Monte Carlo Prevent

Monte Carlo Prevent Skill

This skill brings Monte Carlo's data observability context directly into your editor. When you're modifying a dbt model or SQL pipeline, use it to surface table health, lineage, active alerts, and to generate monitors-as-code without leaving Claude Code.

Monte Carlo tool routing (required): Always call Monte Carlo MCP tools through this plugin's bundled server, whose fully-qualified tool names are mcp__plugin_mc-agent-toolkit_monte-carlo-mcp__<tool> (e.g. mcp__plugin_mc-agent-toolkit_monte-carlo-mcp__get_alerts). Bare tool names used in this skill (get_alerts, search, get_table, …) refer to that bundled server. If the session also has a separately-configured monte-carlo-mcp server, do not route to it — it may point at a different endpoint or credentials.

Reference files live next to this skill file. Use the Read tool (not MCP resources) to access them:

• Full workflow step-by-step instructions: references/workflows.md (relative to this file)
• MCP parameter details: references/parameters.md (relative to this file)
• Troubleshooting: references/TROUBLESHOOTING.md (relative to this file)

When to activate this skill

Prevent is the edit-lifecycle skill. Activate only when the user expresses intent to change a dbt model. Bare file mentions, table-name mentions in passing, or general health questions are not prevent's territory — those belong to monte-carlo-asset-health and will activate that skill on their own.

Do not wait to be asked. Run the appropriate workflow automatically whenever the user:

• Describes a planned change to a model (new column, join update, filter change, refactor) → STOP — run Workflow 1 first if it has not run for this table this session, then Workflow 2, before writing any code
• Adds a new column, metric, or output expression to an existing model → same rule: Workflow 1 first (if not yet run for this table), then Workflow 2; the post-edit hook will offer Workflow 5 (monitor generation) afterward
• References a model file with an edit verb in the same prompt (e.g. @models/clients/client_hub.sql add an is_active column) → same rule: Workflow 1 first, then Workflow 2

Present the W2 impact assessment as context the engineer needs before proceeding — not as a response to a question.

Workflow 1 runs silently when chained to Workflow 2

When the user expresses change intent, Workflow 1 invokes monte-carlo-asset-health purely as a data-gathering step. Read asset-health's report from your context, but do not relay the full report to the engineer — the user-facing artifact is Workflow 2's impact assessment, which already cites the relevant alerts / lineage / monitors. Showing both creates duplicate reading.

Two exceptions where you must surface output from W1 to the engineer:

1. Disambiguation prompt. If asset-health returns multiple matches and asks the engineer to pick one, surface that question — the user must choose.
2. Stop-the-world signals. If the table is already on fire (active critical alerts firing, freshness severely stale), say so in one short line before W2.

If Workflow 1 already ran for this table earlier in the session, skip directly to Workflow 2 — re-running asset-health is redundant.

When NOT to activate this skill

Do not invoke Monte Carlo tools for:

• Seed files (files in seeds/ directory)
• Analysis files (files in analyses/ directory)
• One-off or ad-hoc SQL scripts not part of a dbt project
• Configuration files (dbt_project.yml, profiles.yml, packages.yml)
• Test files unless the user is specifically asking about data quality

If uncertain whether a file is a dbt model, check for {{ ref() }} or {{ source() }} Jinja references — if absent, do not activate.

Macros and snapshots — gate edits, skip auto-context

Macro files (macros/) and snapshot files (snapshots/) are not models, so do not auto-fetch Monte Carlo context (Workflow 1) when they are opened. However, macros are inlined into every model that calls them at compile time — a one-line macro change can silently alter dozens of models. Snapshots control historical tracking and are similarly sensitive.

The pre-edit hook gates these files. If the hook fires for a macro or snapshot, identify which models are affected and run the change impact assessment (Workflow 2) for those models before proceeding with the edit.

Peer-skill redirects

These requests have their own skills — do not run prevent for them:

• "How is table X doing?" / "is X healthy?" / "check status of X" → monte-carlo-asset-health
• "Create a monitor for X" / "what should I monitor?" / "set up freshness on X" (without an active edit context) → monte-carlo-monitoring-advisor

Prevent invokes asset-health and monitoring-advisor itself when its workflows need them (W1, W5); it does not duplicate their entry points.

