Explainer Videos with Claude · Higgsfield MCP and Motion Graphics
Zubair Trabzada shows how a detailed prompt, or a reusable skill file, can coordinate research, scripting, narration, images, graphics, sound and editing for a 60-second explainer. This guide reconstructs the method, checks what Claude and Higgsfield support, explains the costs and shows where human review remains essential.
Video credit: Zubair Trabzada | AI Workshop.Source video: I 100% Automated Vox-Style Motion Graphics With Claude.
Quick Summary
- The video demonstrates a system that uses Claude Opus 5 or another compatible model as the orchestrator, Claude Code as the working environment, and the Higgsfield MCP as the bridge for image, video, and audio models.
- The main request is divided into six blocks of around ten seconds. Claude researches the topic, writes the narration, plans the scenes, generates the resources, synchronizes voice and image and returns an explanatory video of approximately one minute.
- The same process can be saved in a Skill file. Instead of repeating the long prompt, the user attaches or installs the Skill and requests a theme, duration and format. The Skill contains the investigation rules, style, tools, structure and delivery.
- Higgsfield officially confirms that its MCP works with Claude on the Web, Claude Code and other supported clients. The connection gives access to more than 30 models, generational history and a credit system shared with the platform.
- Claude Opus 5 was officially released by Anthropic on July 24, 2026. It is available on Claude and the API, but is not required to reproduce the concept. The video says that the process also worked with Fable 5 and can be tested with other models.
- The expression 100% automated does not mean the absence of human work. The creator chooses the theme, connects and authorizes the MCP, defines the duration and format, provides or approves the voice, reviews the facts, controls credits, corrects errors and decides whether the result is ready for publication.
- The examples presented are selected demonstrations, not a test of reliability. The video does not show all attempts, the total cost per accepted video, an independent factual audit, or the intergenerational error rate.
--what to try first
- Use Claude as the director and coordinator of the process, not as the sole source of truth.
- Connect Higgsfield via the official MCP URL and confirm the domain before authenticating the account.
- Start with a 30-60 second video. A short duration makes it easier to validate facts, style, continuity, voice and cost.
- Structure your video into temporal blocks with one idea, one narrative function, and one visual objective per block.
- Separate research, narration, visual direction, resource generation, assembly and quality control.
- Ask for a table of affirmations with source, date and confidence level before generating the voice.
- Do not treat the instruction to consult two sources as proof of accuracy. Open the sources and confirm each important number.
- For news, war, energy, health, finance, or politics, require human review before synthesizing the narration.
- Save the visual identity as reusable rules: palette, paper texture, typography, composition, photo processing, graphics and transitions.
- Use the reference to Vox style only as a descriptive shortcut. Do not copy Vox Media logos, proprietary graphics, fonts, sequences or editorial identity.
- A Skill is an executable manual. Please read the entire file and associated scripts before installing or authorizing it.
- Explicitly set 9:16, 16:9, duration, language, rhythm, subtitles and safe zone. Do not rely on predefined values.
- The cost combines the use of the Claude model, Higgsfield credits, failed attempts, versions, voice, music and eventual upscale.
- Compare cost per approved video, not cost per click in Generate.
- Use only your voice or a voice with clear authorization. Confirm consent and rights before cloning, sharing or publishing.
- Make at least three short generations before increasing the duration. The best demonstration does not show the normal variance of the system.
- Review each plan for illegible text, wrong numbers, made-up maps, misshapen objects, continuity, synchronization, and excessive audio.
- Keep one editable project or, at the very least, separate resources. A final MP4 is not enough to quickly correct a fact or a sentence.
- The transcript does not contain an explicit statement of sponsorship. The video links to the creator's resources and communities, so always check the tags and any reference links in the original description.
The short answer
Claude and the Higgsfield MCP can coordinate much of an explainer video with editorial graphics. The system can research, script, plan scenes, call image and video models, generate voice and music, and assemble a result from a detailed request.
The important part is the word coordinate. Claude doesn't draw every frame himself or replace the entire infrastructure. Interprets the objective, chooses or calls tools and organizes the results. Higgsfield runs the generations through the available models and discounts the respective credits.
In the embedded video, Zubair Trabzada shows two paths. The first uses a long prompt to create a 60-second video on a specific topic. The second puts the same logic into a Skill file, allowing for shorter requests like creating a one-minute explanation of an AI news story.
Credit for the demo, examples, prompt pack and Skill file presented in the video: Zubair Trabzada and AI Workshop. This article adds independent research, limits, checks, and original production prompts.
