The Content Engine Playbook
A hands-off short-form video system that runs two lanes: faceless clips an AI generates for you, and talking-head clips you film yourself. It sources ideas, writes the scripts, edits every clip, sends each one to Telegram for a one-tap approve, and posts. Built on n8n, Claude, Higgsfield, Submagic, and Blotato. You just film and approve.
Section One
What You're Building
A hands-off short-form pipeline that runs two lanes at once: videos the AI generates for you, and videos you film yourself. You build it once, then it sources, scripts, edits, and posts on a weekly rhythm while you just film and approve. Each stage below is one prompt, written to paste straight into your AI builder, with the landmines baked in as guardrails. Swap any tool for your own; the wiring is what carries over.
Section Two
What You'll Need
Airtable
The record of every post: status, script, keyword, and the finished video URL.
Claude
Writes the scripts, and if you like, wires the build itself.
Higgsfield
Generates the faceless AI-lane videos. Any text-to-video tool can stand in.
Submagic
Edits the clips you film: b-roll, captions, zooms, clean audio.
Google Drive
The drop folder you film into, and the source library you pull from.
Telegram
Where each finished clip lands for a one-tap approve or decline.
Blotato
Posts approved clips out to your socials.
n8n
Holds the wiring together and runs it on a schedule.
Section Three
The Build
01
Step One
Start here: the architecture brief
Most bad builds come from a model that only sees one node at a time. Set the full picture and the two rules that govern everything, then build in sequence.
The rule
Establish the map and the money rules up front. A model that knows the destination wires each node to fit it.
Copy this: architecture brief
I'm wiring a hands-off short-form content pipeline in n8n. Before we build any single piece, hold the whole shape in mind. We build one stage at a time: source and script, split into two lanes, the AI lane, the live lane, review in chat, and publish. Stack: Airtable is the record of every post. Claude writes scripts. Higgsfield generates the faceless AI videos. Submagic edits the clips I film. Google Drive is the drop folder. Telegram is where I review. Blotato posts to my socials. Two rules for the entire build: - Never re-render or re-generate anything already produced. It costs real API money. When testing, read state; do not re-run render steps. - Every paid API bills from its own wallet, separate from any web-app subscription. Before calling a failed job a bug, confirm that API wallet is funded. Acknowledge this architecture back to me, then we start at stage one.
02
Step Two
Source and script
The engine is only as good as its raw material and its rules. Point it at your real library, then hold the script to numbers that actually perform.
Why this works
The hook earns the first six seconds and the word count keeps you at 45 to 55 spoken. Spec beats taste when a machine writes at volume.
Good to know
Dedupe on angle, not just topic. One source can feed a dozen posts if each one opens on a genuinely different idea.
Copy this: source and script
Build stage 1: source and script. Sources: have Claude scan [my Google Drive folder of existing material] plus [the websites my niche reads] and [the forums my audience lives in]. Pull from files that already exist, not just new uploads. For each source, extract up to 3 distinct angles and write one script per angle. Enforce these in the prompt: - Script 110 to 135 words (45 to 55 seconds spoken). Never exceed 140. - Hook: max 8 words. A bold claim or curiosity gap that reads as on-screen text and earns the first 6 seconds. Not a sentence, not clickbait. - Dedupe against angles already used for that source. Write each result to Airtable as a row with status = scripted. Cap the active queue so the system never overproduces.
03
Step Three
Split into two lanes
This is the heart of the half-and-half system. Sort each post by content, hold a weekly ratio, and hand yourself a filename you never have to invent.
The rule
The machine assigns the keyword, not the human. A person guessing filenames is the single most common way this pipeline breaks.
Copy this: split into two lanes
Build stage 2: split every week into two lanes. AI lane: faceless, no person on camera, generated automatically. Live lane: talking-head, I film myself. Classify: audience-question, first-person, opinion, or hot take goes to the live lane. Historical, atmospheric, concept-explanation, or list goes to the AI lane. When unsure, AI lane. Enforce a weekly ratio of [your split, for example 2 AI and 3 live] with a hard cap, topping each lane up based on what is already queued. For every live-lane row, auto-generate a short, filename-safe keyword (for example topic-slug-a1b2) and save it on the record. I will name my video file exactly that keyword so the system can match it later. Never make me name files by hand.
04
Step Four
The AI lane: the half you never touch
These posts write, render, caption, and queue themselves. Your only job is to keep the credits topped up and approve the result.
Good to know
Any text-to-video generator drops in here. What matters is the guard around it: render once, check credits first, honor the cap.
