I have a problem, and the problem is YouTube.
Not the "I watch too much" kind. The other kind. Someone sends me a 40-minute video with the message "you HAVE to watch this", and I spend the next 40 minutes finding out whether I had to. Usually I did not. Usually the good part was four minutes long and started at 23:10.
So I built a workflow that watches the video for me and tells me whether it's worth my time. Then I published it on Edgaze and put a price on it. This post is that whole afternoon, screen by screen. Every click, every field in the publish flow, and the bit where it quietly read the Arabic subtitles and I didn't notice for an hour.
By the end you'll know how to take an idea, turn it into a workflow, test it, publish it, and price it so people actually run it.
What we're building: paste a YouTube link, say how many minutes you have, get back Watch, Skim or Skip, three honest reasons, and the three moments worth jumping to.
Time: about 20 minutes. No code. One of those minutes is me complaining about subtitles.
Step 1: Open Edgaze
Go to edgaze.ai. That's the whole step. I'm including it because I promised every single step, and I'm a man of my word.
If you just want to look around, the marketplace is open without an account. If you want to build and publish, you'll need to sign in. Technically you can start building in Workflow Studio as a guest (your draft lives in that browser), but publishing needs an account, so let's do it properly.
Step 2: Sign in
Hit Sign in anywhere. Google is one click. Email and password works too, and so does "More sign in options" if you're a GitHub person.
New here? Create account at the bottom. It takes less time than reading this paragraph did.
Step 3: Open Workflow Studio
In the left sidebar, the hammer icon is Workflow Studio. Click it and you land on the launcher: start something new, pick up a draft, or open one of your published workflows.
Hit New.
Step 4: The empty canvas (don't panic)
This is the first screen of a new workflow, and it is gloriously empty. A dotted canvas, a box in the middle that says Composer, and a small + underneath for people who prefer to build by hand.
You have two options:
- Describe what you want to Composer, and it builds the workflow for you.
- Build it manually: press
/or double-click the canvas and add blocks yourself.
We're going to do the first one, then go through every block it made, so you understand what's on the canvas and could have built it yourself. Using Composer is fine. Not understanding what it built is how you end up publishing something you can't fix.
Step 5: Ask Composer for the workflow
Type what you want the way you'd explain it to a friend who happens to be very good at building things. Be specific about the inputs (what the person running it gives you) and the output (what they get back).
Here's exactly what I typed:
I want a workflow that tells me whether a YouTube video is worth my time. I paste a link and say how many minutes I've got. It should grab the video's transcript, then give me a verdict (Watch, Skim, or Skip), three honest reasons, and the three best moments with timestamps so I can jump straight to them.Hit send. Composer went off and came back in about twenty seconds with a five-block workflow, a short explanation of what it did, and a little pill at the bottom that says 1 change · Valid · low cost.
Click that pill and you get the change history: what the canvas looked like before (nothing), what it looks like after (something), and how many edits it took to get there. If you don't like a change, the undo arrow next to it rolls it back.
Step 6: Inspect every block
Close Composer and hit Fit graph in the top toolbar to see the whole thing.
Five blocks, left to right:
- Available Time: asks the person running it how many minutes they have.
- Video Transcript: the YouTube integration. Asks for a link and fetches the transcript.
- Viewing Context: stitches the minutes and the transcript into one package.
- Watch Verdict: the AI step. Reads everything and decides.
- Video Recommendation: what the person running it actually sees.
Click any block and the Inspector opens on the right with its settings. Let's go through them in order.
The input: "How many minutes do you have?"
An input block is a question you ask whoever runs the workflow. Whatever you write in Question and Description shows up on their run form, word for word, so write it like a human: "How many minutes do you have?" beats "Enter duration_minutes (integer)" every time.
The integration: getting the YouTube transcript
This is the "simple integration" part, and it's simple: no API keys, no accounts to connect, nothing to authorise.
The settings that matter:
- Source: YouTube video. The same block can read web pages; this tells it to fetch a transcript instead.
- URL source: Ask the user when this runs. The link becomes a question on the run form. You could hard-code a URL instead, but then your workflow only ever judges one video, which is a very niche product.
- Question / Description: "Which YouTube video should we assess?" and "Paste the YouTube video link."
- Transcript style: Timestamped lines. This one is important. Without timestamps, the AI can't point you to moments, and the whole "jump straight to 06:19" promise falls apart.
