Imported from robomotionio/agent-skills (skills/creating-flow/SKILL.md). Install upstream with npx skills add robomotionio/agent-skills --skill creating-flow. Copyright stays with the author.

Robomotion Flow Builder

Robomotion is an RPA platform with a TypeScript SDK and a visual node editor. This skill is a thin index over the reference docs in ./docs/. Read the relevant doc when a topic comes up — don't try to memorize it from this file.

Hard Rules (SDK rejects violations at validate time)

The SDK enforces these. Violations throw at robomotion validate / build with descriptive messages, so the agent never silently produces broken flows:

1. Node IDs MUST be 6-char lowercase hex — /^[0-9a-f]{6}$/. f.node(), .then(), f.edge() reject non-hex IDs ('begin', 'label', 'maps', uppercase) at registration. Pick fresh hex per node. See ./docs/reference/id-format.md.
2. Subflow node ID = subflows/<id>.ts filename, exactly. Both must be 6-hex. The Designer's "enter subflow" UX depends on the match.
3. f.addDependency(namespace, version) is validated against the live package index. version must be concrete ('latest' is rejected) and must exist in the package's published versions list. namespace must exist in https://packages.robomotion.io/stable/index.json. Run robomotion get packages <ns> or robomotion describe package <ns> to resolve real versions before calling addDependency. Never invent a version.
4. Terminal nodes (Debug, Log, Stop, GoTo, End, WaitGroup.Done) have 0 outputs — wire TO them via f.edge(), never .then() from them.
5. Every Core.Flow.GoTo references a Core.Flow.Label id that exists in the same flow file.

Required First Line

Every flow file (main.ts and every subflows/*.ts) starts with this exact import — copy verbatim, including helpers you don't currently use (Bun won't flag dead imports, but missing ones become runtime ReferenceError):

import { flow, Message, Custom, JS, Global, Flow, Credential, AI } from '@robomotion/sdk';

For library files swap flow for library / subflow. Full reference: ./docs/reference/imports.md.

Builder grammar

• f.node(id, type, name, props) — param order. Only emit non-default props (Go runtime fills defaults from pspec).
• .then() for sequential, .edge() for multi-port wiring.
• Message(name) for variables · Custom(value) for literals · JS(expr) for one-line JS · Credential({vaultId, itemId}) for secrets.
• A field takes a scope helper based on its TYPE IN THAT NODE, not its in*/opt* name — and the SAME name can differ across nodes. Every in*/out* port takes a scope helper (Custom('…') / Message()); a bare literal there is silently dropped. For opt* fields, don't guess from the name — check get_node_schema: a field typed object + variableType (even a numeric one) takes a scope helper; a plain number/boolean/enum field takes a bare literal and must NOT be wrapped. The very same property can differ by node: Core.Browser.OpenLink optTimeout is a plain number → optTimeout: 32, but Core.Browser.WaitElement optTimeout is variableType:Integer → optTimeout: Custom('30'). Mismatching either way VALIDATES but FAILS TO LOAD on the robot (flow_error: failed / Config parse error, no nodes run): wrapping a plain field sends an object to a scalar; leaving a variable-backed field bare sends a scalar into a {scope,name} slot. Enums/booleans are always plain (optBrowser: 'chrome', optMethod: 'post', optInsecure: true); in* ports and value fields like optUrl/optDownloadDir/optNofBranches always take Custom()/Message().
• func is a literal string (NOT JS()).
• Common runtime props also take raw values: delayBefore: 2, delayAfter: 0.5, continueOnError: true.
• ES5-only inside func: no =>, no template literals, no const/let, no destructuring. No require() / fs / Buffer / process (pure JS sandbox).
• Loops: Label → ForEach → body → GoTo. Stop is standalone, wired via f.edge() on ForEach port 1.
• Library projects use library.create(id, name, fn) with Begin/End nodes (no .start()). Inline subflows use subflow.create(name, fn).
• Every flow ends with .start(). Every flow has a Core.Flow.Stop node.
• Core.* packages (Core.Trigger, Core.Browser, Core.Programming, Core.CSV, Core.Flow, Core.Vault, Core.Net, Core.Excel, …) are embedded in the robot — NEVER call f.addDependency('Core.*', …). The Designer auto-loads them. Only call f.addDependency(ns, ver) for non-Core.* packages. When updating an existing flow, NEVER bump existing addDependency versions; only add missing ones.
• Comments & canvas layout — Core.Flow.Comment nodes (with an optText markdown string) title the flow and fence its logical phases; the visual arrangement — node positions, comment box colors/sizes, and Sugiyama-style layering — lives in main.designer.ts. Layout is cosmetic (never affects runtime) but it's what makes a flow readable. See ./docs/patterns/comments-and-layout.md.

Full grammar: ./docs/sdk-grammar.md. Architecture: ./docs/architecture.md.

Diagnostic map

Map an error symptom to the doc that fixes it. When validate_flow fails, look up the symptom here before reading the full failure trace.

