Cloudways + Hermes Agent = 24/7 Faceless YouTube Automation
A daily video brief should arrive with enough information to make a decision: a topic, a reason to cover it now, sources, an outline, and a production prompt. You should not have to redo the research before deciding whether to spend money making the video.
This workflow uses Hermes on Cloudways to prepare that brief and deliver it to Telegram. After approval, you copy the production prompt into Higgsfield Supercomputer. Higgsfield handles the script and audiovisual production; you review the finished video before publishing. Google Drive holds the intermediate files so an interrupted generation has somewhere to resume.
The worked example is Lord Yapper, a stickman history channel. Hermes selected a report about a roughly 4,000-year-old tomb at Saqqara, prepared the editorial package, and sent it to Ron without a fresh morning prompt. This guide uses that recorded workflow, with current documentation checked on October 28, 2026. The screenshots are authentic cropped frames from the recording. The reusable prompts below are proposed adaptations, not an exported copy of the original skill.
Cloudways sponsored the original demonstration. Its contribution is managed hosting for Hermes. The research skill, the model’s work, and Higgsfield’s production are separate parts of the system.
Decide which work to automate
Keep four responsibilities distinct:
| Stage | Responsible tool or person | Expected output |
|---|---|---|
| Find and assess topics | Hermes, using available research tools | A source-backed candidate that fits the channel and avoids repeats |
| Decide whether to produce | You, reading the Telegram brief | Approval, a revision request, or rejection |
| Produce the video | Higgsfield Supercomputer | Script, scenes, voiceover, animation, music, captions, thumbnail, and final edit |
| Check and publish | You | An approved video with accurate metadata |
The demonstrated Cloudways instance did not use a persistent browser to operate Higgsfield. The handoff was manual. Do not expect deploying Hermes to connect your Higgsfield account or to start production automatically. Ron regarded that gap as useful: it prevented a research agent from approving every idea and consuming generation credits. Source: handoff and approval boundary, 4:15–6:15.
You can automate preparation before automating publication. A brief that you reject costs research time and model usage; a fully generated video that you reject can also consume substantial production credits.
Prerequisites: prepare the channel brief and accounts
Before deployment, write down the channel’s audience, subject area, visual style, topics already covered, and topics to avoid. For Lord Yapper, the useful constraint was more specific than “history”: stories that can become understandable stickman explainers, including archaeological discoveries, anniversaries, and historical events connected to current interest.
Prepare these inputs:
- A paid Cloudways account and an LLM provider credential supported by the deployed agent. Provider usage is billed separately from hosting.
- A Telegram account if you want briefs delivered to your phone. The browser interface is sufficient for interactive work.
- Higgsfield access with enough credits for a controlled production attempt.
- A Google Drive folder for the project, connected to Higgsfield when you reach production.
- A list of published and already approved topics. A plain document is enough to start if Hermes can read it.
The recorded setup chose Cloudways’ Operator instance because Ron wanted room for research data. It used a single research agent, not a group producing video assets. That is an example configuration, not evidence that every channel requires that tier. Size the instance for the work it will actually run and watch storage growth as research accumulates. Source: instance choice, 6:54–7:35.
Deploy Hermes and confirm that the model responds
In the Cloudways dashboard, open Cloudways AI → Managed AI Agents, then Launch your Agent. Give the instance a project-specific name, select Hermes, choose the region, and continue to the size and provider configuration. Use the price and specifications shown at checkout. Cloudways’ current onboarding article demonstrates OpenClaw and warns that some fields differ for Hermes; the frame below shows the actual Hermes selection from this workflow. Managed AI Agents require a paid account. Official Cloudways onboarding.
After provisioning, open the instance’s details. In the recorded interface, Agent Details contains the workspace access password, and SSH credentials are separate. Open the workspace using the supplied access details. Keep those credentials out of the research skill, Telegram brief, and screenshots. Source: workspace access, 7:35–8:13.
Send a small request before configuring automation: ask Hermes to identify the active project and produce a short reply. A running instance proves that the service started; a successful response also checks the model connection. If the request fails, inspect the provider configuration and error before adding more moving parts. The original demonstration later exposed a provider rejection even while Telegram delivery itself worked.
The provider determines the reasoning and API bill. Cloudways hosts the agent process. Higgsfield has a separate generation bill. Keep those costs separate when estimating whether the workflow is worthwhile.
Connect Telegram and set the delivery destination
Create a bot through Telegram’s official @BotFather using /newbot, choose its display name and unique bot username, and keep the returned token private. Enter it through your Hermes deployment’s Telegram configuration and authorize your own Telegram user. Current Hermes also offers a QR creation path in its dashboard; the manual BotFather route matches the recording. Open the bot conversation, start it, and send a test message. Official Hermes Telegram setup.