- Caption language (optional): leave this blank and you'll learn something about TED talks. More on that in a minute.
The merge: "Viewing Context"
A merge block takes several inputs and joins them into one thing the next block can read. Here, the minutes and the transcript. It's plumbing. Nobody ever sees it. It's the kitchen of the restaurant.
The AI step: "Watch Verdict"
This is where the actual thinking happens.
Type: Generate Text, model GPT 6 Sol (marked Recommended). The prompt Composer wrote is honestly pretty good:
Assess whether the YouTube video is worth the viewer's available time, using only the timestamped transcript and the minutes available in the inputs above. Return exactly one verdict: Watch, Skim, or Skip. Watch means the full video is worthwhile within the available time, Skim means a few segments offer value, and Skip means the likely value does not justify the time. Do not assume the video's total duration unless the transcript establishes it. Give exactly three honest, specific reasons grounded in the transcript, including limitations or uncertainty where relevant. Then identify exactly three best moments with timestamps copied from or reliably derived from the supplied transcript, each with a concise description of why it matters. Prefer timestamps within the viewer's time budget when possible, but do not claim that jumping to a late moment fits the budget. Format the response with headings Verdict, Three reasons, and Three best moments. If the transcript lacks usable timestamps or has fewer than three identifiable moments, say so explicitly rather than inventing timestamps or moments.Notice what makes it good: it defines each verdict instead of hoping the model guesses, it asks for exactly three of things, it forbids inventing timestamps, and it fixes the output format with headings. Those four habits will make almost any prompt better.
The output: "Video Recommendation"
Every workflow ends in an output block. It's the plate the food goes on.
Step 7: Write the prompts like a human (give it a personality)
The prompt tells the model what to do. The Style / System field right under it tells it how to sound. Composer left it empty, so I filled it in:
You're the friend who has watched far too much YouTube and has opinions about it. Be warm, direct and a little funny. No hype and no 'in today's fast-paced world'. If a video is mostly padding, say so kindly, then point to the part that isn't.If a prompt gets long, the little expand icon next to the field opens a proper editor, with character, word and rough token counts at the bottom. Very useful when you're trying to work out why a step costs what it costs.
Step 8: Run it
Top right, Run. A run window opens and asks for the two inputs, exactly as a buyer would see them. There's a Customer view / Builder view switch at the top: Customer view is what buyers get, Builder view shows every block's inputs and outputs for debugging.
I tested it on a video I know well: Tim Urban's TED talk Inside the mind of a master procrastinator. Picked partly because it's great, partly because a procrastination talk is the most on-brand possible test for a "should I watch this" tool. Ten minutes available.
It runs block by block. You can watch it go.
About 55 seconds later:
Skim. The talk runs past 13 minutes, so it won't fit in ten. The best moments are at 04:58 (the Instant Gratification Monkey), 07:34 (the Panic Monster) and 11:10 (where it turns to procrastination with no deadline, the part with the bigger consequences). My favourite line: the Monkey's "easy and fun" logic "explains why a sensible plan loses to a YouTube spiral." That's the style field doing its job, and also a personal attack.
Flip to Builder view to see every block, its status, and what it passed along:
The subtitle story (or: always check Builder view)
My first test run also said Skim. Great verdict, sensible reasons. I was very pleased with myself.
Then I opened Builder view and looked at what the transcript block had actually fetched. It started with "المترجم: Ali Sumait".
That talk has subtitles in dozens of languages. I'd left Caption language blank, so the block took the first track it found, which was Arabic. And the AI, being an AI, read the Arabic transcript and wrote a perfectly good English verdict anyway. It was right, for reasons I hadn't asked for.
The fix is one field: Caption language set to en. Now it reads the English captions every time.
Step 9: Publish (every screen)
When you're happy, hit Publish in the top toolbar. The publish flow has five steps: Details, Pricing, Media, Demo inputs, Review. The card on the right is a live preview of your listing, and it updates as you type.
Details
Going field by field:
- Title. What people search for and click on. "YouTube Watch Verdict" was Composer's name; I changed it to Should I Watch This? because that's the question people actually have.
- Description. One or two plain sentences: what you give it, what you get back. Mine: "Paste a YouTube link and say how many minutes you have. You get Watch, Skim or Skip, three honest reasons, and the three moments actually worth jumping to."
- Tags. Comma-separated, up to ten. Buyers find workflows by these. I used
youtube, video, productivity, summary, time saver. - Category. Leave it on Automatic and Edgaze picks one, or choose yourself. I picked video.
