🦄 So, you’ve flipped the switches, you’ve run a prompt (or ten), and now you’re staring at a Draft PR with an “AI” badge and a rainbow of checks. Perfect! This is the part where we stop hoping for magic and start running a clean, predictable review loop. 🪄
First—Enable the Cool Stuff 🧪
If you’re not already using GitHub’s Feature preview, fix that now. Head to GitHub, click your profile picture (upper-right), then Feature preview. You’ll see a modal like this:
Explore freely, but the one that matters most for Coding Agent reviews is New Files Changed Experience. It’s a serious upgrade for the diff view and PR flow. Sometimes the previews can be glitchy, but this one’s solid.
🦄 Do the civic thing and send the dev team feedback after you’ve used it—even a quick high-five goes a long way!
PR Reviews with Coding Agent 🧭
It’s the PR you already know, with a few Coding-Agent-specific quirks:
-
You are requested as the prompter (and anyone in
CODEOWNERSgets notified too). - The PR opens as a Draft so you can pre-approve before looping in the team.
- Copilot creates branches under the
copilot/prefix; fighting that is wasted energy. - GitHub supports multiple assignees now—expect to see yourself as a co-author/assignee.
- There’s a new right-side box summarizing Copilot’s workflow activity. Mostly a time-saver that eliminates an extra click to the Actions page.
Reminder: your prompt == your code. That’s not a slogan; that’s custody.
🦄 This is not “Copilot’s branch.” It’s yours. Pull it, tweak it, push, and re-prompt. Coding Agent will work from whatever you commit.
Seeing Errors? 🧯
Copilot fills the PR description with its results (usually a checklist; or you could ask for emojis... guilty 🙋♀️). Read that first. Does its conclusion match your intent? That’s your earliest signal of how things really went.
💡 ProTip: If results drift, don’t just patch locally and move on. Take ~20 minutes and quiz Copilot on its thought process from your IDE. Troubleshoot instructions, improve the prompt, then re-run. Iteration now saves you ten runs later. If it’s minor, skip it, fix the branch and keep going.
You might also see a firewall warning in the logs, which is the same kind of warning I mentioned was possible last week. Recently, there was a new Ubuntu runner hiccup that's causing firewall failures. You can fix that with a quick URL in the allow list or you can wait it out. It might already be a non-issue anyway. 🤷♀️
Where’s the “Magic”? 🪄
Short version: your Actions tab turns into a spellbook of what your prompt summoned—some charms pass, some fizzle, one asks for your approval like a polite familiar, and every now and then there's a glitter explosion. 🎉
From here I don’t chase every blinking glyph; I read the page like a logline.
Reverse-chronological lists are sneaky. The real story starts at the bottom, so that’s where I scroll first. The first red I hit is the root cause; everything above it is ripple. I usually delete the later, noisy runs so my eyes stop lying to me. That single red is my anchor. Work from there.
The yellow job with an Approve and run button—that’s GitHub’s protection gate. If the gate guards something I don’t need yet (say, a docs publish or a workflow that's probably not going to pass), I delete it straightaway. If it matters (lint, build, tests), I'll click through and approve. Treat approvals as wards enforced by repo rules, not as shame flags for the AI.
The “magic” doesn't need a wand, just a cadence:
prompt → runs → read from the first red → fix what actually broke → re-prompt with the thing I learned.
💡 ProTip: Beware of the "glitter explosion". Sometimes it's less colorful and more widespread sprinkles of every single file in your repo just changed. 😳
Demystifying Coding Agent 🧚♀️
This is one of my favorite Coding Agent touches: when the timeline gets fuzzy, open the Copilot workflow run and read the logs. The full, step-by-step output you’d expect in chat is right there in Actions—clear footprints from prompt to commit.
💡 ProTip: Open the run → expand the failing step → start at the first sign of red. Fastest path from mystery to fix. 🧭
Reviewing Coding Agent’s PR 🗂️
This is where the preview shines. Switch to Files changed—that’s the doorway to batching.
- Click Start a review and add line comments as Pending.
- After you Submit review with one or more comments including the
@copilotprompt, Copilot treats the whole batch as one session / premium request. It only starts acting once comments are submitted, not while they’re pending.
If you instead post on the Conversation tab with @copilot, each submitted top-level thread is queued up one after another. Either way, the agent will push commits back to the same draft PR as it addresses each comment.
I leave myself notes in the PR, I talk to @copilot where it makes sense, batching in Files changed when I want one premium run; splitting threads in Conversation for simpler asks. Sometimes I pull it into VS Code, make the edits, and send it back to Coding Agent. That’s the difference between fireworks and a repeatable ritual.
💡 ProTip: Concurrency is per-PR; you can’t run two Coding Agent executions at the same time on a single PR. For real parallel work, use this
copilot/branch as the base branch in a new prompt. Then merge the results here before going tomain.
Not Loving the Preview? 🧰
You can disable New Files Changed Experience in Feature preview if it’s not clicking. Once it graduates to GA, the toggle disappears—so if you truly dislike it, say so in feedback while the window’s open.
🦄 My rule: don’t complain unless you’re also contributing to the fix.
Sounds Great… But Now What? 🏁
Start small. Practice prompting on low-risk tasks. Coding Agent can generate any docs you need: README refreshes, contribution guides, tech overviews, user guides (ask for screenshots), systems diagrams (ask for Mermaid), and Gherkin features. Yes, I’ve tested them; yes, it works. Next week I’ll share the exact prompt patterns I lean on and how I keep runs aligned (and cheaper).
AIs Are Benched 💤
ChatGPT has been in time-out all week. Still, it nailed a color scheme and came up with a couple of killer playlist names, so I let it proofread. This is by far the most hand-written post I've made to date. Exhibit A for the thesis: AI edits. The voice is mine. 😇
We do need to have a talk with OpenAI about butchering the poor guy’s arc though. I can’t deal with it like this. 🤨
Top comments (0)