Shared references

Several skills link to the same reference material. Instead of repeating it, the plugin ships it once, in a shared folder beside the skills. A skill points at it with a relative link, and the AI reads a file only when a skill's text sends it there.

The shared folder holds no skill, so it does not load as a skill and costs nothing until a skill sends the AI to read one of its files. It has two parts:

The "Linked from" column lists the skills whose text links to the file.

operations.md

What it is. A generated list of every operation the server exposes, grouped by tool. It is produced from the server's own dispatch tables and is not edited by hand. It exists so that the AI does not guess an operation name from a pattern.

What it contains.

• A header naming the server version it was generated from and the number of operations.
• The call convention: every operation is called by naming the tool and the operation, with all inputs passed together as one group. Passing inputs as separate top-level items is the most common call error.
• A rule: if an operation is not in the list, it does not exist, and nothing should be inferred from a naming pattern.
• One table per tool (ea_model, ea_diagram, ea_analyze, ea_mdg, ea_validate, ea_repository), each row giving the operation name and the first line of its description. ea_validate has one row, audit.

Which skills rely on it. ea-start-here points to it to confirm an operation exists before calling it. ea-help uses it as the second of three checks before concluding that a task has no server route.

Notes. This file is a convenience for the AI. The authoritative list of operations is MCP server: tools and operations, starting at How tools and operations work, and the server answers an unknown operation name with an error that lists the valid names.

westbrook-example.md

What it is. The canonical specification of Westbrook Bank, the fictitious organization behind every example, walkthrough, test fixture and demo in the skills. It exists so that no real customer's name, model or metamodel shape ever appears in an example, and so that two skills that give examples do not describe two different technologies.

What it contains.

Which skills rely on it. The skills listed in the table above link to it when they write an example.

Notes. Westbrook is an illustration, not a prerequisite. The model WestbrookBank.qea and the WBA technology are AI Power Tools' own test material and are not shipped. You substitute your own technology, stereotypes and model. Read the tables as a worked example of the kind of values to expect. The demonstration server tables behind WBA are separate from this file; see Demo languages.

latency.md

What it is. Waiting guidance for any skill that drives the EA desktop interface with screenshots and clicks, or that makes a COM call that can raise a blocking dialog. EA is a single-threaded desktop application. A slow operation blocks its interface, and a screenshot taken too early shows the state from before the operation finished.

What it contains. A wait-time table and a standard pattern.

The pattern: act, wait the interval for that kind of operation (not a flat number), take a screenshot, and never retry an action without first confirming the previous one failed. If EA's title bar shows "(Not Responding)", EA is processing and has not crashed. Wait five more seconds and look again. Do not double-click, re-issue the command or open a second EA instance. For a COM call that can raise a modal dialog at once (the "Profile already exists. Overwrite?" prompt on an MDG Technology import is the example), take one screenshot immediately, before the wait, and a second one after.

Which skills rely on it. The skills in the table above. ea-com cites it for every query whose text has not been validated.

Notes. The intervals are working guidance, and a slow machine or a large model can need longer. They apply to computer-use sessions. In a session without computer use, you take the screenshots and paste them into the conversation instead; see ea-ui-verification.md.

file-encoding.md

What it is. The encoding rule for MDG Technology files, shared by the skills that read, write or deploy them.

What it contains.

• The rule: an MDG Technology file must declare and use UTF-8, and the declaration at the top of the file and the real bytes must agree. EA rejects a file outright when they disagree, so on a hand-edited or received file both must be checked, not just one.
• Read and write patterns that name UTF-8 explicitly.
• Handling of older files that used the Windows 1252 encoding: re-save the file as UTF-8 and update the declaration together, because changing only one recreates the mismatch.
• A note that the Claude Code file-writing tool already produces UTF-8, so a file authored that way needs no extra handling.

Which skills rely on it. ea-mdg-author and ea-mdg-deploy.

Notes. The server's MDG readers try UTF-8 first and fall back to Windows 1252 for older files, so the server can read a file whose declaration and bytes disagree, which EA would reject. A file that the server reads is therefore not proof that EA will load it. See Live MDG reader and ea-mdg-deploy.

ea-ui-verification.md

What it is. The procedure for two things the API will not report honestly: when an operation failed, and when it succeeded but produced the wrong picture. It states the principle that EA reports success it has not earned, and that a success response is a claim and not evidence.

What it contains.

Part 1, blocking error dialogs.

