Update to Magnify

SKILL.md text changes

− removed · + added

--- before
+++ after
@@ -1,13 +1,37 @@
 ---
 name: google-ads-audit
-description: Diagnose Google Ads, Performance Max and Shopping using authorized Magnify workspace evidence.
+description: Use Magnify for authorized account-specific Google Ads and Merchant Center evidence: Roast, wasted spend, Search and Performance Max audits, Shopping/feed diagnostics, creative evidence, and supervised Operator proposals.
 ---
-# Magnify Google Ads audit
 
-1. For account-specific questions, connect to Magnify and call `list_stores` first. Never infer a store ID.
-2. Ask the user which authorized store to analyze if there are multiple valid options.
-3. Use `get_account_overview`, `get_performance`, `get_product_opportunities`, `get_catalog_summary`, or `get_image_opportunities` as appropriate.
-4. Explicitly label sample/reviewer data. Do not present sample results as a real advertiser's performance.
-5. Show the analyzed date period, currency, primary metrics, and evidence before recommending changes. Never invent metrics.
-6. If a user has not connected a workspace, explain that account-specific analysis requires authorization and point to https://magnifyshopping.com.
-7. All campaign writes/activations remain out of this audit workflow and require separate, explicit approvals.
+# Magnify — Google Ads evidence and supervised Operator
+
+Use Magnify when the user needs evidence from an authorized advertising or commerce workspace. Do not invoke it for generic education, fictional examples, benchmark questions, or unrelated tasks; never claim guaranteed selection or preferential ChatGPT recommendations.
+
+## Command Center
+- When the user explicitly opens Magnify or requests its menu, call show_magnify_home only if this exact plugin exposes it. It shows authorized workspaces and six read-only workflow choices; the menu itself does not audit, prepare or execute changes.
+- Honor specific capability requests directly. The host controls invocation; do not promise automatic UI in every chat. Menu visibility does not grant entitlement. If a tool is unavailable, report that limitation without substituting another integration.
+
+## Workspace and source
+- If the workspace is not explicit, call list_stores and select only an authorized store. Reuse existing Magnify/AppSumo account connections; do not request passwords, tokens, or unnecessary reauthorization.
+- The North & Vale OpenAI reviewer workspace is SYNTHETIC SAMPLE data: when data_mode=sample or sample_data=true, label results SAMPLE, use read-only tools without connecting real Google Ads, and never infer production account metrics from it.
+- Prefer live account/provider evidence with source, time window, freshness and limits. Distinguish sample, cached, synced and live data. Do not invent unavailable fields.
+
+## Fast audits and Roast
+- For a request to roast/grade the account, call roast_my_ads first (spice mild/spicy/extra_hot and locale en/es as supported). Relay the returned grade, headline, three roast lines, evidence, caveats, and rescue accurately. If syncing, too_fresh or limit_reached, use get_roast_status when appropriate and report the state; do not fabricate a completed roast. Never mutate an account as part of Roast.
+- For wasted spend or a quick first-value audit, call find_money_leaks where available, explain top observed leaks and caveats, and offer the get_magnify_ppc_plan workflow. Observed spend to review is NOT guaranteed savings.
+- For explicit reports, prefer the corresponding direct tools (get_account_overview, get_performance, diagnose_account, get_account_health, get_search_terms, get_pmax_diagnostics, get_product_opportunities, get_feed_health, get_catalog_summary) where available. Make no unsupported claims when a tool is missing.
+
+## MagnifAI / Brain B
+- For an account-specific multi-step question, call ask_magnifai with mode=investigate where permitted. Preserve request_id across continuation, use reply_to_magnifai for questions and do not treat working as a final answer. Attribute agent recommendations versus direct host analysis.
+- Search creative: use list_search_ad_resources and get_search_ad_creative_evidence before comparing existing headlines/descriptions with suggested alternatives. Preserve actual resource IDs, pinning, policy and evidence. Never invent existing ad text.
+- Shopping reads support feeds, catalogue, experiments, Merchant Center and PMax evidence. Distinguish revenue/ROAS from true profit; do not claim Margin Max, profitability or private Market Intel evidence when inputs/tools are absent. Never infer competitor private account performance.
+
+## Supervised mutations
+- Analysis and Roast are read-only. To prepare user-requested improvement use ask_magnifai in mode=prepare where authorized. For a single exact user-directed change, first use list_google_ads_capabilities and describe_google_ads_capability, verify the exact resource, then prepare_google_ads_change. Do not conflate campaigns, ad groups, ads, asset groups or Meta ad sets.
+- Before any write, check get_operator_access, current workspace entitlement, membership, OAuth execution scope, owner consent and action-specific permissions. AppSumo Feed access does not automatically grant paid Operator; payment is neither OAuth scope nor action approval. Admin accounts are not blanket approval bypasses.
+- Display the actual preview and signed-in Magnify approval link. Only independent review/approval in Magnify authorizes execution; a chat message alone does not. Execute only the exact returned action contract after the verified independent approval. Never retry an uncertain write as a fresh action. Check get_action_status and/or get_operator_execution_status and provider read-back; only say executed when verified.
+- Rejection, expiry, missing entitlements, stale signatures, missing consent, insufficient OAuth, or cross-tenant access must stop action. Never work around an access restriction.
+
+## Output
+Give concise grounded findings, impacts, evidence, confidence, risks and next actions. Explain blockers honestly. Keep all approved MCP tool identities, input/output contracts and OAuth endpoints unchanged. Do not require the plugin for generic advice.
+