How to Use TikTok Symphony Agent: The Prompt-to-Ad Workflow, Step by Step

TikTok announced Symphony Agent at Cannes Lions on June 22, 2026, and the official help article about it runs to roughly nine hundred words: benefits, feature names, one access path, a safety note. Nothing about what the prompt box actually rewards, how the edit loop behaves, or what the moderation gate rejects. Seven weeks later, that gap is being filled mostly by announcement digests that reword the press release.

This is the working version, assembled from TikTok's own documentation in English and Chinese, its For Business blog, and the early hands-on coverage out of Japan and China, with the single-source claims flagged as such. We build AI product video on Seedance-class engines for a living, so where engine behavior matters, we can speak from production experience rather than the press kit — while being clear we are describing Symphony Agent from the record, not from inside it.

Where it lives, and which part this guide covers

Symphony Agent is one brain across three surfaces: Symphony Creative Studio (chat-based ad creation), Content Suite (AI search over creator content), and TikTok One (brief generation and creator matching, where per trade coverage a campaign brief can be built from text, a PDF, or a product URL, and comes back with a shortlist of about twenty creators). This guide covers the Creative Studio side, the part that turns a prompt into an ad video.

Access is plain: a TikTok for Business account, desktop only. Log into Symphony Creative Studio, hover over the left navigation panel, click Agent. Creative Studio is described by TikTok's own help center as a free creation tool; TikTok has published no Agent-specific pricing, so treat "free" as inherited rather than promised.

What the prompt box actually accepts

The documented inputs: one or more product images, a TikTok trend as inspiration — a template, a top ad trend, or a custom reference video you upload — and natural-language direction. TikTok's examples extend that to selling points, target audience, brand tone, and creative constraints ("no humans", "no medical claims"). The Chinese-language help article lists the same five: 商品图片、卖点、目标受众、品牌风格、创意限制.

Two field notes worth stealing. Cross-border seller coverage out of China recommends arriving with three to five core selling points and three to five clean product images. And the Japanese hands-on walkthrough found the Agent produces meaningfully better results from product images, brand assets and reference compositions than from text-only prompts. That matches our own experience operating Seedance-class engines: these models follow concrete visual anchors far more reliably than adjectives.

What you cannot do, despite what several digests imply: paste a product URL into the video flow. URL input is documented only on the TikTok One brief side.

The four deliverables, and the edit loop

The flow is a fixed sequence, with editing available at every step:

1. Product brief — the Agent's reading of your product and goal. Correct it here; errors compound downstream.
2. Insight report — produced when your inspiration is a top ad trend or an uploaded reference video, explaining why the reference performs.
3. Storyboard — the step where conversational editing earns its keep: rewrite storylines, adjust scenes, modify voiceovers, change tone or CTA, highlight different selling points.
4. The video — request up to three variations at once for A/B testing, with different characters, scenes or creative directions.

TikTok's For Business blog quotes about three minutes to generate a 12-second video and five for a 20-second one; its help center says about five minutes for 15 seconds. The numbers disagree, so plan on "a few minutes per render" and attribute any specific figure to its page.

Beyond the prompt box there is an edit button opening a side panel for precise script, scene, voiceover, structure and CTA adjustments, and a separate AI Edit for fixing visual defects in specific scenes. Sessions save and resume, several run in parallel, and "Use the previous version" is an undo.

The moderation gate, before you waste renders

Moderation runs before generation. Per TikTok's help center: non-compliant images or text are blocked pre-generation, and human faces in product images are automatically flagged for review. The Seedance-era rollout notes go further: images or videos containing real faces are restricted as generation references, prompts referencing celebrities are rejected, and all generated content undergoes further review before it can publish on TikTok.

If your ad concept depends on a real spokesperson's face, this tool is structurally not where you make it. Feed it product-only shots and let it generate its own people.

Labels, rights, and what leaves the platform with you

The good news, from TikTok's own terms: outputs generated from the GenAI tools belong to the user unless otherwise stated, inputs and outputs are not used to train other models at this time, and exports go to your computer or sync to Ads Manager. Stock assets in the wider Studio library come licensed via partners like Getty Images and are cleared for commercial use.

The catch: an AI-generated label is added to all exported videos per TikTok ad policy, along with an invisible watermark and C2PA Content Credentials. TikTok auto-applies AI labels when it detects C2PA markers regardless of self-disclosure, and other platforms increasingly read the same credentials. No official source addresses removing the label. Assume it travels with the file.

What's underneath, and the 2.5 fine print

TikTok's help center is explicit that Creative Studio's video generation is powered by Dreamina Seedance 2.0 — the Agent-specific pages never name a model, so the safe statement is that the Agent operates inside a Studio whose engine is Seedance 2.0. On August 3, 2026, TikTok announced a Seedance 2.5 upgrade for Creative Studio and Symphony Automations: maximum AI video length doubled from 15 to 30 seconds, reference uploads raised from 9 to 50 across image, video and audio. Two caveats before you plan around that: the upgrade is documented for Studio and Automations, not explicitly for the Agent workflow, and it is rolling out to select paid advertisers in select markets.

We run Seedance 2.0 and 2.5 directly in production (our cost notes on the 2.5 tier cover what the 30-second single take is actually worth), so one calibration from experience: Seedance-class models reward specific, timestamped direction and concrete product handling, and they punish vague mood words. The Japanese coverage's advice to human-review product shape, logos and dialogue accuracy in every output is correct, and it applies to every engine in this class, ours included.

Where the Agent stops being the right tool

Symphony Agent starts from a brief and a trend, inside TikTok's walls, with TikTok's labels on the output and its restrictions on real faces. That is a genuinely useful free drafting surface for TikTok ads.

It is the wrong shape for three jobs. When you want to rebuild a specific winning video — not a trend category, but that one video whose formula you can see working — you need a tool that deconstructs the reference and rebuilds its structure around your product, which is exactly what Riffkit does. When you need your own locked character with a consistent face across a campaign, the real-face restriction rules the Studio out. And when the deliverable has to live beyond TikTok as an asset you fully control, you want output whose engine, ratio and rate you chose yourself — the pricing page shows ours before you render. Same model family underneath; the difference is whose walls it ships inside.