ea-start-here

Logic: AI + AI Power Tools checks. The AI routes and judges; read-only server calls establish the facts.

What it is for

Use it to begin every session on solid ground, and to route a request when you are not sure which skill fits. You end with the server, the open model and the loaded technologies confirmed, and your request routed to the right skill.

ea-start-here is the entry point for any Sparx EA task. It does two things. It runs a short check, the session preflight, so the AI knows what it is actually connected to, and it routes the request to the right skill so the AI does not guess.

Its premise is that most EA failures are not modeling errors. They are sessions acting on a model that is not open, a technology that is not installed, or a repository different from the one assumed. The session preflight is four steps and costs a handful of small calls. It is a different thing from the reporting preflight, the dry run that ea-reporting-database makes before a build (ea-reporting-database).

Capabilities it uses

When to use it, and when not

Use it first, whenever a session involves an EA model and before reaching for a more specific EA skill. It is cheap, and the more specific skills assume the facts it establishes. The full routing map is in The routing map. A skill that is not named in this skill's own table still triggers from its own description.

• Not for Prolaborate. Prolaborate is a separate product driven through a browser. Start at prol-start-here.
• Do not use ea-help early. It is the last resort: it works from Sparx's user guide and drives the interface, which is slower and less verifiable than any operation.
• Choosing between writing a technology directly and building it from a profile package is the most common routing mistake. Use ea-mdg-author when writing the technology directly. Use ea-mdg-model-build when the stereotypes already exist as a profile package inside a repository and EA should generate the technology from them. Building by hand what already exists as a model is wasted work, so the AI checks for a profile package first.

How it works

The session preflight. Four steps, run in order, stopping at the first that fails.

1. Is the server alive, and does it match these skills? AI Power Tools The AI calls ping with the release of the skills, which returns the server identity, its version and an update summary. When the server or the skills are behind the current release, the reply carries a notice with the steps to update, and the AI shows it to you word for word. A reply without an update summary comes from a server older than the skills, so the AI gives you the steps to update the Claude Desktop extension. The AI notes the version, because skills depend on a minimum server version and a mismatch explains a missing operation faster than debugging. If there is no ping at all, the AI says in one sentence that the session has no AI Power Tools tools, and works out which of three causes applies, asking you where the session runs if it cannot tell:

   • A Claude Cowork or other cloud session. A cloud session cannot reach a local MCP server. You move to the Code tab of the Claude desktop app, or another session that runs on the Windows machine where EA runs.
   • Claude Desktop timed out connecting its local servers when the session started. This is the usual cause in a local session. You reconnect the extension in Settings, Extensions, or start a new session.
   • The extension is not installed, or is turned off. The AI gives you the download link and the install steps.

   It then stops the model work and attempts no EA call. What it can still offer is reading what the server or its engines already wrote into a folder you have connected: XMI and image exports, generated reports, Parquet files and generated Power BI projects. Files a cloud session reads from a connected folder are processed on Anthropic's servers, and a session without the server cannot change the model.

2. What repository is open? (AI Power Tools, then AI) The AI calls get_repository_info, which says whether EA is running and which model is open, and acts on the answer:

   • EA is not running: the AI says so and does not attempt model operations. It offers to start EA, with launch_ea or, where computer use is available, by operating the desktop. Launching an application on your desktop is your decision, so the AI asks. You Once you agree, launch_ea starts EA AI Power Tools, or the AI starts it through the desktop. AI with computer use; otherwise you
   • A different model is open than expected: the AI stops and confirms which is intended. Writing to the wrong repository is the most expensive mistake available, and it is silent. You
   • No project is open: open_project with a path AI Power Tools. Opening the model through EA's own File menu, and dismissing any dialog the open raises, is AI with computer use; otherwise you.

3. Which modeling technologies are live? (AI Power Tools, rule-based) One call, assess_mdg_situation, reports which technologies are loaded and enabled, how much of the model conforms to them, and how many stereotypes are used outside any of them. It classifies the situation by rule: a client-built technology that covers all but a small share of stereotyped elements (under 5 percent outside it) counts as conforming, and anything above that as accumulating ad hoc usage. This is a rule-based judgment and it can be wrong. The AI then asks a finer question if it needs one:

   Question	Operation
   What is the overall situation?	assess_mdg_situation
   What is registered on this machine?	list_registered_technologies
   What does this model store, and at which Location (Project or Model)?	get_embedded_mdgs
   Is this specific technology live?	get_mdg_from_runtime, which needs a technology id and answers about that one technology only

   These calls answer different questions, and the difference between them is usually the bug: a technology can be embedded but not loaded, or registered but not enabled (EA itself reports a technology disabled in Manage Technology as loaded, so get_mdg_from_runtime reports it as registered but not enabled), or stored at both Locations of the model so that one copy shadows the other. Installed is not active.

4. Confirm before writing. (AI, you) If the task changes the model, the AI states what it is about to change and in which repository, and gets your agreement before the first write.

Routing.

5. Pick the skill. AI The AI matches the request to the routing map. It treats ea-help as a last resort.

When it goes wrong.

6. Escalate to ea-diagnostic instead of guessing a third time. AI The triggers are: an operation fails twice with the same error after the input was corrected; EA runs but behaves inconsistently (an element that exists is not found, a tagged value does not stick, a diagram does not show what the model contains); a technology passes every check but its stereotypes are still unavailable; or two operations contradict each other.
7. Check three common causes first. AI EA may be busy and not hung: "(Not Responding)" in the title bar means EA is processing, so wait and do not open a second instance, which can corrupt work. Looking at the title bar and any open dialog is AI with computer use; otherwise you. The technology may not be loaded, which get_mdg_from_runtime answers in one call. Or the write went somewhere else, so re-run get_repository_info and confirm the open model.