In the intended destination chat, send /sethome. The home channel tells Hermes where to deliver scheduled results and cross-platform messages. It is not a substitute for provider authentication, and it is not what makes all ordinary chat replies possible.
If /sethome reports that the agent is busy, wait for that turn to finish and retry. Confirm the destination afterward. Then request a harmless delivery test from the web workspace and check that it reaches the intended chat. Test an ordinary message from Telegram as well. These checks distinguish outbound notifications from a working two-way model conversation.
Treat the web conversation and the Telegram conversation as separate sessions unless you have deliberately configured shared context. The recording explicitly distinguishes them. Keep the reusable channel instructions in a saved skill or project file so a daily run does not depend on remembering something said only in a browser chat.
Write a research skill that produces a decision-ready brief
A useful skill contains the channel rules and the research procedure. It also defines what the agent should do when the evidence is weak. Otherwise a daily schedule can pressure it into delivering a confident idea even when it found nothing suitable.
Hermes skills use a file named SKILL.md, with name and description frontmatter. The recording sometimes calls it “skills.md”; use the actual singular filename when creating a skill. Current Hermes stores user skills under ~/.hermes/skills/ and can create or update them through the agent. Ask it to save a named skill and show you the resulting contents and location. Confirm the skill appears in /skills and load it by name for the first trial. Official skill creation and use.
The original lordyapper-production skill has a useful editorial boundary: Hermes proposes the historical subject and factual outline, while Higgsfield chooses the creative production details. It asks for deduplication, a reason the topic matters now, title alternatives, a description, and one self-contained production prompt.
Use this proposed brief to create your own skill, replacing the channel details with your actual requirements:
Create a reusable skill named channel-research-brief.
Channel: a stickman history explainer channel.
Read the saved channel profile and published/approved topic list first.
Research current signals using available YouTube, Google Trends,
Reddit, and primary historical or archaeological sources.
If a source is unavailable, say so. Never invent trend measurements.
Check the date of the underlying event as well as the article date.
Compare candidates for channel fit, factual support, visual potential,
and overlap with existing videos. Select at most one idea. If none
has sufficient support, return a short no-pick report with the reason.
Produce:
- The date and one-sentence idea.
- Why this subject is timely, with links and dates.
- A recommended angle and the historical question it answers.
- A 6–10-section factual outline.
- One primary title and two alternatives.
- A description with source links and clearly marked uncertainties.
- One self-contained Higgsfield production prompt.
- A short creator review checklist.
Keep factual research separate from creative production choices.
Do not generate a full script or dictate every scene here.
Do not launch production, spend generation credits, or publish.
Save the brief and its sources in the project research folder.
Run the skill manually once. Read the actual source pages, not just the proposed title. Check that the outline follows from the evidence and that the selected topic differs meaningfully from previous uploads. If the output is wrong, revise the saved skill before scheduling it; correcting one chat reply does not necessarily correct tomorrow’s instructions.
The Saqqara example shows why date checking matters. A historical subject may be old while the news hook is recent, but a recently published article can also recycle an earlier discovery. Ask the agent to distinguish the discovery date, announcement date, and reporting date where available. Do not call a story “newly discovered” merely because it appeared in yesterday’s feed.
Schedule one daily run and verify the first delivery
After the manual brief is useful, ask Hermes to create one daily scheduled job that loads the named skill and delivers its final output to Telegram. State the time zone explicitly. For example:
Schedule channel-research-brief every day at 09:00 Asia/Hong_Kong.
Load that saved skill for each run and deliver the final brief to my
configured Telegram home chat. Do not create a second schedule if
this job already exists. Show the saved job ID, schedule, time zone,
next run, attached skill, and delivery destination.
This is a suggested natural-language setup request, not a shell command. Hermes’ current scheduler supports recurring schedules, attached skills, and explicit delivery targets. A scheduled run has its own context, so attach the skill and give it persistent project inputs. The scheduler delivers the final response; the prompt need not tell the agent to send a second copy. Official scheduled-task reference.
Read back the saved job. A chat response saying “done” is insufficient if it omits the destination or next run. Check that the deployed version interpreted the time zone as intended. Run a single test through the configured job where supported, then wait for an actual scheduled delivery before assuming the workflow is ready for unattended use.
If the brief arrives twice, inspect the job list and delivery settings before changing the research prompt. If it never arrives, check whether the job ran, whether the model returned a final response, and whether delivery failed. Those are different faults. Avoid creating another recurring job as a troubleshooting shortcut.
Approve the idea and hand off a complete production prompt
Read the Telegram package as an editor. Open its strongest source, check the central claim, compare the outline with the proposed title, and decide whether the video deserves production credits. A title can be engaging without promising a discovery the source never made.
