Writing has always been a passion for me, but with everything else going on there is not much time left to sit on my desk, grab a cup of coffee, and just write. Writing became a work activity, not something I feel like doing when I have some free time. But the thinking never stops, and sometimes I feel like I have something valuable to share.

Writing on the Go

So one of these days I tried writing on my notes app while I’m on the treadmill, at my gym. That worked surprisingly well, so I used some of my gym sessions to write down LinkedIn posts and random ideas for videos. I also used this time to edit videos from my phone.

Eventually I wrote way too much and ended up finding out that LinkedIn has a character limit for posts. “What a shame”, I thought, as I found it difficult to cut down on words and I really wanted to share that piece of content on its entirety.

“What about my blog?”, I thought. That would be great, but from the gym I cannot really post anything to my Git-based static blog. Or can I? 🤔

Copilot to the Rescue

“Maybe Copilot can help me. I already have the content, after all!” - so it would be only an operational thing: create a pull request on my blog repo including the content I would provide it, and then I could preview, review, and merge the PR directly from the treadmill.

I went to the “Agents” tab of my repo using the GitHub mobile app. My first prompt was:

Bootstrap a new blog post with the title “Use the Right Tool for the Job and your Skillset” and create a pull request so that I can add the content. Tags: “career”, “development”

Then after a little while I got a notification from an open PR on my repo. The PR had exactly what I asked for: a markdown skeleton I could edit to include my text. I was able to finish the post from my phone, but it was still quite a bit of work. This was the first published post.

A couple days later I improved the process and instead of using the “Agents” tab, I created an issue with the content already in a code block, with instructions to create a blog post using that content. The, I assigned the issue to Copilot, and a similar process followed, but this time the PR already contained everything. This was the result from the original issue.

Turning Videos into Quickstart Tutorials

Another blogging struggle I had was around turning a video tutorial into a text version. Writing a detailed tutorial for things like 3D design software is lot of work, as I’d need to show the interface with screenshots and all. Making a video is much easier, and takes me much less time. But I understand that some people prefer to follow along with a text version. So I picked a video transcript from Google Drive (auto-generated when you upload a video there) and asked Copilot to create a step-by-step summary of the video and include the embed in a blog post. It also generated a short intro and conclusion, which was fine. You can see the result here and here. I was actually pretty happy with this use case, because I could focus on creating the videos and still make them acessible. Also, I wasn’t using any AI to author the original content, so I felt good about that too.

Migrating to Copilot Agentic Workflows

I was trying to figure out the best way to handle this automation without having to repeat the same prompt over and over, and that was around the same time I heard about Copilot Agentic Workflows for the first time.

Luckily for me, as part of the GitHub Stars program I get to join some meetings with GitHub engineers where they share the latest features and things that are not yet in general availability. Shortly after one of these meetings I started my experiments to create Agentic Workflows with the goal of streamlining my blog publishing process, and here’s where I am today: this post was 90% made from my phone while on the gym. Copilot created the PR for me, and I am just finishing the other 10% with some links to demonstrate the process. The best thing is that the “Blog Post Creator” workflow was also created by Copilot through the “Agents” tab!

After I got the Blog Post Creator workflow up and running, I set up another workflow to deal with the video tutorials, this time triggered by the label youtube-todo. Instead of creating a PR directly (which would sort of duplicate the process), I asked the agent to prepare the content and create a new issue with the content-todo label, in order to trigger the “Blog Post Creator” workflow. So, basically I:

• create an issue providing a YouTube URL and the transcript of the video in a code block, label with youtube-todo
• the “Youtube to Blog Post Creator” workflow starts working on the issue, generates a blog post, then creates a new issue to trigger the “Blog Post Creator” workflow (label content-todo )
• the “Blog Post Creator” workflow is triggered and proceeds to create a PR with the new blog post.

If the content is just text, I create an issue to trigger the Blog Post Creator workflow directly.

The Prompt

The “Blog Post Creator” workflow was created based on the following prompt:

Create a workflow for GitHub Agentic Workflows using https://raw.githubusercontent.com/github/gh-aw/main/create.md

The purpose of this workflow is to create a PR with a new blog post using information provided in a GitHub issue. The workflow should be triggered when an open issue is labeled with the “content-todo” label. Based on the content provided, create a SEO friendly title and description. Check the tags already used in the blog and tag the post apropriately with a maximum of 3 tags. Do not include any additional content in the body of text.

Took only a couple minutes to get the PR ready for review with the new workflow. After some tweaking, I got it up and running.

One important thing to notice is that Agentic workflows are actually markdown (.md) files with a YAML frontmatter that defines some metadata, permissions, and rules for triggering the workflow. You’ll then run a gh aw compile command (or instruct the agent to do so, when needed) to “compile” the agentic workflow markdown and generate a lock file that is the actual workflow that will be executed. That will be a much longer file with several instructions and guardrails in place. For reference, you can have a look at the agentic markdown file and its generated lock file.

Workflow Setup

To make everything work, I needed to:

• Update a few settings in the repository to allow GitHub Actions to create new pull requests (Settings -> Actions -> General -> Workflow Permissions, then check the box for “Allow GitHub Actions to create and approve pull requests”).
• Create a fine-grained PAT (personal access token) to authenticate to Copilot CLI from the workflow.
• Update Copilot security settings (Settings -> Copilot -> Agent -> Custom Allowlist) to allow requests to https://api.github.com/repos/github/gh-aw, so that the agent is able to fetch the gh aw extension, which is needed to compile the markdown workflow file and generate the lock file (the actual workflow that you’re not supposed to manually change).

Conclusion

Agentic Workflows are really handy for my use case since I don’t have much time to spend on the operational tasks required to get my ideas published here. I can focus on my original content and let Copilot handle the boring and time-consuming tasks without having to explicitly tell him what to do - just create an issue, add the content, and label it appropriately.

If you want to try out Copilot Agentic Workflows, check the excellent official docs to get started quickly.