Helena - AI Marketing Agent

View saved version →

---
name: helena-marketing-references
description: Find and inspect bounded Marketing Intelligence reference previews with Helena, then apply relevant patterns to original marketing work.
---

# Use marketing references

Call `search_marketing_intelligence` with a repository, focused query, and up to five results. Supported repositories cover email templates, competitor ads, scaling ads, brand identity, tactical case studies, campaigns, and growth playbooks. Select a relevant result and call `get_marketing_intelligence_preview` with its repository and ID.

Use sanitized text and any low-resolution visual as inspiration. Identify the relevant hook, structure, or audience insight and create original content grounded in the user's brand and verified product facts. Do not copy another brand's claims or imply a preview grants rights to its original assets.

Search and preview tools return results directly. Respect quota and access errors. No tool exports raw source HTML, original media, storage paths, private adaptation instructions, or the corpus. Do not attempt to retrieve those through workspace tools or invented operations.

View saved version →

---
name: helena-publishing
description: Publish, schedule, or save CMS drafts with Helena's fixed platform variants; use separate email sending and advertising tools only with explicit approval, never for review or reporting requests.
---

# Publish approved content

Confirm the exact content, destination account/site/audience, and publication time. Show drafts in chat unless the user asks to save one to a connected CMS or email platform. Generating media, reviewing copy, or saying “looks good” is not permission to publish, send, schedule, or change ad spend.

Use `get_connected_integrations` to check connections and access modes. Use `get_memories` when saved Helena facts or preferences are relevant; these are not a complete brand profile or authorization to publish. Discover identities and posting targets with `get_social_destinations` or `get_cms_destinations`; never guess IDs. If an integration is absent or read-only, explain that and stop.

Call `publish_social_post` or `publish_cms_post` with the declared `platform` and its platform-specific fields inside `input`. Omit `input.scheduled_date` for immediate publication, or supply a future timezone-qualified timestamp to schedule it. Threads currently supports immediate publication only. For an explicitly requested unpublished CMS draft, set `input.draft: true` and omit the date; omitting `draft` means publication. To publish an existing CMS draft, provide its `post_id` instead of new-post content: preserve the existing ID, never recreate it. Existing drafts support immediate publication only. Use `update_cms_post` for revisions, preserving publication state; Webflow stages changes and requires separate approval of the returned version hash before publication. TikTok inbox uploads use the separate `create_tiktok_draft` tool.

CMS platforms include WordPress, Ghost, Webflow, Shopify, Blogger, and HubSpot. Blogger requires `blog_id`. For HubSpot, use `get_cms_destinations` with `include_authors: true` when an author is needed; immediate publication of a new post requires a slug, author, and meta description. HubSpot scheduling is not exposed. If creation succeeds but publication fails, reuse the returned `created_post_id` instead of creating a duplicate.

For email, use the separate email family with `platform: mailchimp`, `klaviyo`, or `brevo` and its declared `input`. Read audiences/templates, create an unsent campaign with `create_email_campaign`, and revise it with `update_email_campaign`. `update_email_template` changes a reusable template; Klaviyo campaign copies must then be refreshed by passing that template ID to `update_email_campaign`. `preview_email_template` saves a public preview artifact for Mailchimp/Klaviyo; Brevo preview and cancellation are not exposed. Brevo audience discovery also returns verified senders; use its closed `updates` object for draft revisions, never a raw API payload. Inspect audience, sender, subject, and actual campaign content before `send_email_campaign`, with optional `input.scheduled_date`. It sends to configured recipients, not an arbitrary person. SMS/push are unsupported. Cancel a scheduled send in the supported tool or provider dashboard before editing; only editable drafts can be revised. Set `user_confirmed` only after explicit approval of the exact action; budget and activation changes also require approval.

Social/CMS/email publishing tools return results directly. A failed or timed-out request is not proof that nothing was sent or published; verify in the connected platform before retrying to avoid duplicates. Scheduling confirms the schedule was created, not that future publication has happened. Use `get_content_calendar`, `reschedule_content`, and `cancel_scheduled_content` for scheduled social/CMS items. Report actual results, links, and errors.

