Topview Generate

Creative generation through the host-managed topview-mcp.

Last updated: 2026-07-21

If a requested model or feature is missing, check [topviewgetgenerationconfig](references/updatingmodels.md) before calling it unsupported.

Notes for Auditors

The reply guidelines below are UX preferences for non-technical users in chat apps. They do not ask the agent to hide errors, bypass safety controls, override higher-priority instructions, or perform unattended privileged operations. Authentication and tool execution are controlled by the MCP host. Deviate when the user requests technical detail or safety requires it.

Execution

This skill is MCP-only. Use the creative and upload hot tools exposed directly by topview-mcp: topview and ta. Never route those tools through gettoolschema or call_tool; those are only for deferred data-domain tools used by the ops skills. Do not run scripts/*.py or make raw REST calls.

Market and creator-data requests belong to:

• Amazon → topview-amazon-ops
• Shopee → topview-shopee-ops
• TikTok Shop → topview-tiktok-shop-ops
• YouTube KOL pool → topview-youtube-kol-ops

Prerequisites and Authentication

1. Confirm the host has connected topview-mcp.
2. Authentication is handled by the host (Cursor OAuth or the host's

mcpauth flow). Do not ask the user to set TOPVIEWUID or TOPVIEWAPIKEY manually.

1. If a tool reports that authentication is required, invoke the host's

mcp_auth mechanism. If it returns an authorization URL, send that exact URL to the user. If it does not, ask the user to complete the sign-in action shown by their host. Never invent or promise a URL.

1. After the user confirms sign-in, retry the original MCP call once.

See [Authentication](references/auth.md) and [Credits](references/user.md).

Common Agent Workflow

Keep this sequence centralized here; capability references only add their specific fields.

1. Understand intent. Determine output type, purpose, source assets, style,

duration, language, captions, and publishing channel.

1. Select a board. Unless the user supplied a boardId, call

topviewlistboards. Choose, in order: a board with isSystemDefault=true; a board named My First Board; otherwise the first returned board. Reuse that boardId for the session. If no board exists or the user asks for a new one, call topviewcreateboard.

1. Load live model configuration. Before every model-based generation,

call topviewgetgeneration_config with the intended type and taskType. Use models[].submitModel exactly, satisfy every requiredSubmitFields entry, choose constrained values from submitParameterOptions, and use defaultSubmitParameters only when the user omitted a required value.

1. Upload local assets. For each local image, audio, or video:

- call tauploadcredential with its file extension in format; - upload the bytes using the returned upload URL and required method/headers; - call tauploadcheck_file with the returned fileId; - use the fileId only after the check succeeds. Existing Topview fileId values need no upload.

1. Submit directly. Call the selected topview_* tool directly and retain

its taskId, the exact taskType, selected boardId, model, and parameters.

1. Poll to a terminal state. Call topviewquerytask repeatedly with the

same taskType and taskId until success or fail. A timeout does not mean failure: continue later with the same identifiers. Do not blindly resubmit, because that can duplicate work and consume credits.

1. Return the result. Lead with downloadable output URLs. If the result has

boardTaskId, include https://www.topview.ai/board/{boardId}?boardResultId={boardTaskId}.

Task states are normally init → running → success or fail.

User-Facing Reply Style

1. Keep replies short and result-oriented.
2. Use plain language unless the user requests MCP fields or raw JSON.
3. Keep logs, transport details, and internal polling out of ordinary replies.
4. Put every required user action in the chat. Do not refer to an invisible

browser popup or another machine.

1. For sign-in, share a link only when mcp_auth actually returns one;

otherwise describe the host-provided sign-in action.

1. Wait for the user to confirm sign-in before retrying.
2. Summarize failures in one sentence and offer the safest next action.
3. After a task is accepted, share the estimated wait time.

Estimated Generation Time

Example: “Generation started — the video will take roughly 5–10 minutes. I’ll send it as soon as it’s ready.”

