Affiliate disclosure: this post contains an affiliate link to Tube Magic. If you subscribe through it, I earn a commission at no extra cost to you. I have not bought the tool — this series checks tools against my actual workflow instead of pretending to review them. Everything else listed here is what I actually run, with no affiliate relationship.

Search "youtube automation tools" and you'll get lists of forty apps sorted by nothing. Tools aren't interchangeable — each one does a job. Here is the stack behind my daily Shorts pipeline, organized by job, including the jobs I deliberately don't give to a tool at all.

This is part 3 of a series: part 1 mapped where an episode spends its time, and part 2 drew the line between what I automate and what I refuse to.

Job 1: Research & picking the material

What I run: my own reading, plus an agent that fetches and files source texts.
Tool verdict: no dedicated tool. My material is a fixed shelf of classic books; the "research" is judgment about what a 40-second video can honestly carry. Topic-hunting tools solve a problem I don't have.

Job 2: Script draft → final script

What I run: a general LLM chat for structure, then I write the final ~90 words myself.
Tool verdict: a script generator would slot in here, but at 90 words the gap between "draft from a chat" and "draft from a dedicated tool" is too small to pay for.

Job 3: Metadata drafts (titles, thumbnails, tags)

What I run: LLM first drafts + my rewrites. This is my highest-frequency rewrite loop.
Tool verdict: this is the one job where I'm actively checking a dedicated candidate — Tube Magic (affiliate link: https://tubemagic.com/ds#aff=jpyun1999d9bf&cam=), because titles/thumbnails/tags are its core, not a side feature. Still a check, not a purchase: its marketplace numbers ($47/mo, 8.94% cancellation) are sane for the category, but sane numbers aren't usage.

Job 4: Assembly & rendering

What I run: a scripted rendering pipeline (cards + audio → vertical file), executed by agents, with the publish step held for a human check.
Tool verdict: solved. If your assembly is manual, almost any editor with templates beats none — this job is mechanics, and mechanics should be boring.

Job 5: Publishing & bookkeeping

What I run: upload checklist + agent-maintained log of what shipped and what it did.
Tool verdict: scheduling tools are fine; the log matters more than the scheduler. You can't fix a bottleneck you didn't record.

The pattern

Every "AI stack" post I read sorts tools by logo. Sort yours by job instead: write down the five jobs, mark which ones are actually costing you time, and only then look at tools — one job at a time. Half the tools on those listicles will eliminate themselves.

Which job eats your time? For me it was Job 3 by a wide margin — I'm curious whether that holds for other small channels.