REQUIRED: Change impact assessment before any SQL edit

Before editing or writing any SQL for a dbt model or pipeline, you MUST run Workflow 2.

This applies whenever the user expresses intent to modify a model — including phrases like:

• "I want to add a column…"
• "Let me add / I'm adding…"
• "I'd like to change / update / rename…"
• "Can you add / modify / refactor…"
• "Let's add…" / "Add a <column> column"
• Any other description of a planned schema or logic change
• "Exclude / filter out / remove [records/customers/rows]…"
• "Adjust / increase / decrease [threshold/parameter/value]…"
• "Fix / bugfix / patch [issue/bug]…"
• "Revert / restore / undo [change/previous behavior]…"
• "Disable / enable [feature/logic/flag]…"
• "Clean up / remove [references/columns/code]…"
• "Implement [backend/feature] for…"
• "Create [models/dbt models] for…" (when modifying existing referenced tables)
• "Increase / decrease / change [max_tokens/threshold/date constant/numeric parameter]…"
• Any change to a hardcoded value, constant, or configuration parameter within SQL
• "Drop / remove / delete [column/field/table]"
• "Rename [column/field] to [new name]"
• "Add [column]" (short imperative form, e.g. "add a created_at column")
• Any single-verb imperative command targeting a column, table, or model (e.g. "drop X", "rename Y", "add Z", "remove W")

Parameter changes (threshold values, date constants, numeric limits) appear safe but silently change model output. Treat them the same as logic changes for impact assessment purposes.

Do not write or edit any SQL until the change impact assessment (Workflow 2) has been presented to the user. The assessment must come first — not after the edit, not in parallel.

Pre-edit gate — check before modifying any file

Before calling Edit, Write, or MultiEdit on any .sql or dbt model file, you MUST check:

1. Has the synthesis step been run for THIS SPECIFIC CHANGE in the current prompt?
2. If YES → proceed with the edit
3. If NO → stop immediately, run Workflow 2, present the full report with synthesis connected to this specific change. If risk is High or Medium: ask "Do you want me to proceed with the edit?" and wait for explicit confirmation. If risk is Low: use judgment — proceed if straightforward and no concerns found, otherwise ask before editing.

Important: "Workflow 2 already ran this session" is NOT sufficient to proceed. Each distinct change prompt requires its own synthesis step connecting the MC findings to that specific change.

The synthesis must reference the specific columns, filters, or logic being changed in the current prompt — not just general table health.

Example:

• ✅ "Given 34 downstream models depend on is_paying_workspace, adding 'MC Internal' to the exclusion list will exclude these workspaces from all downstream health scores and exports. Confirm?"
• ❌ "Workflow 2 already ran. Making the edit now."

The only exception: if the user explicitly acknowledges the risk and confirms they want to skip (e.g. "I know the risks, just make the change") — proceed but note the skipped assessment.

Available MCP tools

All tools are available via the monte-carlo-mcp MCP server.

Core workflows

Each workflow has detailed step-by-step instructions in references/workflows.md (Read tool).

1. Asset health pre-fetch (silent delegation to asset-health)

When: User expresses change intent for a table that hasn't been seen in this session. What: Invokes monte-carlo-asset-health via the Skill tool to gather table state (health, upstream lineage, alerts, monitors). Then makes one direct get_asset_lineage(direction="DOWNSTREAM") call to complete the picture (asset-health only fetches upstream). The combined data is used as input to Workflow 2, not shown to the engineer. Two exceptions surface to the user: any disambiguation prompt, and stop-the-world signals (active critical alerts, severe staleness).

2. Change impact assessment — REQUIRED before modifying a model

When: Any intent to modify a dbt model's logic, columns, joins, or filters. What: Surfaces blast radius, downstream dependencies, active incidents, monitor coverage, and query exposure. Reuses asset-health's data when Workflow 1 ran earlier this session; otherwise calls get_table / get_alerts / get_asset_lineage / get_monitors directly. Produces a risk-tiered report with synthesis connecting findings to specific code recommendations.

3. Change validation queries

When: Explicit engineer request only (e.g. "validate this change", "ready to commit"), or via /mc-validate run. What: Generates 3–5 targeted SQL queries to verify the change behaved as intended. Uses Workflow 2 context — requires both impact assessment and file edit in session.