- Edgaze code. Your workflow's short name and its link:
edgaze.ai/your-handle/your-code. Anyone can type the code on edgaze.ai and land straight on your workflow, no link needed. It checks availability as you type. Mine isshould-i-watch. - Visibility. Public shows up in the marketplace and on your profile. Unlisted works for anyone with the link but stays out of search. Private is just you.
- Release notes. Shown to buyers in their library. Optional on version 1, useful from version 2 onwards.
- Allow remixing. Other creators can use your workflow as a sealed block inside theirs, and you earn 25% every time their remix runs. They never see your prompts. I leave this on; it's free distribution.
Pricing
This step gets its own section below, because it's the one people get wrong. Quick tour first.
- Free or Paid per run. Free means anyone can run it and you don't earn a margin. Paid per run means buyers pay a fixed price every time it runs.
- Model failover. If a model provider has an outage, Edgaze substitutes a nearby model so runs still finish. On by default; keep it on unless you have a reason.
- Provider credentials. Who pays the model providers. Edgaze hosted (the default: Edgaze runs the models and builds the cost into the price), your connected keys, or buyer brings keys.
Switch to Paid per run and the real numbers show up:
I set my margin to $0.20:
Media
- Thumbnail. The first thing people see in the marketplace. If you don't upload one, Edgaze makes a clean title card from your title and description, which is what I kept. If you upload your own: 16:9, ideally 1280 × 720, up to 10 MB.
- Example output images. Up to six images of what your workflow produces, shown on your listing. A screenshot of a real result does more than another sentence of description ever will.
Demo inputs
Every listing gets a free demo, so people can try it before paying. This step decides what the demo form starts with.
Fill a field and that's what visitors see. Leave it empty and Edgaze writes a realistic example (marked "AI example"), and it never changes a field you filled. I filled both with the procrastination talk and 10 minutes, so the first thing anyone sees is a run I know works.
Review
One last look at everything. Each line has an Edit button that jumps back to that step. Below it are the Terms of Service, Creator Terms and Acceptable Use Policy you're agreeing to by publishing. Then Publish, and it's live.
Step 10: Your product page
Here's what a stranger sees when they open the link:
They get a free first run, then a fixed per-run price. Off to the side are buttons to Remix it, call it through the API, or use it from an AI assistant over MCP. That's the same workflow, sold three more ways, and you didn't do anything extra.
Scroll down and there's your title card, description and tags, plus a Q&A thread for buyer questions:
Updating a published workflow
Remember the Arabic subtitles? I found them after publishing, so version 1 went out reading Arabic. Fixing a live workflow is easy:
- Open Workflow Studio and click your workflow under Your workflows on the launcher. Not the draft with the same name under Continue: the published one. (Ask me how I know.)
- Make your change.
- Hit Update workflow listing.
The publish flow opens in update mode and says Updating to v2 at the top.
Write a release note. Mine says "Reads English captions now. Version 1 had been reading the Arabic ones." Honesty is a feature. Then Update, and everyone who opens the link gets the new version.
How per-run pricing actually works
This is the part I'd want someone to explain to me properly, so here it is.
The formula
Every run has one price, and buyers see it before they pay:
Price per run = your margin + compute
- Your margin is what you charge for the workflow: anything from $0.05 to $50 per run. The default is $0.50.
- Compute is what the run costs Edgaze to execute: the model calls, the transcript fetch, the infrastructure. Edgaze estimates it from your workflow and you can't edit it. It's not profit for anyone; it's the bill.
- You keep 80% of your margin. Edgaze keeps 20%. That's the whole split. You don't pay for compute out of your share, because the buyer covers it in the price.
For Should I Watch This?: $0.20 margin + about $0.13 compute = $0.33 per run, and I earn $0.16 every time someone runs it.
How compute is priced
A ten-second clip and a three-hour podcast cost very different amounts to process. Edgaze doesn't charge buyers a surprise price after the fact. Buyers see one price before they run and pay exactly that. Inputs above your maximum allowed input (the "Input limits" setting on the Pricing step; mine is 100k characters per input) are refused before a run starts.
Two consequences:
- Set your input limit to what your workflow genuinely needs. It is the ceiling on what one run can cost.
- Your margin never moves unless you change it. People who bought runs in a pack keep the price they paid.