The Lord Yapper example used the tomb story to form a historical explanation about the person, burial, and surrounding period. Hermes supplied the subject and structure. Higgsfield retained responsibility for narration, scene choices, voice, music, captions, pacing, thumbnail, and editing. Source: creative division and completed handoff, 5:47–6:15 and 9:27–10:38.
Open Higgsfield’s Supercomputer, start a task, and paste the approved prompt. Its official documentation describes natural-language tasks and a Connectors area that includes Google Drive. Check the running credit cost as generation proceeds. Current Supercomputer also has workspace storage and a Manage Storage control, so the older temporary-file failure should not be read as a claim that every current task loses its files. Official Supercomputer guide.
Before starting, confirm that the handoff includes the channel style, factual outline, source links, uncertainties, output requirements, storage instructions, and a stop for creator review. If it says only “make the idea above,” it depends on context that the new Higgsfield conversation may not have.
Troubleshooting interrupted production and lost assets
The earlier Lord Yapper production repeatedly lost temporary working files or became stuck during long processing. Ron reports that the previous workflow took roughly twelve hours; the later research-to-video attempt took under two. These are observations from two different attempts, not a controlled hosting benchmark or a guarantee of production time. The changes included better preparation and durable storage. Source: sandbox problem, 0:36–1:47; timing comparison, 10:03–10:38.
Connect Google Drive in Higgsfield and select a project folder you can inspect. The demonstrated handoff asks for a unique dated run folder, immediate saving of intermediate assets, and an ordered manifest. That manifest records each scene or block’s index, generation job, clip, image, audio, duration, and completion state. It makes the intended sequence explicit when the task resumes.
A proposed storage clause for your production prompt is:
Use the connected Google Drive project folder for durable storage.
Create one unique dated folder for this approved production.
Save each intermediate and final asset as it is created. Maintain an
ordered manifest linking scene indexes to files, durations, generation
job IDs where available, and completion status.
If the working environment resets, inspect the saved manifest and
files before resuming. Reuse completed assets that remain valid.
Verify the assembled scene order before export. Save the final video,
thumbnail, title, description, manifest, and completion report.
Stop for my review before uploading or publishing.
After the first few assets, inspect the Drive folder yourself. A promise to save is not proof that the files are present. The manifest should refer to accessible files, and completed entries should correspond to usable outputs. After an interruption, ask for a recovery report identifying what was reused and what still needs generation. If the task cannot find its previous files, resolve the folder or permissions problem before paying to rebuild everything.
Review the finished video and measure the operation
Watch the export from beginning to end. Check the history against the approved sources, the order of scenes, narration timing, captions, music level, and thumbnail claim. Confirm that the final saved file matches the version you reviewed. Only then publish and add the topic to the channel’s completed list so tomorrow’s researcher can avoid it.
Ron compared a Lord Yapper upload with about 221 views after two weeks against an established animation channel’s example with about 423,000 views in five days. That comparison motivated the search for timely topics; it does not isolate why one video succeeded. Channel size, topic, title, distribution, and audience history differ. The Saqqara video’s eventual performance was not established in this demonstration. Source: audience and monetization problem, 1:47–4:15.
The source mentions a rough advertising-revenue assumption of $2–$5 per thousand views. Do not use it as a revenue promise or a break-even calculation for your channel. Record actual hosting, model, and production costs, including rejected generations, and compare them with your own measured results. Faster production can reduce repetitive work while the channel still loses money.
Start with a small number of approved videos. Track how many research briefs were usable, why you rejected others, how much repair each production needed, and which topics attracted viewers. Update the skill with those findings. Scheduling five or seven weak ideas a week only makes the review queue grow faster.
Choose Cloudways, a raw VPS, or your own computer
Hermes can run this research workflow outside Cloudways. A local machine must remain available for its scheduled work. A raw VPS provides an always-on host, while you take responsibility for deployment, security, updates, backups, resource monitoring, and recovery. Cloudways sells managed infrastructure around the agent. Its current Hermes page describes managed security patches, backups, and validated updates. Confirm the current backup and recovery terms for your chosen service before relying on them. Official managed Hermes hosting.
Managed hosting does not decide whether the historical claim is true, whether the chosen story is interesting, or whether the finished video deserves publication. Your saved research procedure and review still determine those outcomes.
The first useful milestone is modest: one daily brief arrives in the correct Telegram chat, cites sources you can open, avoids a topic already covered, and can be handed to Higgsfield without reconstructing missing instructions. Once that works, test one approved production and confirm that its assets survive in the chosen storage folder. Those two checks establish the workflow you will repeat.