For paid ads, read `get_ad_accounts`, `get_ad_campaigns`, and `get_ad_performance` first. Each create/update tool targets one declared entity with fixed platform inputs; no generic API payload is accepted. Confirm account, currency, exact changes, budget, targeting, and launch. Google values use micros; Meta budgets use currency subunits; TikTok uses currency units; LinkedIn requires amount/currencyCode. Never assume USD. Campaign creation requires `input.initial_delivery`: `paused` for Meta/Google, `draft` for LinkedIn, or `live` for TikTok, which can start spending immediately. Google creation also selects `campaign_type`; TikTok asset tools select `media_type`. Adding a Google search ad group to an active campaign can start delivery. Google copy edits can replace/remove the old ad; confirm replacement, then use the returned ID for any separate status update. When an ads tool returns a job_id, poll get_operation_status; never repeat the mutation to check progress. After failures/timeouts/expired jobs, verify provider state before retrying: multi-step changes can partially succeed. Do not claim delivery or provider approval merely from ACTIVE/ENABLED status.

Refunds, return labels, arbitrary shell commands, arbitrary integration APIs, and unrestricted agent requests are unsupported. Do not call Helena for those requests or repurpose an email campaign tool to email a return label.

View saved version →

---
name: helena-video
description: Generate footage or author HTML compositions in the workspace and render them with Helena. Advanced routing and orchestration happen in the Helena web app.
---

# Create or revise a video

Capture verified product references, exact copy/dialogue, aspect ratio, duration, and audio requirements. Use `generate_video` for new footage; put supplied speech verbatim in `script`. Always pass explicit `duration` and `aspect_ratio` matching the user's request (landscape is `16:9`, portrait is `9:16`). Supported clip durations are 4, 5, 8, or 12 seconds, including reference-guided clips with up to nine images. A closing-frame image requires 5, 8, or 12 seconds. For a longer video, propose multiple clips and compose them together; do not silently shorten the user's request. Do not combine non-temporal reference images with first/last-frame anchors.

For deterministic editing, stitching, or exact text/branding, author a self-contained HyperFrames HTML composition. Save it with `write_workspace_file` under an account-relative path such as `files/video-compositions/product.html`, obtaining approval for the file write. Then call `compose_video` with `composition_path` and optional render settings. The renderer does not write HTML or accept a brief/assets payload. To revise, read the saved HTML, edit it, save to the same path, and render again.

Minimal format (replace the example copy and styling with the requested content):

```html
<!doctype html>
<html><head>
<script src="https://cdn.jsdelivr.net/npm/gsap@3.14.2/dist/gsap.min.js"></script>
<style>
  body { margin: 0; }
  #root { position: relative; width: 1080px; height: 1920px; overflow: hidden; background: #111; color: white; }
  .clip { position: absolute; inset: 0; display: grid; place-items: center; font: 80px sans-serif; }
</style>
</head><body>
<div id="root" data-composition-id="main" data-width="1080" data-height="1920" data-duration="5">
  <section class="clip" data-start="0" data-duration="5" data-track-index="1"><h1 id="title">Your headline</h1></section>
</div>
<script>
  const timeline = gsap.timeline({ paused: true });
  timeline.from("#title", { opacity: 0, y: 40, duration: 0.6 }, 0);
  window.__timelines = { main: timeline };
</script>
</body></html>
```

Use unitless pixel dimensions and timing in seconds, not milliseconds. Timed layers need `data-start`, `data-duration`, and numeric `data-track-index`; overlapping layers use different tracks. Media elements must be direct children of the composition root. Workspace media paths can be used in `src`; the renderer makes them fetchable. `data-media-start` selects a source offset. Video elements are muted: preserve sound using a separate root-level `<audio>` with matching source/timing, its own track, and `data-volume`. Omit audio for silent output. Do not guess source durations or speech timing; inspect supplied media with `analyze_media` and use `transcribe_media` for spoken captions. Register a paused GSAP timeline keyed by the root composition ID.

Proper routing across generation, editing, composition, and providers happens within the [Helena web app](https://agent.enrichlabs.ai). Use it for advanced or hybrid workflows instead of reconstructing internal creative playbooks in ChatGPT.

Footage generation consumes video credits; composition rendering does not. Both return a `job_id`; poll `get_operation_status` until done or error without repeating the operation. Inspect the returned URL, warnings, and any quality critique before claiming success. Fix and rerender a flagged composition rather than publishing it. Creating media does not authorize social publication; publish only through a separate tool after explicit user approval.