Why the compute price can change after you publish
The publish flow shows a starting compute price. After you publish, Edgaze measures what each run actually uses and adjusts compute automatically at current model rates, so the price follows what the workflow really costs to run. It also follows model providers when they change their prices. Before you publish a paid workflow for the first time, you confirm you understand this. The product page always shows the current price.
What should you charge?
Some honest guidance, from staring at a lot of listings:
- Price the outcome, not the effort. Nobody cares that your prompt took three hours to write. They care that you saved them 40 minutes. Should I Watch This? saves people the length of a video. Anything under a dollar feels like a bargain for that, and that's exactly the feeling you want.
- Small, frequent jobs want small prices. A tool someone runs five times a week should feel like an impulse, not a decision. $0.20–$0.50 margins live here. Big, rare jobs, like a full report or a pile of bank statements reconciled, can carry a few dollars.
- Look around the marketplace. The "Up next" column on any product page is a free market survey. Comparable workflows near mine sit between $0.08 and $0.75 per run, with heavy-duty ones over $2. Land somewhere that makes sense next to them.
- Do the boring maths once. At $0.20 margin I earn $0.16 a run. 100 runs a day is $16 a day, about $480 a month, from a workflow that took twenty minutes to build. 10 runs a day is $48 a month. Pick a price you'd be happy with at the realistic number, not the dream one.
- The free first run is your best salesperson. Everyone gets one free run. Make it count: fill the demo inputs with an example that shows off the result (Step 9, Demo inputs).
- You can always change it. Publish a new version with a new price. Don't let pricing keep you from shipping.
The template: Should I Watch This?
Here's the whole workflow, so you can rebuild it in five minutes (or hand this section straight to Composer).
Blocks, in order:
| # | Block | Type | Key settings |
|---|---|---|---|
| 1 | Available Time | Input (number) | Question: How many minutes do you have? · Description: Enter the time you can spend watching. |
| 2 | Video Transcript | Get Content | Source: YouTube video · URL source: Ask the user when this runs · Question: Which YouTube video should we assess? · Caption language: en · Transcript style: Timestamped lines |
| 3 | Viewing Context | Merge | Inputs: blocks 1 and 2 · Mode: concatenate text |
| 4 | Watch Verdict | Generate Text | Model: GPT 6 Sol · Prompt and style below |
| 5 | Video Recommendation | Output | Connected from block 4 |
Watch Verdict: prompt
Assess whether the YouTube video is worth the viewer's available time, using only the timestamped transcript and the minutes available in the inputs above. Return exactly one verdict: Watch, Skim, or Skip. Watch means the full video is worthwhile within the available time, Skim means a few segments offer value, and Skip means the likely value does not justify the time. Do not assume the video's total duration unless the transcript establishes it. Give exactly three honest, specific reasons grounded in the transcript, including limitations or uncertainty where relevant. Then identify exactly three best moments with timestamps copied from or reliably derived from the supplied transcript, each with a concise description of why it matters. Prefer timestamps within the viewer's time budget when possible, but do not claim that jumping to a late moment fits the budget. Format the response with headings Verdict, Three reasons, and Three best moments. If the transcript lacks usable timestamps or has fewer than three identifiable moments, say so explicitly rather than inventing timestamps or moments.Watch Verdict: style / system
You're the friend who has watched far too much YouTube and has opinions about it. Be warm, direct and a little funny. No hype and no 'in today's fast-paced world'. If a video is mostly padding, say so kindly, then point to the part that isn't.Listing settings I used: Paid per run · $0.20 margin · Edgaze hosted · model failover on · remixing on · category video · demo inputs filled.
Or skip the rebuild and run the real one: Should I Watch This?. The first run is free, obviously.
Things I'd tell myself before my first publish
Check Builder view before you trust a result
Name it after the question, not the mechanism
Fill your demo inputs
Don't overthink the price
Keep remixing on
Your turn
The pattern here, take an input, fetch something real, judge it, return a decision, covers a ridiculous number of useful products:
- Podcast link → "which 10 minutes should I listen to?"
- Article URL → "is this worth reading, and what's the one idea?"
- Job posting → "should I apply, and what will they ask me?"
- Long email thread → "what do I actually need to reply to?"
Same five blocks. Different prompts.
If you build something with this, send it over. The good ones get featured, and the great ones get an unreasonably long blog post.
Build yours
Open Workflow Studio, describe what you want, and publish it the same afternoon.