Copy this: the AI (faceless) lane
Build stage 3: the AI (faceless) lane. For every AI-lane row, automatically produce a faceless video with no person on camera. Send the script to Higgsfield to render the visuals, add a voiceover from the script, and burn in captions. Save the finished video URL back to the row and move it to review. Rules: - Render only rows that have never been rendered. Re-rendering a finished video is pure wasted spend. - Check Higgsfield's credit balance before submitting. If it is short, leave the row queued for the next run instead of failing. - Respect the weekly cap from stage 2. Don't drain credits making more than you'll post.
05
Step Five
The live lane: drop a file, it gets dressed
You film in your window, drop the clips in one folder, and the system does the rest: match, edit, save. The care is all in handling more than one at a time.
The rule
A folder poll hands you a batch, not one file. If any node assumes a single item, every clip after the first vanishes silently.
Good to know
The filename is a contract. Let the machine name it and you never think about this again.
Copy this: intake and edit the clips
Build stage 4: intake and edit the clips I film. Watch one Google Drive folder. When a video lands, take the filename minus its extension as the keyword, match it to the queued row, and send the video to Submagic to add b-roll, captions, zooms, clean audio, and remove bad takes. Save the finished video URL back to the row. Wire these correctly the first time: - Share the folder "anyone with the link" so Submagic can fetch the file. - One folder poll returns all new files at once. Every downstream step must handle N files in a single run, not just one. - Filename to keyword must match exactly. A space vs a hyphen, or a "(1)" suffix, breaks the match. - Poll only during my working window (for example twice on my filming evening) to save executions.
06
Step Six
Review each clip in chat
Both lanes land here. The review step is where most builds quietly fail on a batch. The fix is to stop waiting for one reply and start listening for taps.
The rule
Send-and-wait is one-at-a-time by design. Drop five clips at once and four are stranded. Use callback buttons plus a listener for real batch review.
Good to know
A decline shouldn't be a dead end. Send it back to scripted so it re-enters the queue instead of rotting as a rejected row.
Copy this: review in Telegram
Build stage 5: let me review each clip in Telegram. Send each finished clip as its own message with a Watch link and its own Approve and Decline buttons. I approve or decline each one independently, in any order. Do not use a "send and wait for one reply" step. It handles only one item and silently strands the rest of a batch. Instead: - Put callback buttons on each message carrying that row's id. - Build a separate listener that catches the taps and updates each row. - Approve sets status = approved. - Decline sends it back to the queue (status = scripted) so I redo it. Don't leave dead rejected rows lying around. Telegram parses message text as Markdown, so a value with an underscore (like youtube_shorts) throws "can't parse entities". Strip Markdown characters from any dynamic text, or send in plain or HTML mode.
07
Step Seven
Publish to the right brand
If you run more than one brand or channel through one publisher, the default behavior will post to the wrong one. Name the target explicitly and it never happens.
The rule
First account per platform is a coin flip on a multi-brand publisher. Match the account by name, or you post one brand's clip to another brand's channel.
Copy this: publish through Blotato
Build stage 6: publish approved clips through Blotato to [my platforms, for example TikTok and YouTube]. Read this before you pick an account: my Blotato account holds profiles for several brands and channels. Do not select "the first account per platform"; that posts to the wrong one. Select the account whose name matches [this brand's name or handle] and post only to those. Skip any platform with no matching account. Build the caption as caption, then a blank line, then hashtags. Post one approved clip per run on a daily drip, then mark it scheduled so it never double-posts.
08
Step Eight
The guardrails
Eight landmines, each one an hour to learn. Paste them into the build as standing rules and they cost you nothing.
▸ The 8 guardrails
- Separate API wallets — a failed render is usually an empty API wallet, not a bug. Fund it, then debug.
- Never re-render finished work — verify by reading status, not by re-running the expensive step.
- Paused runs are frozen in time — a waiting run uses its start-time snapshot; start a fresh run to test.
- Assume batches, not items — one trigger can carry many records; a single-item assumption drops everything after the first.
- Pick accounts by name — on a multi-brand publisher, never take the first account.
- Pin schedules to one clock — a workflow timezone plus a local time double-shifts every trigger. Convert once.
- The filename is a contract — keyword matching is exact; let the system name files.
- Sanitize chat text — underscores and asterisks break Markdown parsers mid-send. Clean dynamic text or send plain.
“The filename is a contract. Let the machine name it and you never think about this again.”
— The Content Engine Playbook, p. 6
§ Fin.
Rather have it built for you?
This guide is one system from the Soulwire Studio workflow library. We build AI automations for businesses that run on process: intake, scheduling, proposals, compliance-heavy operations. If you'd rather spend your time closing deals than debugging content, book an audit and we'll find the automation hiding in your business.
