How I Automated 90% of My Content Workflow (With ChatGPT + Riverside)

Peter Yang introduces a system he says saves at least five hours a week and automates much of his podcast workflow. The demonstration is more specific than an unattended content factory: he repeatedly changes drafts, chooses clips and reviews what will be scheduled.

The percentages and time savings are his estimates, not measurements independently reproduced here. He also discloses Riverside as the tutorial's sponsor and uses it throughout the example. That commercial context belongs alongside the workflow description.

The episode used as an example concerns personal-finance software, but this article is about media production. It does not reproduce financial instructions or invent the exact prompts that are not given in the source transcript.

Yang lists fifteen steps rather than beginning with a tool purchase. They range from guest research and an interview guide to editor instructions, final transcripts, packaging, descriptions, social posts and scheduled clips.

01:52View full image

The list reveals three phases: preparing the conversation, coordinating the edit and producing the materials around the finished episode. Asking an AI which tasks it can take on is then a discussion grounded in an actual process, rather than a vague request to automate everything.

An inventory should include small handoffs and checks too. If they disappear from the map, the resulting system may produce a draft while leaving someone to rediscover the missing operational work.

Yang creates a preparation skill, an editing skill and a production skill. Production is an orchestrator that calls narrower skills for a title and thumbnail, show notes, a newsletter, social material and clips.

That structure gives each step an identifiable purpose and artifact. A proposed interview guide is different from an editor's package; a final-video description is different from a draft based on a raw recording.

The skill text was largely developed with AI from his workflow instructions. He presents it as a process he specified, not a requirement that a user write every line of a large instruction file unaided.

The preparation example gathers guest information, links and comparable videos, then proposes an interview goal and possible packaging. Yang uses that early work to ask whether the subject can make an interesting episode.

He rejects some of the initial emphasis and asks for more practical use cases. The guide is revised into questions, supporting notes and demonstrations he wants to see, then placed in a document for review.

The important action is the correction. Research is not useful merely because links were gathered; it has to support the actual conversation the host wants to have. Suggested answers are preparation material, not a license to attribute unspoken statements to the guest.

After recording, Yang uses the recording tool's connection to create instructions in a Linear ticket. The package includes the proposed title and thumbnail, an opening reel, cut notes and clip candidates.

The agent suggests two versions of the opening from the written transcript. Yang combines or changes them and adjusts where clips begin and end. The ticket then records the selected quotations, timestamps, logistical cuts and accompanying copy.

This is a transcript-based selection workflow. It can reduce searching, but the demonstration does not prove that text alone captures every visual issue or performance detail in the recording. The editor and host still have to verify the chosen moment and the finished result.

For the newsletter, Yang asks for a practical draft centered on the guest's examples. The first version gets important prompt details wrong. He supplies the exact six prompts from the conversation and asks for them to be incorporated instead.

The title-and-thumbnail skill examines the episode and comparable videos, then offers combinations. Yang changes the options and says he still spends substantial time choosing the package. The planned A/B test is not reported as a completed winning result.

This preserves a distinction between making options cheap and making the final judgment automatic. A reusable template can accelerate execution while leaving a difficult editorial choice unresolved.

The show-notes step uses the uploaded final video rather than assume raw-recording timings remain correct. Yang asks for named use cases in the chapters and checks links and descriptions through further iterations.

After agreeing on the material, he demonstrates using the browser to place it in the video description and handle associated settings. Those actions occur after a chosen output, not simply because a draft exists.

The general handoff is useful beyond a podcast: identify which version of the source an artifact refers to, verify the details and make the destination explicit. Tool access and authorization still determine what a particular setup can actually perform.

Yang's feedback loop begins before he edits the newsletter manually. He asks for a snapshot of the AI draft, makes his changes, then has the system compare the two versions.

13:50View full image

He asks for numbered proposed changes to the skill so he can accept or decline each one. He reads them, gives feedback and only then tells the system to update the instructions. That separates observing an edit from assuming every local preference should become a permanent rule.

The result is explicit maintenance, not automatic learning hidden inside the model. It also supplies an artifact someone can inspect when the workflow behaves differently later.

For social material, Yang supplies examples of his previous writing and asks for alternatives. He revises the opening and chooses an image before reviewing the scheduled teaser.

The clip step likewise starts with suggestions, then a chosen moment. He checks the edited clip, removes pauses and unnecessary words, reviews the accompanying post and approves scheduling through his social tool.

That sequence matters more than the exact app names. Candidate generation, source verification, editing and the decision to schedule are distinct stages. A successful draft in one stage does not establish that the entire publication has been completed correctly.

At the end, Yang acknowledges that even this system leaves him spending considerable time on production. He considers whether another person should help operate the workflow rather than continuing to do everything himself.

His suggestion is to begin with one or two recurring manual workflows, map their steps, create phase skills and connect them. More integrations can make additional actions possible, but do not automatically make the content accurate or the process effortless.

The demonstrated improvement is a set of more reusable handoffs and review loops. Its usefulness depends on keeping source details, human decisions and the actual state of the destination visible, so that saved effort does not become unnoticed editorial error.