Do something the tools cannot reach

Goal

Carry out an EA task that no operation covers, by finding the written procedure in Sparx's user guide for the exact EA version and driving it in EA's interface, or by scripting it through EA's automation interface. For any user or administrator who has reached the edge of the operations. You end with the task done, a record of the guide version the procedure came from, and the result verified through a different means from the one that acted.

Where it runs

This scenario needs the MCP server and, for the procedures it drives in EA's interface, computer use. Scripting through EA's automation interface needs local code execution.

Before you start

• EA running with a model open. Without it there is no version to match.
• Computer use, if you want the AI to drive EA's interface. Without it, the AI hands you the steps and you tell it what you see.
• Local code execution, with a Python environment on the machine where EA runs, only if the task goes through scripting (ea-com).
• For a task that changes the model: a baseline of the affected package first.

The server covers a large part of EA, and EA's automation interface covers more. What is left is real: dialog-only settings, wizard pages, and anything EA exposes only through its interface. The fallback is ea-help. It reads the procedure from Sparx's own guide for the running version, then drives EA's interface, or hands the steps to you.

This is the last resort, and the scenario starts by checking that it is the right one. Three ways exist to do any EA task: an AI Power Tools operation, EA's automation interface, and EA's own interface. A limitation is real only when all three fail. "There is no operation for this" is a reason to check the other two, not a conclusion.

An illustration: you need to see every technology that EA lists, including disabled ones, to find out why a Westbrook stereotype is not available. EA offers no list of technologies through automation, so the operations can read a technology you name, and the copies the model stores, but cannot enumerate every disabled registration on the machine. The Manage Technologies dialog, with "Hide disabled Technologies" unticked, shows all of them. The same approach applies to any dialog-only setting.

Steps

1. AI ea-start-here, ea-help and ea-com confirm that no skill, operation or script covers the task. The AI says plainly that this is the fallback. If scripting can do it, ea-com provides a short script instead.
2. AI Power Tools ea-help establishes the EA version from the repository read.
3. AI Power Tools ea-help establishes the repository type: a local file, a server connection, or a cloud connection. The same task has different instructions for each.
4. You Resolve any ambiguity in the version or repository type. Three outcomes stop the work: EA is not running, the version has no guide, or the version was given and not read. The AI does not fall back to the nearest version.
5. AI Power Tools ea-help resolves Sparx's published user guide for that exact major and minor version and finds the page, using two to four plain words in the guide's search, then the version's sitemap.
6. AI ea-help presents the procedure: which guide version it came from, which repository type it assumes, and the steps in EA's own wording. The Notes section of a page is read first, because "unavailable in a cloud repository" tends to live there.
7. You Approve the procedure before it runs.
8. AI with computer use; otherwise you ea-help carries it out in EA's interface, one step at a time, taking a screenshot before the first click and after each dialog or wizard page. A ribbon path that does not lead where the guide says is the version mismatch surfacing: the AI stops and re-checks the version rather than hunting for a similar control. Without computer use, it gives you the steps and asks you to report what the final dialog shows.
9. AI Power Tools ea-help verifies the result through a different means from the one that acted: reading the element back, a direct model query, the language-state reading, or a screenshot. If the two disagree, the verification wins.

Two traps in the guide's web addresses both return success. Requesting a bare version folder redirects to the current version's welcome page, and an unknown page under a real version does the same. A success reply proves nothing, so the AI checks that the final address, after redirects, still carries your version and the page asked for. Never read the guide in bulk: some pages run to tens of thousands of characters. Extract the block you need.

To bring EA to the front, click its taskbar button. Never launch EA again: that opens a second instance.

Prompts to use

Opening

Follow-ups

The approval in step 7.

The AI handing the steps to you.

The verification in step 9.

Recommended models

Ratings are guidance based on the kind of reasoning the skill needs, not measured benchmarks.

Checking the result

• Accurate. The reported EA version and repository type match what the repository read returned. The page the AI followed carries your version in its final address. The verification used a different means from the one that acted, and it agrees.
• Complete. The write-up states the guide version, the repository type, the pages read, the steps taken and what verification showed. If the guide did not answer the question, the write-up says exactly that and lists what was searched.
• Nothing unintended changed. When the task changed model content, compare the model with its baseline. Read the final dialog for any option you did not choose: a screen action can complete against the wrong object without complaint.

Variations

• Technology deployment. The "Save Package as MDG Technology" step in Build an MDG Technology from a source model, and the "Import to Model" step that puts a technology file into a model in Deploy an MDG Technology and prove it works, follow the same pattern: each is done in EA's interface.
• Automation, not the interface. When the function is exposed by EA's automation interface, a short script replaces the clicking, where local code execution is available. Prefer it for a task you will repeat.
• A format, not a dialog. When you need a file format rather than a dialog, the answer is usually on the page of the automation class that consumes it, not on the task page.
• No guide for your version. The AI stops and says so. It does not fall back to a neighboring version.

Limitations

• Slower and less verifiable than any operation. The skill is a documented fallback, not a supported path.
• The guide documents dialogs thoroughly and file formats unevenly. "The documentation does not specify this" is a legitimate answer. The AI records what was searched.
• A malformed query sent through scripting raises a modal dialog, not an error in the script, and it blocks until the dialog is dismissed. An unattended script hangs. With computer use, the AI reads the dialog, uses the message to correct the script, and dismisses it. Without computer use, you take a screenshot of the dialog, paste it into the conversation so the AI can analyze it, and dismiss the dialog.
• EA reports success it has not earned, and reports failure as a dialog. After any screen action, the AI screenshots, reads the title of what is on screen, and verifies.
• A server connection or a cloud connection can offer different or missing options compared with a local file. Establish the repository type before reading the procedure.