Drift-prone reminders before every Write / Edit of flow code:

• Never output TypeScript as chat text — always use Write / Edit. Plans and explanations stay in chat.
• Hex IDs from the start. Cross-references (optNodes.ids, Catch.optNodes.ids, subflow filenames) must use the same hex.
• For browser flows: explore the live page first (Skill(exploring-browser) or mcp__browser__* after ToolSearch warmup). Don't guess selectors. Core.Browser.* element nodes (ClickElement/TypeText/GetValue/SetValue/WaitElement/Select) default inSelector to XPath — translate CSS handles you find (#email, input[type="email"]) to XPath (//input[@id='email']) and omit inSelectorType; use a CSS string ONLY with inSelectorType: 'css' (plain literal — inSelectorType is an enum, so NEVER Custom('css')). A CSS string with the default engine fails at runtime with "element not found". Never write inSelectorType: 'xpath' (invalid; the value is xpath:position). Also: enum/dropdown opts (optBrowser, optProxy, optProxyAuth, optClickType) take a PLAIN string/boolean — NEVER Custom(); wrapping an enum in Custom() emits a {name,scope} object and the robot rejects the node at load with Config parse error (flow never starts). Custom()/Message() are only for variable value fields (selectors/URLs/text/paths). See ./docs/patterns/browser.md.
• For any flow that READS or WRITES tabular data (CSV / Excel / Google Sheets / Excel 365 / SQLite / Airtable / Pandas / DataTable / DOMParser) — read ./docs/patterns/data-tables.md BEFORE adding the node, both for the Function that builds the table AND for the reader/writer node. That doc names the exact node and shows its properties (e.g. write CSV = Core.CSV.WriteCSV with inFilePath + inTable; write Sheets = Robomotion.GoogleSheets.SetRange; etc.) and the {columns: [...], rows: [{key: value}]} format (never {header: ...}, never rows-as-arrays). Do NOT unified_search / search for data-output nodes — search returns TEMPLATES, not nodes, and looping on it wastes the turn. The node names are in data-tables.md; once you know the node, use get_node_schema for its exact properties. When a search returns templates instead of the node you need, stop searching and read the relevant pattern doc.
• For any Robomotion.ChatAssistant flow in conversational mode — read ./docs/patterns/conversational-chat.md BEFORE writing it. One user message is one ChatIn → ChatOut run and ChatOut is the only thing that unlocks the composer, so a branch that ends anywhere else (an error, a missing ChatOut) freezes the chat until the page is reloaded — the most common bug in these flows, and it looks like a product fault rather than a flow fault. That doc also carries the streaming wiring (Callback In stream_delta → Streaming Text, and no Text node repeating the answer) and the attachments wiring (GetAttachments, because msg.payload.files is names and versions, never files on disk).
• Validate BEFORE save — save_flow only compiles, it does NOT pspec-validate.

Pattern reference

Read these docs before writing the corresponding code:

References:

For schemas, examples, and package docs, use the robomotion CLI (it's already on PATH, call it by bare name):

Public templates repo: github.com/robomotionio/robomotion-templates is the canonical source. Prefer cloning/forking a matching template over building from scratch.

Workflow

Full step-by-step: ./docs/workflow.md. Outline:

0. Gather requirements (interactive only) — credentials (commit to a vault-item pick, don't quiz the user), URLs, files, iteration, error handling.
1. Discover — robomotion search, robomotion get nodes, robomotion docs <namespace> (MANDATORY for every non-Core.* package).
2. Plan — output plan as chat text, then AskUserQuestion(["Build it", "Modify plan"]).
3. Write — read 1-2 relevant ./docs/patterns/*.md, verify property names with robomotion describe node, then Write main.ts (and any subflows/<id>.ts). For browser flows: explore live first.
4. Validate — call validate_flow MCP tool. Pspec-checks AND dependency-checks. MUST run BEFORE save.
5. Save — save_flow if registered (Designer / pi); else git commit && git push from inside the flow dir. This is the terminal step. Stop here and report success — do NOT chain into running the flow. Running is a separate user request handled by the running-flow skill.

If invoked in direct mode ("Write main.ts for X", "Generate a flow that does Y"), skip 0-2 and jump to 3.

Browser caveat: if code changed after the initial exploration (different selectors, new actions), re-verify selectors against the live page before saving. Selectors are owned by Step 3, not a post-save step.

Canonical example (simple chain)

For loop / conditional / subflow / catch examples, see the corresponding pattern docs — they have richer working snippets.

import { flow, Message, Custom } from '@robomotion/sdk';

flow.create('main', 'Simple Flow', (f) => {
  f.node('42ec21', 'Core.Trigger.Inject', 'Start', {})
    .then('7dbafc', 'Core.Programming.Function', 'Setup', {
      func: `msg.url = 'https://example.com'; return msg;`
    })
    .then('a06926', 'Core.Browser.Open', 'Open Browser', {
      outBrowserId: Message('browser_id')
    })
    .then('8e1c4b', 'Core.Browser.OpenLink', 'Navigate', {
      inBrowserId: Message('browser_id'),
      inUrl: Message('url'),
      outPageId: Message('page_id')
    })
    .then('d52f73', 'Core.Browser.Close', 'Close', {
      inBrowserId: Message('browser_id')
    })
    .then('b9a841', 'Core.Flow.Stop', 'Stop', {});
}).start();

CLI & MCP

• robomotion — self-sufficient CLI. Builds, validates, runs, searches, inspects. robomotion help for the full verb list.
• robomotion-browser-mcp — MCP server for interactive browser exploration (used by exploring-browser and mcp__browser__* tools).

The robomotion CLI shells out to robomotion-sdk-mcp internally for search-backed commands and calls api.robomotion.io directly for run/stop/vault/robot operations. No additional MCP servers required.

Regression suite

This skill ships with an automated eval suite at ./evals/ — Tier A pinpoint regressions (handcrafted fixtures, one rule each) plus Tier B integration tests (live main.ts from the public robomotion-templates repo). Run bun run skills/creating-flow/evals/run-evals.ts from the agent-skills root before committing edits to this SKILL.md or the ./docs/ files. See ./evals/README.md for adding new cases and the assertion grammar.

Related skills

• validating-flow — schema validation
• testing-flow — behavioral tests
• running-flow — execute on robot
• searching-packages — find packages, nodes, templates
• exploring-browser — interactive browser automation
• reversing-network — convert a browser flow to HTTP after capturing traffic