The three-tier system
The first layer is the Claude model. In the video, Zubair uses Opus 5 and discusses experiences with Fable 5. Anthropic officially launched Opus 5 on July 24, 2026 and positions it for extended work, tooling, scanning, and multi-step tasks.
The second layer is Claude Code, in the desktop application or other compatible environment. This is where the agent reads the prompt or Skill, researches, plans, calls tools and tracks tasks until delivery.
The third layer is the Higgsfield MCP. MCP is an open protocol that allows an AI application to call external systems. In this case, Claude can request images, videos, voice, music and editing operations without the user manually jumping between multiple applications.
The result looks like a single conversation, but it remains a chain of services. Each tier has different permissions, limits, costs, policies, and potential failures.
Required tools
The minimum demo set is a Claude account with support for connectors or Claude Code, a Higgsfield account and authorization to turn on the MCP server.
The official Higgsfield page features more than 30 models through the connector. The current list includes image and video options like Nano Banana Pro, Seedance, Kling, Soul and others. Availability may vary by plan, region and time.
The video also uses a prompt pack and a `vox-motion-graphics.skill` file distributed in the free AI Workshop Lite community. These resources belong to the creator and must be obtained through the original links indicated in the video.
It is not mandatory to use Zubair's prompt or Skill to apply the method. The principle can be recreated with your own script, authoritative style rules, an audited Skill, and any toolkit that returns separate resources.
How to turn on the Higgsfield MCP
Starting at 1:51, the video shows the connector configuration. The official documentation indicates the URL `https://mcp.higgsfield.ai/mcp`. Always confirm the address directly on the official page before pasting or authenticating.
In Claude, open the customization or connectors area, choose to add a custom connector, name it Higgsfield, paste the URL and start the connection. The flow opens a Higgsfield authentication page to authorize access.
After returning to Claude, confirm that the connector appears active. In Claude Code environments, configuration can also be managed by the interface or by MCP commands, depending on the version installed.
Don't set broad permissions for convenience without understanding the effect. Start with call confirmation, note what tools are used, and then consider more permissive rules for a known flow.
The connection does not eliminate data sharing. The required requests, reference images, audio and results are processed by the services involved. Do not send confidential material, voices, faces or client work without checking contracts and policies.
What does a single prompt really mean?
The phrase a single prompt describes user input, not a single system operation. Between the request and the MP4 there is a sequence of searching, writing, planning, generating, selecting, synchronizing, composing and exporting.
In the 60-second example, Claude starts by researching the news. The prompt asks that numbers be compared between at least two sources, that disputed values be treated conservatively, and that sources be listed.
Then, the agent creates the narration, divides the story into six blocks, generates images and other resources, chooses a voice and calls Higgsfield to produce the necessary elements.
The merit of the prompt is to transform a long chain into an execution contract. The risk is hiding important decisions within this chain. If the investigation is wrong, automation can quickly produce a false claim with a convincing image, voice and music.
The six-block architecture
The transcription indicates six blocks for a video of approximately 60 seconds. This division creates units of about ten seconds, close to the length supported by many video models and short enough to maintain a clear idea.
A useful block includes five elements: narration phrase, visual information, movement type, on-screen text, and transition. Add the source of the fact and the status of continuity with the previous block.
A possible editorial structure is: hook, context, mechanism, evidence, consequence and conclusion. Not all themes need the same formula, but each block must have a narrative function.
The aesthetic presented combines cropped photographs, paper textures, strong typography, maps, arrows, graphs, zooms and simulated camera movement. The value is in consistent visual grammar, not the literal reproduction of an editorial brand.
Play the 60-second video
Starting at 5:18, Zubair copies a prepared prompt and replaces the theme field. It shows that the topic can come from your own notes, a personal assistant or recent news.
After submitting the request, Claude presents investigation and generation steps. At 8:39, the interface shows a call from Higgsfield to generate images with Nano Banana Pro. The agent also prepares voice and other resources.
The video states that, in the end, the user can download the result and publish it. This is the shortest route, but it should not be the production route. Before publishing, export the narration, fonts, images, audio, and final version to a review folder.
Do a first run with marked water, low resolution or economy settings. Only increase quality after the script, rhythm and visual direction are approved.
The Skill turns the method into a reusable shortcut
Starting at 9:11, the demo replaces the long prompt with a Skill file. Zubair attaches the resource and asks for a one-minute video on a relevant AI topic.
Claude reads the Skill and retrieves the saved instructions: how to investigate, what structure to use, when to call the Higgsfield, how to generate voice, how to organize six blocks and what format to deliver.
Anthropic describes Agent Skills as folders of instructions, scripts, and resources that provide procedural knowledge to the agent. They are suitable for repetitive flows because they can be discovered and loaded when the task matches the description.