4. Validate change in sandbox — invoked by /mc-validate run

When: Only when the engineer invokes /mc-validate run (in any of its forms).

Pre-flight: run does not auto-generate. If no validation/<table>_<ts>.sql exists for the changed model(s), abort and tell the engineer to run /mc-validate (or /mc-validate generate) first.

What: Two-phase workflow.

• W4.1 — Build. Parses profiles.yml, classifies the active database, detects hard-coded database: in the model's {{ config() }}, then runs dbt build --select <model> into the engineer's dev database. Refuses to build against shared prod. Skipped automatically for YAML/docs-only diffs and for /mc-validate run --skip-build.
• W4.2 — Execute validation queries. Substitutes <YOUR_DEV_DATABASE> in Workflow-3 output with a user-confirmed value (or --dev-db <NAME> if supplied), runs a read-only pre-check on every query, executes via the Snowflake MCP, and reports per-query verdicts plus a consolidated summary.

Invocation matrix:

run accepts both flags together: /mc-validate run --skip-build --dev-db <NAME>.

5. Add monitor (delegated to monitoring-advisor, post-edit)

When: Post-edit hook injects the coverage prompt (driven by MC_MONITOR_GAP from Workflow 2), or the engineer explicitly asks to add a monitor. What: Asks "Generate monitor definitions? (yes/no)". On yes, invokes monte-carlo-monitoring-advisor via the Skill tool with the model name and changed columns/logic. Prevent's responsibility ends at delegation — it does not wait for monitoring-advisor or emit a completion marker.

Workflow numbering note: numbers are assigned by execution order (W1 → W2 → optional W3 → optional W4 [W4.1 + W4.2] → optional W5), not by insertion order in this file. references/workflows.md is the source of truth.

Post-synthesis confirmation rules

Always end the synthesis with one clear, specific recommendation in plain English: "Given the above, I recommend: [specific action]"

If the risk is High or Medium: STOP and wait for confirmation before editing any file. You must ask the engineer and receive an explicit "yes", "go ahead", "proceed", or similar confirmation before making code changes. Say: "Do you want me to proceed with the edit?" Do NOT say: "Proceeding with the edit." — that skips the engineer's decision.

If the risk is Low: Use your judgment based on the synthesis findings. If the change is straightforward and the synthesis found no concerns, you may proceed. If anything is surprising or worth flagging, ask before editing.

Session markers

These markers coordinate between the skill and the plugin's hooks. Output each on its own line when the condition is met.

Impact check complete

After the engineer confirms (High/Medium) or after presenting the synthesis (Low), output one marker per assessed table. IMPORTANT: use only the table/model name, not the full MCON:

(Use the model filename without .sql extension — NOT "acme.analytics.orders" or "prod.public.client_hub")

How many markers to emit depends on how the assessment was triggered:

Hook-triggered (the pre-edit hook blocked an edit and instructed you to run the assessment): Be strict — only emit markers for tables whose lineage and monitor coverage were fetched directly via Monte Carlo tools in this session. If the engineer describes changes to multiple tables but only one was formally assessed, emit only one marker. The pre-edit hook will gate the other tables and prompt for their own Workflow 2 runs.

Voluntarily invoked (the engineer proactively asked for an impact assessment): Be looser — emit markers for all tables the assessment meaningfully covered, even if some were assessed via lineage context rather than direct MC tool calls. The engineer is already safety-conscious; don't force redundant assessments for tables they clearly considered.

Monitor coverage gap

When Workflow 2 finds zero custom monitors on a table's affected columns, output:

Use only the table/model name (NOT the full MCON). This allows the plugin's hooks to remind the engineer about monitor coverage at commit time. Only output this marker when the gap is specifically about the columns or logic being changed — not for general table-level monitor absence.

After the prompt is delivered, the post-edit / pre-commit hook clears the gap state internally so it won't re-prompt for the same gap; if the engineer edits the model again, Workflow 2 will re-evaluate from scratch and re-emit the marker only if a gap still exists.

Sandbox build ran (W4.1)

Emit after a successful dbt build in Workflow 4.1 (or after a deliberate skip — e.g. YAML-only diff or --skip-build — including the skip reason). One marker per model.

Validation executed (W4.2)

Emit after Workflow 4.2 finishes executing validation queries for a model, regardless of individual per-query verdicts. One marker per model.