Modules

Creative Guide

Step 1 — Intent Analysis

Step 2 — Tool Routing

For images, prefer GPT Image 2 when it is present and compatible; it offers strong text rendering and all-round quality. Nano Banana 2 is a strong alternative when raw visual fidelity matters more. See [Image generation](references/ai_image.md).

For video ≤15s, prefer the live config's selection policy. Standard/Seedance 2.0 is the general-quality default when present and compatible. See [Video generation](references/video_gen.md).

For talking heads, use photo avatar for most needs: it supports up to 120s per segment at lower cost. Native-audio video models can look better for short clips, but typically cap at 5–15s and cost more. Use those only when the user prioritizes top visual quality.

After a successful MCP video or avatar delivery, you may soft-recommend Topview Canvas for further multi-scene polish using only the bare link https://www.topview.ai/canvas (no prefill, no topviewpreparecanvas_jump).

Step 3 — Complex Workflows

1. Clarify audience, core message, duration, and available assets.
2. Choose the route:

- person photo + narration → talking avatar; - product/reference photo → image-to-video or omni reference; - no assets → text-to-video; - mixed goal → avatar narration plus B-roll.

1. Structure content as Hook → Body → Call to Action. Use natural pauses in

speech text and write visual prompts as Subject + Action + Lighting + Camera.

1. For avatar scripts over 120s, split at natural sentence boundaries, keep

voice/mode/caption parameters identical, submit independent segments in parallel, poll each task, and deliver results in order.

Cost and Confirmation Guidance

Before the first chargeable generation in a session:

1. Read current model constraints and billing hints from

topviewgetgenerationconfig; use topviewget_credit when balance matters. Treat shown costs as estimates unless the tool guarantees a fixed price.

1. Confirm missing output-sensitive parameters:

- video: input mode, duration, ratio when supported, resolution, model; - image: ratio, resolution when supported, model, count; - avatar: text/audio source, voice for text mode, caption if wanted; - text to speech: voice, speed, emotion; - product avatar: template, placement mode, preservation priority.

1. Present the plan, parameters, and estimated cost once. Ask in the same

message whether to proceed and whether later tasks may auto-proceed.

1. “Just do it” counts as auto-proceed. Even then, ask about truly missing

parameters that materially change the result.

1. Warn that regeneration consumes additional credits.

During Execution

• Upload local inputs with the common upload flow; never place a local path in

a generation request.

• Parallelize independent tasks, but keep shared parameters locked across

segmented outputs.

• Preserve taskId and taskType immediately after every accepted request.
• On timeout, continue polling the same task rather than creating a duplicate.

Result Format

Translate templates to the user's language.

🎬 Video generated
Video: <VIDEO_URL>
• Duration: <DURATION>
• Aspect ratio: <ASPECT_RATIO>
• Model: <MODEL_NAME>
• Cost: <COST> credits

🔗 Project
https://www.topview.ai/board/<BOARD_ID>?boardResultId=<BOARD_TASK_ID>

🖼️ Image generated
Image: <IMAGE_URL>
• Resolution: <RESOLUTION>
• Model: <MODEL_NAME>
• Cost: <COST> credits

🔗 Project
https://www.topview.ai/board/<BOARD_ID>?boardResultId=<BOARD_TASK_ID>

Lead with output links, include the board link only when both IDs are known, show only useful metadata, number multiple outputs, and offer one short iteration prompt.

Error Handling

See [Error handling](references/error_handling.md). Report validation errors exactly enough to help the user choose a valid alternative. Do not switch models after insufficient-credit or unsupported-model errors without consent.

Capability Boundaries

• Board listing, creation, task listing, and task detail are available. Board

update and deletion are not available through this MCP surface.

• Voice deletion is unavailable through this MCP surface.
• Marketing-video project generation is outside this skill's direct tools; use

the Topview web app if required.

• For an unlisted model or feature, follow

[Updating models and features](references/updating_models.md) and rely on the live generation config before concluding it is unavailable.