A Skill also increases the risk surface. It may contain scripts, dependencies, network requests, or instructions that are not visible in the short request. The official guidance is to install only trusted resources and audit all files before use.
For client work, keep the Skill in a versioned repository. Record author, origin, version, date, authorized tools, changes and quality tests.
Format, duration and framing
On the first run with the Skill, the system chooses a vertical format. Zubair interrupts and explicitly asks for 16:9, demonstrating that predefined values can defeat the purpose.
Set at the beginning: 9:16 or 16:9, total duration, maximum duration per plan, language, reading rate, subtitles, safe zone, resolution and target platform.
Higgsfield claims that the MCP can generate clips of up to 15 seconds, depending on the model. A 60-second explainer will typically be made up of multiple generations, animated images, graphics, and audio, not a single continuous clip.
For two-, three-, or five-minute videos, increase the editorial structure first. Automatically multiplying the number of blocks without chapters, progression and visual variety produces repetition and increases cost faster than quality.
How to calculate the actual cost
Zubair warns that longer videos use more Higgsfield credits. Official documentation confirms that each generation uses credits according to the model and resolution and that the same account balance is used through the MCP.
Claude also has its own cost or limits. On the API, Opus 5 was launched at 5 dollars per million tokens in and 25 dollars per million tokens out. In Claude subscriptions, usage limits and credits follow different rules.
The full budget includes research, Claude tokens, images, clips, voice, music, versions, glitches, upscale, storage and review time. A video that looks cheap in the first generation can be expensive after ten fixes.
Record four metrics: total execution cost, number of attempts, seconds passed, and cost per second passed. Also compare the human time saved and the human correction time.
Do not publish fixed plan values in the prompt itself or in the Skill. Consult the pricing page before each production and maintain a maximum credit limit per run.
AI voice and responsible cloning
At 13:44, the demo shows a catalog voice named Sterling. Zubair explains that the user can request another voice or let the system choose.
In the Higgsfield audio area, the video shows voice options and a flow for creating a custom voice. The official Higgsfield page features integration with vendors like ElevenLabs and asks for confirmation of consent during cloning.
If you use your own voice, record in a quiet space, maintain consistent distance and interpretation, and save the original file. A short sample may work for prototyping, but quality and consistency depend on the reference audio.
Don't clone a voice found in a video, podcast or call. ElevenLabs documentation requires rights and consent and applies additional verification to professional clones.
After generation, confirm pronunciation, proper names, numbers, pauses, intention and synchronization. A natural voice can make a factual error more convincing, so the audio should only be finalized after the script has been approved.
What the results demonstrate
From 15:54 onwards, Zubair plays a video about a conflict and another about data centers and electricity. Both combine narration, cuts, zooms, maps, graphs, text and transitions into a recognizable editorial format.
The second example was started by a short request and Skill chose the topic, planned the angle and called Higgsfield. This demonstrates the portability of the process, not the accuracy of the statements produced.
The video also shows an example with sound effects and music. The absence of music in the first two was a choice, not a stated limitation of the system.
The selected results show an interesting quality ceiling. They don't show the complete distribution: how many generations failed, how many features were rejected, how much each version cost or how much assembly work took place off-screen.
Factual control is the most important step
Examples deal with war, energy prices, attacks, infrastructure and predictions. These are topics where dates, numbers and sources can change quickly and where one wrong sentence can cause damage.
Before generating the voice, convert the script into a table of statements. For each factual sentence, record primary source, independent source, event date, publication date, supporting citation, confidence level, and final wording.
Opening two pages is not the same as confirming a fact. Two news stories may repeat the same original source, be out of date, or use incompatible estimates.
Do not use the video model to validate the investigation. The visual plan must only receive facts that have already been approved. If a number is in dispute, explain the uncertainty or remove it.
Make a final check after assembly. On-screen text, graphics and maps can introduce errors that did not exist in the narrated script.
Learn grammar without copying the brand
The Vox-style expression helps the public imagine a quick editorial explanation, with clippings, data, maps, typography and documentary rhythm. Does not imply collaboration or approval by Vox Media.
Transform this reference into an original specification. Choose your palette, licensed typographic family, cropping proportions, texture library, graphics treatment, icon system, transition speed and sound language.
Avoid logos, headers, proprietary fonts, graphic signatures, recognizable sequences, or shot-by-shot copying of a published video.
Consistency improves when visual rules are concrete. Saying documentary and editorial is less useful than providing colors, margins, thicknesses, grid, shadow treatment, rules for photographs and authorized examples.
