מסמך מקור מתקדם. דפי ההתקנה, פרוטוקול ההפעלה וששת הסקילים זמינים לקריאה בעברית; קובץ המקור נשמר כפי שהוא.

--- name: orchestration description: Use when coordinating a Hebrew Meta lead-generation pipeline, checking marketing quality, or producing an evidence-limited report. ---

Coordinate the bounded marketing workflow

Follow brief → messages → creative → campaign → lead → status; report is a derived artifact from explicit source snapshots. Preserve the three layers: canonical knowledge, tenant profile, and separately protected executor. A skill file is knowledge, not a running agent or native host discovery.

For each new customer rollout, connector change, or marketing claim about the product, first consult docs/operator-onboarding-v1.md at the package root and the dated capability audit it links. Follow its installation, per-client isolation, approval, evidence and revocation gates. Do not describe conditional or untested integrations as working.

Version and authority boundary

Read the selected profile and the exact contract from the package root. Unconverted V1 uses contracts/pipeline.contract.json, mode simulation, demo IDs, supplied synthetic records, geometric placeholders and the existing offline validator only. V1 restrictions remain intact. For V2 explicitly select contracts/pipeline.v2.contract.json; reject unconverted V1 with an explanation. Conversion creates a separate simulation artifact, preserves its source, enriches missing fields, rebuilds all digests, and never promotes approve_demo to human authorization. Neither mode names nor schema validation enable a connector. This build uses no network, accounts, secrets, installs or customer records; integration-test and customer-pilot are separate future gates, not claims of readiness.

Use approved API/MCP/connectors first when integration is separately enabled; browser fallback is only for a capability unavailable in the authorized programmatic path, at the same authority boundary. Never launch external scripts, installers, model CLIs or services from this Skill. Source rule IDs below resolve to docs/knowledge-register-v1.md (checked 2026-10-24). Untested real behavior remains unverified.

For each client, discover the agent host's already authorized Meta and CRM connections before recommending new setup. Verify the exact account/file, operations, scopes and readback supported by each. Reuse a suitable connection within its authority; otherwise describe the smallest client-owned setup and approval required. A supervised connector handoff is an explicit action with result verification, not an autonomous integration or proof that leads arrive while chat is closed. Do not require a new vendor-owned Cloud project for customer onboarding.

Handoff and review [A1, A2, N1, N2, R1, M1]

Brief: supplied offer/audience, evidence classes, contradictions and unknowns. Block only the dependent stage; missing budget need not stop bounded copy.
Messages: exactly two materially different reasons to care about the same supported offer. Record one comparison variable, expected outcome, metric and stop condition; no forced sample size or benchmark.
Creative: distinct post/story copy and four final assets, one per angle per 1:1 and 9:16 ratio. A test placeholder cannot satisfy final asset acceptance.
Campaign: new test entities PAUSED at campaign/adset/ad, preflight and exact protected approval. Null budget blocks execution. Existing active assets are outside this pipeline.
Lead/status: private synthetic Sheets first, tenant/source identity, independent quality assessment and observed handling status. Real sale requires independent evidence.

Route cardinality is a user decision. If the brief specifies both website and native Meta form, maintain two distinct campaign/source artifacts and route-specific approvals through the full pipeline, sharing only the explicitly common lead/status contract. Never turn an explicit two-route request into an either/or clarification. A route blocked by missing evidence is reported as blocked for that route; independent drafting for the other continues. Preserve the user's already supplied common answers without asking again.

Use the contract envelope fields, fresh revision, and hash of the entire immediately preceding envelope. A changed stage invalidates downstream digests and approvals. Reject unknown tenants, changed identity mappings and contradictions. A separate human reviewer assesses audience/offer fit, substantive angle distinction, support for implications, promise→CTA→form consistency and asset usability. A passing schema cannot replace that review.

Report [N2, R1, M1]

Record as-of, snapshot ID, lead-created cohort, timezone, attribution window, refreshed source window and qualified-definition version. Refresh the agreed source window when reporting, within approved quota. Exclude synthetic tests from business metrics; leave missing ad sources unknown. Separate Meta leads, CRM leads, intake gap, reviewed/qualified/unknown, handling status and independently verified sale/revenue. Do not sum platform attribution as sales or treat UTM as conversion proof; use no PII in UTM.

Calculate CPL only with comparable spend and a labeled lead denominator; quality among reviewed = yes/reviewed, review coverage = reviewed/all, and verified quality share = yes/all (verified so far). CPQL = comparable spend/yes; zero or missing denominator yields null, never zero or infinity. Mark immature cohorts or partial review provisional. Compare only equivalent windows/quality versions; a breakdown average is observation, not causal evidence or a budget instruction. No autonomous optimization. Report artifact paths, commands and observed evidence class: design / local simulation / real integration / customer environment. No claim extends to Google Ads, newsletter or video.