record-workflow

name	record-workflow
description	Record an automation workflow by performing actions, then replay it later as a playbook. Capture clicks, typing, navigation into reusable sequences. Use when: "record this", "capture steps", "make repeatable", "save automation", "create playbook", "record and replay", "macro", "record workflow".
disable-model-invocation	true
allowed-tools	["mcp__sh__playbook_record","mcp__sh__export_playbook","mcp__sh__platform_guide","mcp__sh__memory_recall","mcp__sh__memory_save","mcp__sh__memory_query_patterns","mcp__sh__job_create","mcp__sh__job_run","mcp__sh__job_status"]

Workflow Recording & Replay

You are recording a desktop or browser automation workflow into a reusable playbook, then replaying it via the job system.

Intelligence Wrapper

Every tool call returns automatic hints. The recording system (playbook_record) works with the intelligence wrapper — when recording is active, mcpRecorder.captureToolCall() captures each tool call as a playbook step. Watch for [HINT], [WARNING], and [STRATEGY] lines for guidance.

Existing Playbooks

Before recording a new workflow, check if one already exists:

• memory_recall(task="record {workflow}") — check past recordings
• memory_query_patterns(scope="{platform}") — check verified patterns

Available playbooks: x-twitter, linkedin, instagram, reddit, threads, youtube, discord, davinci-color-grade, davinci-edit-timeline, davinci-render, google-flow-create-project, google-flow-generate-image, google-flow-generate-video, google-flow-edit-image, google-flow-edit-video, google-flow-open-project, google-flow-search-assets, codex-desktop.

Recording a Workflow

Start Recording

playbook_record(action="start", platform="{platform_name}")

• platform is a label (e.g., "gmail-compose", "slack-message", "figma-export")
• From this point, tool calls made within the SAME MCP server session are captured as playbook steps via the intelligence wrapper's mcpRecorder.captureToolCall().

Important: Recording works within a single MCP server session. Because Claude Code may restart the MCP server between tool calls, recording is best suited for workflows done in quick succession. For long multi-step workflows, consider using export_playbook after the session instead — it reconstructs the playbook from session memory (actions.jsonl).

Perform the Workflow

Now execute the workflow normally using any ScreenHand tools. For example:

• browser_navigate → captured as navigate step
• browser_click → captured as browser_click step
• browser_fill_form → captured as browser_type step
• ui_press → captured as press step
• key → captured as key_combo step
• menu_click → captured as menu_click step
• type_text → captured as type_into step

Check Recording Status

playbook_record(action="status")

Shows how many steps have been captured so far.

Stop Recording & Save

playbook_record(action="stop", name="{workflow_name}")

• Saves to playbooks/{workflow_name}.json
• The playbook is immediately available for replay

Replaying a Workflow

As a Job (recommended)

job_create(task="Replay: {description}", playbookId="{workflow_name}")
job_run(jobId="{id}")

The job system executes each playbook step through the PlaybookEngine with:

• Variable substitution ({VAR_NAME} in step fields)
• Optional popup detection (if observer is running)
• Automatic verification after each step
• Error handling with retry

Check Replay Progress

job_status(jobId="{id}")

Shows which steps completed, which are pending, any errors.

Exporting Platform Knowledge

After exploring a platform (manually or via platform_explore), export what you learned:

export_playbook(platform="{name}", domain="{domain.com}")

This extracts from your session memory:

• URLs visited
• Selectors that worked
• Errors encountered and resolutions
• Strategies used

Output is saved to references/{platform}.json — usable by platform_guide in future sessions.

Editing Playbooks

The saved playbook JSON can be hand-edited. Structure:

{
  "platform": "my-workflow",
  "steps": [
    {
      "action": "navigate",
      "url": "https://example.com",
      "description": "Open the site"
    },
    {
      "action": "press",
      "target": "Login",
      "description": "Click login button"
    },
    {
      "action": "type_into",
      "target": "input[name='email']",
      "text": "{EMAIL}",
      "description": "Enter email"
    },
    {
      "action": "key_combo",
      "keys": "enter",
      "description": "Submit form"
    }
  ]
}

Supported Action Types

navigate, press, type_into, extract, key, key_combo, menu_click, scroll, wait, screenshot, browser_js, browser_click, browser_type, browser_human_click, cdp_key_event

Variables

Use {VAR_NAME} in any step field. Pass values via job_create(vars={VAR_NAME: "value"}).

Optional Steps

Add "optional": true to a step — failure continues instead of aborting.

Verification

Add "verify": "selector" to a step — after execution, checks if the selector exists.

Best Practices

1. Name playbooks descriptively: slack-send-message, figma-export-png, gmail-compose-reply
2. Use variables for dynamic data: {POST_TEXT}, {EMAIL}, {FILE_PATH}
3. Add descriptions to each step: Makes the playbook readable and debuggable
4. Record on a clean state: Start from a known state (fresh page load, app just launched)
5. Test replay immediately: After recording, replay once to verify it works

$ARGUMENTS