A safer production workflow
First, approve the theme and objective. Define audience, platform, duration, intended conclusion, permitted sources, and prohibited statements.
Second, do research and freeze the script. No visual generation begins until the assertion table is approved.
Third, create the scene sheet. For each block, associate narration, visual resource, movement, text, source and continuity.
Fourth, generate a two-block sample. Validate style, voice, rhythm and cost before ordering the full minute.
Fifth, generate separate resources and maintain stable names. One folder per block simplifies replacements and auditing.
Sixth, assemble and do quality control. Review at normal speed, without sound, with just sound and frame by frame in trouble spots.
Finally, save the production package: briefing, sources, approved script, Skill used, versions, definitions, costs, resource rights, voice consents, project and export.
Predictable crashes and fixes
If the format is wrong, lock the aspect ratio, resolution and preset at the beginning of the Skill. The demo itself shows a vertical run that had to be stopped.
If your branding varies, reduce the number of templates, use consistent references, and first generate a stylesheet with approved examples.
If the graphs are wrong, generate the data with code or a deterministic tool and animate the result. Don't ask an image model to simultaneously invent and draw a factual graphic.
If the voice stumbles over names or numbers, write pronunciation, separate sentences, and regenerate only the affected audio.
If the cost skyrockets, set limits per block, use animated images where generative video doesn't add value, and approve the script before generating in high quality.
If the video seems generic, improve the thesis and point of view. More movement doesn't make up for a story with no question, tension or conclusion.
Verdict: Automate execution, preserve judgment
Zubair's demo is valuable because it shows an interface change. The creator stops operating search, image, video and voice tools separately and starts directing a chain through natural language and reusable rules.
The Higgsfield MCP makes this orchestration technically plausible, and the Agent Skills make the method portable. Opus 5 improves the ability to maintain long tasks, call tools, and check work, but none of these pieces guarantee factual truth or good editorial taste.
The best use is not to publish everything that comes out of a prompt. It's about reducing mechanical work so that the person can devote more attention to research, structure, originality, rights, rhythm and quality control.
Watch the embedded video for complete setup and results. Use the process as a starting point, not as authorization to skip review or replicate another publication's identity.
Prompt Pack for AI Explainers
These original prompts complement Zubair's demonstration. They are written to transform a generic request into an auditable process with research, storytelling, visual direction, cost control and human review.
Prompt pack credit: Zubair Trabzada | AI Workshop.
1. Editorial briefing and limits
Use before the survey to define what the video should explain and what the agent cannot invent.
Create a production brief for an explanatory video with motion graphics. THEME [theme] OBJECTIVE [what the public should understand or do] AUDIENCE [public] DELIVERY Duration: [duration] Format: [9:16 or 16:9] Platform: [platform] Language and variant: [language] Tom: [tone] PERMITTED FONTS [primary sources, press, reports] RESTRICTIONS [prohibited statements, brands, people, private data, visual limits] Returns: 1. Central question 2. Provisional thesis 3. Narrative structure in six blocks 4. Facts that need verification 5. Editorial, legal and security risks 6. Necessary Visuals 7. Approval criteria 8. Cost limit and attempts Don't write the narration yet. Do not invent numbers, quotes, sources or events.
2. Investigation and statement table
Runs before voice and images to prevent an error from running through the entire automation.
Investigate this topic for an explanatory video: [theme and briefing approved] Create a table with: - Claim ID - Proposed formulation - Primary source - Independent source - Date of the event - Publication date - Excerpt or supporting data - Agreement or conflict between sources - Confidence level - Safe formulation for narration Rules: 1. Open and check each source 2. Do not count two news items that repeat the same origin as independent sources 3. Distinguish fact, estimate, forecast and opinion 4. Flags numbers that have changed or are in dispute 5. Remove claims without sufficient support 6. List final URLs 7. Indicates the investigation cutoff time In the end, it presents only the approvable facts and questions that require human decision. Do not proceed to visual generation.
3. Six-block visual guide
Convert approved facts into a ready-made scene sheet for Claude and Higgsfield.
Transforms the briefing and the table of approved statements into a 60-second video divided into six blocks. ORIGINAL IDENTITY Palette: [colors] Licensed typography: [typography] Textures: [textures] Photographic treatment: [rules] Graphics: [rules] Transitions: [rules] Sound: [rules] For each block, it returns: 1. Entry and exit time 2. Narrative function 3. Narration with a maximum of [number] words 4. ID of each statement used 5. Visual composition 6. Resources to generate or obtain 7. Camera movement and animation 8. On-screen text 9. Sound effect or music 10. Transition 11. Continuity with the previous block 12. Approval criteria In the end, it creates: - Full narration - List of resources by folder - Separate generation prompts - Assembly instructions - List of what cannot be changed Uses original editorial grammar. Do not copy logos, fonts, layouts or plans from an existing publication.
