund automation. Stripe / Shopify / Square / similar dashboards all follow the same pattern: search by order ID → open the payment → click Refund → confirm amount → verify.

curl -fsS -X POST "$MANTIS_ENDPOINT/v1/predict" \
  -H "Authorization: Api-Key $BASETEN_API_KEY" \
  -H "X-Mantis-Token: $MANTIS_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<'JSON'
{
  "detached": true,
  "state_key": "refund-order-ord-91234",
  "max_cost": 2,
  "max_time_minutes": 15,
  "micro": [
    {"intent": "Go to https://payments.example.com/login",
     "type": "navigate", "budget": 3, "section": "setup", "required": true,
     "params": {"wait_after_load_seconds": 20}},
    {"intent": "Enter the email",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Email", "value": "ops@example.com"}},
    {"intent": "Enter the password",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Password", "value": "<password>"}},
    {"intent": "Click Sign in",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Sign in"}},
    {"intent": "Verify the dashboard loaded",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows the payments dashboard with a top-nav 'Payments' link"},
    {"intent": "Click Payments in the navigation",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Payments"}},
    {"intent": "Type the order ID in the search box",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Search by order, payment, or customer", "value": "ord_91234"}},
    {"intent": "Click the matching payment row to open it",
     "type": "click", "budget": 8, "grounding": true, "section": "setup",
     "required": true,
     "hints": {"layout": "single"}},
    {"intent": "Verify the payment detail loaded",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows the payment detail for ord_91234, including amount and customer"},
    {"intent": "Click Refund to open the refund dialog",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Refund", "aliases": ["Issue refund", "Refund payment"]}},
    {"intent": "Enter the refund amount",
     "type": "fill_field", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Refund amount", "value": "29.99"}},
    {"intent": "Set the reason",
     "type": "select_option", "budget": 6, "section": "setup", "required": false,
     "params": {"dropdown_label": "Reason", "option_label": "Requested by customer"}},
    {"intent": "Click Confirm refund",
     "type": "submit", "budget": 4, "section": "setup", "required": true,
     "params": {"label": "Refund", "aliases": ["Confirm refund", "Issue refund"]}},
    {"intent": "Verify the refund posted",
     "type": "extract_data", "claude_only": true, "budget": 0, "section": "setup",
     "gate": true,
     "verify": "Page shows a refund confirmation: status reads Refunded or a refund event appears in the payment timeline with amount $29.99"}
  ]
}
JSON

A note on action recipes vs. read recipes. Action recipes (post / write / refund) carry real-world consequences — a successful run mutates state on a third-party system. Two safety practices:

1. Set max_cost low (≤ $2.50) and max_time_minutes tight (≤ 15) so a runaway run can't burn the whole budget on one mistake.
2. Always include a final gate: true extract_data step that verifies the action actually happened (toast, redirect, status change). Without that gate, a run can succeed (no error thrown) while the underlying action silently failed.

---

Picking the right loop_count and max_cost

Rough calibration (Holo3 / Claude / IPRoyal proxy on the Baseten H100):

Items wanted	Recommended loop_count	Wall time	Typical cost
3-5	5-8	6-10 min	$0.50-$1.00
10	15	15-20 min	$1.50-$2.50
20	30	30-40 min	$3-$5
50+	Chunk via state_key resume	n/a	n/a

Bump loop_count higher than your target (1.5×) — it absorbs failures on spam-skipped listings and dead detail pages. Hard caps: MANTIS_MAX_LOOP_ITERATIONS=50, MANTIS_MAX_STEPS_PER_PLAN=200.

For >50 items, split into multiple /v1/predict runs sharing the same state_key and pass resume_state: true — the runner picks up at the last checkpoint.

---

Anti-patterns to avoid

Don't	Do instead
navigate_back (browser back button) — Holo3 misclicks it ~30%	Use a fresh navigate step to the search URL
Plans without a gate step after navigate	Always add a Claude-verified gate so you halt fast on Cloudflare/wrong-page
max_cost: 0.50	Hits the cap mid-extraction; budget at least $1.50 for a 10-listing run
Hardcoding "Sponsored" skip logic in intent strings	Set spam_indicators in the schema; the runner skips automatically
claude_only: false on extract_data	Holo3 isn't a vision-language model; always use Claude for extraction

---

Next

• Generic CUA usage — the request envelope + tenant setup.
• Embedding MicroPlanRunner — when you need pause/resume, host tools, or library-level control.
• Plan formats — every micro-step field documented.