Nitrosend

Nitrosend

Use this as the single public Runx surface for Nitrosend customer operations. It calls the live Nitrosend MCP boundary through a bounded provider adapter; API keys are delivered as credentials, never accepted as skill inputs or returned in receipts.

This skill is not for Nitrosend customer-support administration or team Slack work. Those are product-operator concerns owned by the Nitrosend repository.

Choose the runner

• status (default): live account, brand, sender, domain, provider, warmup, and deliverability readiness.
• analytics: live account, campaign, flow, or message insights.
• review-delivery: read-only content and preflight review.
• plan-campaign, plan-flow, plan-transactional, and plan-import: bounded agent judgment that produces a reviewable request without provider completion.
• compose-email: read the live Nitrosend authoring contract, let one bounded agent author its exact next call, and return authoritative MCP validation. It never persists a draft or gains delivery authority.
• apply-draft: apply exact reviewed arguments for a campaign, flow, template, or segment draft. It never sends or activates.
• approve-delivery: approve a reviewed campaign or flow without delivering.
• send-campaign: send or schedule an already-approved campaign after a fresh provider review and explicit approval.
• activate-flow: activate an already-approved flow after a fresh review and explicit approval.
• send-transactional: dry-run or send one idempotent message to one recipient.
• import-contacts: dry-run or import at most 100 inline consented records.
• import-contacts-csv: validate or upload a local CSV through Nitrosend's authorized direct-upload path. File bytes and signed URLs never enter the agent packet or receipt.
• import-status: make one bounded status read for an asynchronous import.
• segment-from-prose: internal planning lane for the current supported filter catalog; unsupported filters are rejected rather than approximated.

Use the current public https://nitrosend.com/SKILL.md, nitro_get_status, and the live MCP schema as product truth. Do not copy onboarding or tool schemas into another repo-local skill.

Safe operating sequence

1. Run status and stop on sender, domain, suspension, warmup, or account blockers.
2. For email authoring, use compose-email so Nitrosend supplies current brand and memory context before the model writes. Treat its MCP validation as authoritative; repair in another bounded turn when requested. For other planning, use the matching planning runner.
3. Apply only validated arguments through apply-draft; that separate runner retains the approval gate and is the first persistence boundary.
4. Run review-delivery before approval. Use approve-delivery separately so retries never combine approval-state mutation with recipient delivery.
5. Use send-campaign or activate-flow only after provider approval state is established. A fresh review and Runx approval gate are mandatory.
6. Give every real transactional send, campaign delivery, and import a stable idempotency key. Reuse that key after a timeout; do not mint a new one.
7. Treat completion as real only when the sealed receipt contains Nitrosend provider evidence. A plan receipt is not proof of send, schedule, activation, or import.

Contact import rules

Every import requires a stable source_id and a plain-language consent_basis. Purchased, scraped, or data-broker lists are refused.

For CSV imports, pass an absolute .csv path. The adapter computes metadata and checksum locally, reserves an authorized upload, streams the file directly to the returned public HTTPS host, finalizes with the signed ID, and discards the signed URL. The import is asynchronous; call import-status again as needed rather than keeping a resident polling loop.

Stop conditions

• Missing provider credential or brand context.
• Unsupported operation, audience, segment filter, or lifecycle transition.
• Missing consent source, recipient, schedule time, or idempotency key.
• Failed provider review or preflight.
• Missing or denied approval.
• Any request to expose credentials, signed upload URLs, raw contact files, or unbounded provider responses.
• Any claim of completion without provider readback evidence.

Agent task contracts

The planning acts prepare exact downstream operations and never call the provider. compose-email-from-contract is different: the enclosing graph has already read authoritative MCP intent evidence, and the agent only authors a candidate for the graph's read-only MCP validation. No agent act persists or delivers.

send-campaign

Build one campaign plan from the objective, current account-status JSON, and audience brief. Distinguish a campaign from a flow or transactional message. Require an explicit bounded audience, compliant sender readiness, review before approval, and a separate confirmation before send or schedule. Do not invent list or segment ids and do not claim delivery from this planning runner.

build-flow-plan

Build one event-triggered automation plan with a supported trigger and ordered steps. Route one-off broadcasts to send-campaign; never disguise them as a flow. Preview and creation may be planned before confirmation, but approval and activation must remain a separate confirmed delivery-control operation.

send-transactional-plan

Plan exactly one email or SMS to one named recipient. Require channel-specific content, a stable idempotency key, dry-run validation, and confirmation before a real send. Reject audience or list broadcasts and route those to send-campaign.

compose-email-from-contract

Read the composition_intent provider evidence supplied as the declared input or current graph context. It must contain a successful Nitrosend result with a composition contract, contract id, and next call. Return campaign_candidate with decision (ready, needs_input, or reject), exact arguments, rationale, and blockers.

For ready, begin from the contract's next_call, preserve its contract_id, set composition_mode to validate, and fill only the requested creative fields. Follow the supplied brand, memory, examples, hard constraints, design mode, and repair guidance. Never reconstruct omitted context, invent claims, add an audience or schedule, or request draft, approval, activation, testing, or sending. If the intent evidence or required contract fields are missing, return needs_input without fabricated arguments.

import-contacts-plan

Plan only consented contacts with a stable source, consent basis, bounded record count, channels, compliance checks, dry run, and confirmation before import. Refuse purchased, scraped, or brokered lists. Choose inline import only for the bounded inline path; use the CSV runner for larger local files, without copying file contents or signed URLs into the plan.

segment-from-prose-plan

Translate the brief only through filters and predicates present in filter_catalog_json. Return a concrete segment_request when every condition is representable. Reject unsupported event history, attribution, or compound semantics rather than approximating them with superficially similar fields.