4. Final audit, cost and decision
Applies to all results, including rejected attempts.
Audit this explanatory video before publishing. ENTRIES Briefing: [paste] Affirmation table: [paste] Approved script: [paste] Scene sheet: [paste] Claude model and mode: [details] Higgsfield models: [details] Costs and credits: [details] Video version: [ID] Rates from 1 to 10: - Factual accuracy - Correspondence between font, voice and text on screen - Narrative clarity - Visual originality - Consistency of style - Continuity - Readability - Voice quality and pronunciation - Audio synchronization and mixing - Visual artifacts - Rights and consents - Format and safe zone Returns: 1. Approve, correct or reject 2. Factual errors with timecode 3. Resources to replace 4. Minimal fix for next version 5. Total Attempts 6. Approved seconds 7. Total cost 8. Cost per approved second 9. End credits for people, tools and sources 10. Publication checklist Do not approve a visually beautiful video that contains an unauthorized fact, voice, brand or graphic.
Video Timestamps
FAQ
Is it even possible to create a complete explainer video with a single prompt?
It is possible to start a complete chain with one order. Claude divides the work into research, script, resources, voice and editing and calls tools through MCP. There continue to be multiple operations, costs, and decisions behind that single order.
Is the process really 100% automated?
Not in the literal sense. The user sets up accounts and permissions, chooses or approves theme, format, voice and style, controls credits, checks facts, reviews the video and decides what to publish.
Who created the demo?
The video, examples, prompt pack and Skill presented belong to Zubair Trabzada, from the AI Workshop channel.
What is the Higgsfield MCP?
It is a Model Context Protocol server that allows compatible clients, including Claude and Claude Code, to call image, video, and audio tools and models available in Higgsfield.
What is the official URL for the Higgsfield connector?
The official page features `https://mcp.higgsfield.ai/mcp`. Please confirm the address in Higgsfield documentation before authenticating.
Is it mandatory to use Claude Opus 5?
No. Zubair uses Opus 5 and reports results with Fable 5. The concept may work with other models capable of holding long tasks and calling tools, although quality, cost and persistence vary.
Does Opus 5 officially exist?
Yes. Anthropic released Claude Opus 5 on July 24, 2026 and made it available on its platforms and API.
What is an Agent Skill?
It is a structured folder or file with instructions, resources, and sometimes scripts that teach the agent a repeatable process. It must be audited before installation because it may execute code or connect to external services.
Can I create videos in 9:16 and 16:9?
Yes, but state the proportion explicitly. The demo shows a vertical run that was stopped and restarted at 16:9.
How much does a 60 second video cost?
There is no single price. The cost depends on the Claude model, number of tokens, Higgsfield models, resolution, voice, music, number of resources and attempts. Measure the total cost per approved second.
Does Higgsfield charge separately when used by MCP?
The connector uses the same credit system as the Higgsfield platform. Each call consumes credits depending on the model and resolution.
Can I use my own voice?
Yes. The video shows creating a custom voice in Higgsfield. Only use your voice or material with authorization and confirm the rules of the selected supplier.
Can I clone someone else's voice with permission?
Rules vary by vendor and clone type. ElevenLabs, for example, requires confirmation of rights and restricts professional clones to the voice verified by the owner himself. Consult the applicable policy before creating or sharing the clone.
How to avoid errors in graphics and news?
Approve the script and a table of statements before generation, create graphs from verified data, and separately review voice, on-screen text, maps, and values.
Is this lawsuit affiliated with Vox Media?
There is no indication of affiliation. The Vox-style expression is used as a descriptive reference. Create your own identity and do not copy proprietary assets or graphics.
Was the video sponsored?
The transcript provided does not include an explicit statement of sponsorship. The video links to the creator's community and external tools, so check the original description for tags and any referral links.
--sources and credits
- Zubair Trabzada | AI Workshop: I 100% Automated Vox-Style Motion Graphics With Claude
- Zubair Trabzada Channel | AI Workshop
- AI Workshop Lite: Community and Free Resources
- Higgsfield: MCP official page
- Higgsfield: official guide to generating video through Claude
- Higgsfield: current prices
- Higgsfield: voice cloning
- Anthropic: official launch of Claude Opus 5
- Anthropic: Claude Code
- Model Context Protocol: official introduction
- Anthropic: Agent Skills for Reusable Flows
- ElevenLabs: how voice cloning works