Article
Plan Image & Video Prompts with Claude Fable 5
Claude Fable 5 is Anthropic's first generally available Mythos-class model — a tier above Opus, built for demanding reasoning, vision, and long-running agentic work. It does not generate images or video pixels. What it does exceptionally well is read briefs, analyze reference visuals, and plan structured prompts that dedicated generation models can execute.
Recommended workflow: use Claude Fable 5 in the Claude app or API (model ID:
claude-fable-5) to plan your prompts, then paste them into HiArt and generate with GPT Image 2 or Seedance 2.0. This guide maps Fable 5's verified strengths — vision SOTA, 1M-token context, long-horizon agents, and document reasoning — to a practical prompt-planning workflow for brand, e-commerce, and video teams.Why Fable 5 for planning, HiArt for execution
Layer | Claude Fable 5 | HiArt |
|---|---|---|
Core job | Reason over briefs, screenshots, and references; output prompt specs | Generate and iterate images/video on a shared canvas |
Vision | SOTA — rebuild apps from screenshots, read charts and layouts | Reference images feed into GPT Image 2 / Seedance @ syntax |
Context | 1M tokens default; persistent memory across a campaign | Project canvas, nodes, and generation history |
Output | Structured prompts, shot lists, @ role assignments | Pixels — hero shots, product loops, social crops |
Best models | API: claude-fable-5 |
Anthropic reports that the longer and more complex the task, the larger Fable 5's lead over prior Claude models. Campaign planning — multiple assets, consistent look, platform-specific crops — is exactly that kind of task.
Fable 5 strengths that matter for prompt planning
1. Vision SOTA
Anthropic positions Fable 5 as state-of-the-art for vision: it can rebuild a web app's source code from screenshots alone, extract precise numbers from scientific figures, and complete vision-only agentic tasks that earlier models needed heavy scaffolding for. On GDP.pdf, a benchmark for reasoning over rendered documents without tools, Fable 5 leads the field at 29.8%.
For creative teams, the practical version is simpler: upload competitor ads, mood boards, product photos, or storyboard frames — Fable reads composition, lighting, typography, and color, then writes prompts your generation model can follow.
2. 1M context + memory
Fable 5 supports a 1M-token context window by default and up to 128k output tokens per request. Combined with the memory tool, it can hold an entire brand kit, shot list, and prior prompt revisions in one session — and keep character look, palette, and camera language consistent across ten or more assets.
3. Long-horizon agentic reasoning
Fable 5 is built for multi-step work that runs longer than a single chat turn. Early customers report fewer tool calls and fewer correction rounds on complex jobs. For prompt planning, that means: brief → shot breakdown → per-asset prompts → acceptance criteria — in one coherent pass instead of five fragmented chats.
4. Document and chart reasoning
On knowledge-work benchmarks, Fable 5 shows strong gains in document reasoning and chart interpretation. Paste a client PDF, Excel shot list, or media plan — Fable can extract deliverables, aspect ratios, and copy requirements, then map them to model-specific prompt fields.
5. Spatial reasoning
Spatial reasoning scores roughly tripled versus Opus 4.8 in Anthropic's evaluations. That translates to better composition specs — where the product sits in frame, how negative space balances headline copy, which camera angle fits a 9:16 crop.
The two-step workflow
- Step 1 — Plan in Claude Fable 5: Upload brief + references. Ask for structured prompts, shot lists, and
@role maps. Review and edit the plan before spending generation credits. - Step 2 — Execute on HiArt: Paste prompts into image or video nodes. Attach reference files Fable flagged. Iterate on the canvas; use Seedance 2.0 Fast for motion previews before final renders.
A customer quote Anthropic published captures why this matters: apps that took a hundred prompts a year ago now get one-shotted. The win is not magic generation — it is getting the prompt right before you open HiArt.
Copy-paste system prompts for Fable 5
Image prompt planner (GPT Image 2)
Paste this into Claude Fable 5, then attach your mood boards, product photos, or competitor references.
You are a visual creative director planning prompts for HiArt's GPT Image 2 model. I will upload reference images (competitor ads, mood boards, product shots, brand guidelines). For each deliverable I need: 1. Analyze references: subject, materials, lighting, composition, color palette, typography style 2. Output a paste-ready GPT Image 2 prompt using this structure: [Subject] + [Action/Pose] + [Environment] + [Lighting] + [Camera/Composition] + [Style/Medium] + [Text overlay if any] 3. List which elements should be uploaded as reference images (GPT Image 2 supports up to 16) 4. Specify aspect ratio and resolution tier (1K–4K) 5. Keep visual consistency across all assets in this campaign Do not generate images. Output prompts and planning notes only. Ask clarifying questions if the brief is ambiguous.
Video prompt planner (Seedance 2.0)
Use this when planning omni-reference video prompts with the
@ syntax.You are a short-form video director planning prompts for HiArt's Seedance 2.0 model. I will provide: campaign brief, reference images/videos/audio, duration, and aspect ratio. For each clip (4–15 seconds): 1. Build a shot list with camera movement and action beats 2. Assign @Image/@Video/@Audio roles for each reference file 3. Write the final omni-reference prompt with explicit @ syntax 4. Respect Seedance limits: ≤12 files total, ≤9 images, ≤3 videos, ≤3 audio; no realistic human faces in uploads; audio cannot be used alone 5. Recommend preview order: Seedance 2.0 Fast first, then Seedance 2.0 for finals Output: shot table + paste-ready prompts per clip. No video generation.
End-to-end example: e-commerce summer campaign
Input to Fable 5
- PDF brief: 1 hero banner (16:9), 3 product shots (1:1), 2 vertical ad clips (9:16, 6s each)
- Upload: brand color swatch, 2 competitor ad screenshots, 1 product flat-lay photo
- Attach HiArt model constraints (reference limits, aspect ratios)
Fable 5 output (planning layer)
- Hero banner prompt — subject, beach environment, golden-hour lighting, 16:9, headline placement zone
- Product shot prompts (×3) — same palette, alternating angles; which photo to attach as reference
- Video clip A —
@Image1product look,@Video1camera push-in reference, 6s, 9:16 - Video clip B — text-to-video fallback if no motion reference; beat-matched caption safe zones
- Consistency sheet — hex codes, font mood, photography style keywords shared across all prompts
Execute on HiArt
- Open HiArt canvas → image node → select GPT Image 2 → paste hero + product prompts
- Attach references Fable listed → generate → edit in-place for copy tweaks
- Video node → Seedance 2.0 Fast for motion preview → switch to Seedance 2.0 for finals
- Keep all nodes in one project so HiArt's design agent reads full campaign context
Effort and cost tips
Task | Suggested effort | Why |
|---|---|---|
Single hero image prompt | medium | Fast turnaround; vision analysis is still strong |
Full campaign (5+ assets) | high | Long-context planning benefits from deeper reasoning |
PDF brief + 10 references + shot list | high or xhigh | Multi-step agentic breakdown |
Quick prompt polish on one clip | medium | Avoid over-spending on rewrites |
Fable 5 API pricing is $10 per million input tokens and $50 per million output tokens. Planning is token-expensive — but one accurate prompt plan usually costs less than three failed generations on HiArt. Route hard planning jobs to Fable; route routine single-asset tweaks to a cheaper model or direct iteration on canvas.
Ready to try it? Plan your next campaign in Claude Fable 5, then open HiArt and paste the prompts into GPT Image 2 or Seedance 2.0 nodes. Explore all generation models on the models hub, or browse use cases for workflow inspiration.