Corrections to common assumptions. The skill ends with things people assume that are not true:

• Validation runs through audit. ea_validate checks the model against a ruleset with audit. See ea_validate.
• Query, do not remember. Answers about model state come from a fresh read. A verification call is almost always under 1,000 tokens.
• A direct query is always available (execute_sql) and is the authority when no operation covers a question. EA's table and column names are inconsistent, so the AI consults the schema reference in ea-modeling, or describe_table, and does not guess. See SQL guard.
• Examples use Westbrook Bank, a fictitious organization with the WBA technology. See westbrook-example.md. Substitute your own.
• Verify in EA's user interface. AI with computer use; otherwise you If a call seems to hang, the AI looks at EA's screen, reads the dialog, uses its message to correct the call and dismisses it. After any diagram create or edit it reloads the diagram and looks at it. Without computer use it says so: you take a screenshot of the dialog and paste it into the conversation so the AI can analyze it, then you dismiss the dialog so processing can continue, and you look at the diagram and say what you see. See ea-ui-verification.md.

What you get

• A plain-language summary of the session: server version, open model, EA version, repository type, and which technologies are live.
• A statement of what will be changed, and where, before any write.
• A hand-off to the skill that fits the task, or an escalation to ea-diagnostic.

The skill changes nothing in the model.

Tools it relies on

• ping (ping): is the server alive, which version, and is an update due.
• get_repository_info, launch_ea and open_project (Project): which model is open; start EA or open a project, only after you agree.
• assess_mdg_situation, list_registered_technologies, get_embedded_mdgs and get_mdg_from_runtime (Discovering technologies): technology state.
• execute_sql (Direct queries): the fallback for any question no operation covers.
• audit (ea_validate): validation against a ruleset.
• Shared references: operations.md, westbrook-example.md and ea-ui-verification.md.

Sample prompts

1. "Before we start, check that EA is open on the right model and tell me which modeling technologies are loaded." Runs the full session preflight and summarizes it.
2. "I want to add a new application to the Westbrook Bank model." Runs the session preflight, states what it will change and in which repository, then routes to ea-modeling.
3. "Is the WBA technology actually live in this session, or only installed?" Interrogates that one technology and explains installed versus embedded versus loaded. Substitute your own technology id.
4. "We have a profile package for our metamodel inside the repository. How do I turn it into an MDG Technology?" Checks that the profile package exists, then routes to ea-mdg-model-build, not ea-mdg-author.
5. "I need our application inventory in Power BI." Routes to ea-power-bi, which asks where the data should come from before building anything: views in the EA database, which needs direct access to a SQL Server database and the right to create views in it; a separate reporting database, which needs a SQL Server to run or a SQL database in Microsoft Fabric; or Parquet files, which need neither.
6. "This keeps failing with the same error, and the diagram doesn't match what the model says." Meets two escalation triggers and routes to ea-diagnostic, after the three common-cause checks.

Recommended models

Ratings are guidance based on the kind of reasoning the skill needs, not measured benchmarks. A weaker model is most likely to skip the "confirm before write" step or to route to ea-help too early, so review those two behaviors first.

Checking accuracy and completeness

1. Session preflight facts against EA. Compare the AI's summary with the EA window: version, project file and repository type. Correct: each value matches what EA shows. Complete: the summary covers server version, open model and live technologies. Red flag: a model name the AI never read from get_repository_info.
2. Wrong-model guard. Open a different model from the one your prompt names. Correct: the AI stops and asks. Red flag: any write before the question is answered.
3. Technology state against EA's own view. Compare the AI's report with Specialize > Technologies > Manage Technology (the Location, Project or Model, and the enabled state). Correct: they agree, or the AI explains the difference, and a technology you see disabled there is reported as registered but not enabled. Red flag: "all technologies are loaded" when the AI asked about no technology in particular. In some EA builds these checks can disagree with that dialog, in which case the disagreement is a finding to report, not an AI error.
4. Routing correctness. Give ten requests with a known right skill, including two MDG requests (one with a profile package, one without) and one request only EA's interface can do. Correct: the first skill named matches the routing map. Red flags: ea-help named before skills, operations and COM were considered; ea-mdg-author chosen when a profile package exists.
5. No invented operations. Correct: validation is run with audit under ea_validate. Red flag: the AI names a validation operation that ea_validate does not list.
6. Write confirmation. Read the transcript before the first write. Correct: the message names the repository and what will change, and waits. Red flag: a write with no such message.
7. Escalation threshold. Correct: after two identical failures the AI switches to ea-diagnostic. Red flag: a third variant of the same call.

Limitations

• The routing table inside the skill is not exhaustive. Seven skills (ea-change-management, ea-model-hygiene, ea-mdg-assess, ea-diagram-advisor, ea-diagram-composition, ea-import-export, ea-stakeholder-reporting) are chosen from their own descriptions, and Prolaborate has its own entry skill. The routing map gives the full map.
• The technology checks can disagree with EA's Manage Technology view in some EA builds. get_embedded_mdgs reads both Locations where a model stores a technology, Project and Model, so an empty answer means the model stores no technology. A technology loaded from a file on the workstation is not stored in the model and does not appear in that answer. Cross-check in EA's user interface. See ea_mdg.
• Windows and a running EA are required. The server attaches to a running EA and reports when none is open. The session preflight starts EA only with your agreement, through launch_ea or, with computer use, through the desktop. See Limitations and known issues.