• The symptom. An operation never returns and nothing reports an error. EA reports a failed operation by opening a native modal dialog and waiting for a person to click OK, holding the connection. A long wait, Windows reporting EA as "Responding", and near-zero processor use all look the same as a deadlock and prove nothing.
• The rule. Look at EA's screen before concluding anything. Do not report a hang, kill a process, restart EA or retry until you have.
• The procedure. Take a screenshot first. Read every dialog. Identify the cause. With computer use, the AI reads the message, uses it to correct the call and dismisses the dialog. Dismiss from the front, because dialogs stack and a later call can be queued behind one raised by an earlier call. Fix the cause, then retry. Confirm recovery with ping and then get_repository_info.
• Reading query failures. The dialog text names the failed query. A table lists common causes: a function the backend lacks (a local .qea file is SQLite, so use the SQLite function for byte counts), syntax from another database dialect such as TOP against SQLite, and a wrong column name (check with describe_table).
• Other frequent dialogs, each with what to do: an XMI import colliding with existing content, an export path that expects another workspace, the expected overwrite prompt on an MDG Technology import, and a duplicate-profile warning. The procedure also points to EA's System Output panel for warnings that never become dialogs.

Part 2, verifying that a diagram looks right. Every diagram create or edit ends with a screenshot. The procedure: make the change, open the diagram, reload it if rows were written directly (otherwise EA can show a cached render of the old picture), wait, then screenshot and compare with the intent. A checklist covers elements visible, no overlap, expected connectors present and suppressed ones absent, styling applied, real containment rather than visual overlap, and readability at normal zoom.

Without computer use. Say so. For a blocked call, 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. For a diagram, you look at it and say what you see. The file gives wording for a blocked call and for a diagram, and forbids reporting "EA is hung", "EA crashed" or "the diagram is ready" without either looking or asking.

Which skills rely on it. Sixteen skills link to it: ea-start-here, ea-modeling, ea-navigation-diagrams, ea-model-hygiene, ea-change-management, ea-mdg-assess, ea-mdg-author, ea-mdg-deploy, ea-mdg-model-build, ea-validation, ea-import-export, ea-stakeholder-reporting, ea-reporting-database, ea-com, ea-help and ea-diagnostic. The Prolaborate session reference points to it as the EA-side counterpart. Most of those skills carry a short "Verify in EA's interface" section that summarizes it.

Notes.

• The procedure needs a way to see EA: computer use, or a person who pastes a screenshot into the conversation. Without either, the file says to hand the check to you and not to guess.
• The server's SQL guard prevents the most common query dialog, a wrong column name on a known table. It does not check unknown tables.
• For diagrams, verify_diagram always reloads before it renders, which removes the cached-render trap. It still reports geometry, not whether the picture reads well.

prolaborate-session.md

What it is. The shared guide for every skill that drives the Prolaborate web interface. Prolaborate is a single-page web application over an API that requires an access token. The file's premise is that most confusing failures come from one of three assumptions: that the session cookie authenticates the API, that the page holds a live token, or that an administrator's experience resembles a normal user's. Its governing principle is that the AI rides your session and never creates one: it does not register an application, generate a security token or ask for a client secret.

What it contains.

1. The cookie does not authenticate the API. Pages use a session cookie. The API accepts only a bearer access token. A "not authorized" reply on a cookie-only request proves only that no token was sent.
2. Where the token is. The application keeps its sign-in record in the browser's session storage. A short script finds the record at run time and narrows immediately to the access token. The rule "read through the API, act through the interface" follows: structured data from the API for anything the AI needs to know or verify, the interface for anything it needs to change.
3. The token expires in minutes and the page does not keep it fresh. The file reports a roughly four-minute lifetime and renewal only on navigation, which is an observation of the hosted Prolaborate service and not a documented contract, so the stored token is routinely stale. The AI re-reads it immediately before each burst of calls. The file lists the error codes to recognize: ID2019 (aged out: navigate, re-read, retry once), ID2095 (authenticated but not permitted), ID2025 (a service token and not a session token), and an HTTP 422 reply saying a user identifier must not be empty.
4. Read the claims before attempting anything. The token's payload says whether the person is a Super Admin, whether they are read-only, which groups they belong to (which drives Access Permissions), and who an action will be attributed to. A Super Admin session is not a test of what a restricted user sees.
5. The server can end the session. A live connection can push a logout. If a page becomes the login screen mid-task, the AI stops and tells you. It does not retry.
6. Version pinning. Screens move between releases. Each prol- skill states the version it was verified against. If the running version is newer, the AI prefers API reads and says the interface flow is unverified.

Three rules sit under section 2:

Which skills rely on it. prol-start-here, prol-dashboards, prol-matrix and prol-impact-analysis. The first tells the AI to read the file before driving anything.

Notes. The token lifetime and renewal behavior are what the file reports from observation of the hosted Prolaborate service. The Prolaborate skills were verified as a Super Admin. See Limitations and known issues.