In the event you’re feeling overwhelmed by AI proper now, you’re not alone.
On daily basis it appears there’s a new device, new MCP, new mannequin, new ability, new workflow, new characteristic, new social submit that’s some type of “Hey look! I’ve fully discovered AI with this one bizarre immediate.”
I…don’t consider you.
I work with AI each single day, and what I’m discovering is that much less is far more. It’s not about what I set up or configure or trick the agent into doing that makes any actual distinction. That stuff is attention-grabbing, however on the finish of the day it seems like gimmicks.
I see the most important positive factors in my productiveness from how I take advantage of the harness and the way effectively I perceive it.
So on this submit, I’m sharing you a easy workflow that you should utilize to drastically enhance your effectiveness with AI simply by utilizing current options of GitHub Copilot. No bizarre prompts. No ability everybody else appears to learn about. Simply the harness. The harness is all you want—principally.
Disclaimers
I’m utilizing the time period “harness” interchangeably with “GitHub Copilot.” The purpose of this submit is to maintain issues easy, so simply know that GitHub Copilot is an agent harness.
I don’t imply to insinuate that you simply received’t ever want any abilities or MCPs or directions or customized brokers, and so on. In truth, these issues will develop into fairly vital as you progress and must outline advanced workflows and automate issues to your groups. In truth, I take advantage of a couple of all through this weblog submit!
What I’m declaring right here is that you do not want any of these issues to be extremely profitable with AI.
Additionally, there may be lots of slop on the market. In the event you don’t consider that, ask the agent to create a ability to do something in any respect. It would fortunately oblige. Whether or not or not that generated ability truly works, it may be simply printed to any variety of ability or MCP registries.
That is an apparent one, proper? Decide a device! It’s really easy!
However even throughout the GitHub Copilot household, there are lots of choices. These embody the CLI, the brand new GitHub Copilot app, VS Code, Visible Studio, and JetBrains, simply to call a couple of.
The excellent news is that these experiences are more and more being centralized on the identical harness. The small print can differ by device, however the core workflow is constant. Be taught the harness as soon as, use it in all places.
That mentioned, I do consider that studying the harness is essential, and one of the best ways to be taught it’s to be as near it as doable. So in case you are simply beginning out, I’d suggest starting with the GitHub Copilot CLI. It’s a terminal interface, which implies it’s simply textual content. There isn’t a lot UI to be taught. You enter a immediate. The agent does issues. However the interplay is extra direct, instant, and, frankly, very satisfying.
For this demonstration, I’ll be utilizing the brand new GitHub Copilot app. However the harness that app makes use of is the very same factor you’ll be utilizing in case you are utilizing the GitHub Copilot CLI, Visible Studio Code and lots of different locations yow will discover GitHub Copilot.
2. Activate YOLO mode
YOLO mode is also referred to as “Permit All.” This lets the agent execute any command with out asking permission. This may range relying on the device you might be utilizing, however for many it’s merely an /allow-all command within the chat. In any other case, the agent goes to cease and wait to your approval each single time it must do some work.
Brokers want autonomy so that you can see a rise in productiveness. If it’s important to approve every little thing the agent does, you would possibly as effectively simply do it your self. Moreover, that’s a depressing consumer expertise. No one needs to be relegated to sitting at a desk urgent the “Approve” button all day. And urgent “Approve” time and again simply trains you to not learn what you might be being requested to approve, which defeats the aim.
You need to be protected with brokers, although. Unhealthy issues occur to good folks. When utilizing YOLO mode, you don’t need to run the agent in your native machine. That is very true when you find yourself utilizing them at work—knowledge is personal in your group’s methods, and errors might be expensive.
Thankfully there are a bunch of choices for operating brokers in sandboxes. A straightforward one to get began with is GitHub Codespaces or improvement containers.
3. Begin with a prototype
One of the vital magical issues about AI is you could simply prototype something and every little thing up entrance. Traditionally, this was not the case. Prototyping was a full part of a undertaking, and had been usually a luxurious. Now, you may make one with a immediate.
Let’s have a look at a couple of examples.
Let’s say we need to construct a date picker net part. That appears straighforward, but it surely’s truly fairly advanced. Consider all of the totally different belongings you would possibly need to do with it.
How do you navigate throughout the part?
What does the chosen date seem like?
What does a particular vary seem like?
How does the consumer navigate between days, months, and years?
Begin with a easy prototype and get a number of variations. I often begin with one thing like this:
On this case, the AI generated a bunch of various layouts, however considered one of them is a mock the place it begins with the 12 months view. That’s attention-grabbing. I would really like my date picker to allow the consumer to zoom out to the 12 months, then into the month, and at last to the day. These are the sorts of belongings you don’t take into account till you see them.
As people, we course of sensory-rich fashions like photographs, shapes, and tangible layouts a lot sooner than dense textual content. Creating low-effort prototypes early on helps make advanced ideas instantly intuitive.
And this is applicable to non-visual duties as effectively.
As an example, if I need to add a brand new API endpoint, I’ll nonetheless create a visible prototype to grasp the necessities and constraints earlier than diving into the implementation.
For the reason that GitHub Copilot app helps Mermaid diagrams, the agent renders this as Markdown, mapping out 5 other ways we might implement this API endpoint.
When working with brokers, it’s simple to overlook that every little thing is nuanced. Prototyping helps uncover the nuances up entrance, so that you keep away from spending invaluable time and tokens on rework.
I like to recommend utilizing a medium-sized mannequin, comparable to GPT 5.6 Terra or Claude Sonnet, on medium reasoning for many work. I additionally suggest you follow no matter mannequin you select right here at some stage in this specific characteristic, bug, or enhancement. Immediate caching will prevent tokens. So long as you don’t swap to a unique mannequin or reasoning stage, your earlier chats stay cached with the mannequin, providing you with a reduction on future requests.
4. Plan methodically
Now that you understand what you truly need versus what you initially thought you wished, it’s time to plan out the implementation.
Swap to plan mode in GitHub Copilot with out beginning a brand new session.
“/plan Construct a date picker net part. I would like the consumer to have the ability to zoom out and in of years, months, and days.”
That’s a reasonably obscure immediate, and also you’ll probably have extra context for the mannequin than I do right here, however that is only a demonstration. In the event you don’t have extra context, it’s OK. That’s precisely what this step is for.
In principle, you will get a mannequin to one-shot something for those who compose the right immediate with the right context within the good order. In principle.
However none of us can do this. Planning helps you get nearer to that ultimate, although, by asking the entire questions that you’d must reply your self alongside the way in which for those who had been to construct this out by hand:
Can the beginning and finish date be the identical?
Are partial alternatives legitimate?
Ought to customers be capable to clear the date?
Ought to “at this time” at all times be a visual choice?
Is handbook entry allowed?
What format is the date saved in?
Ought to pasting in dates be allowed?
The listing goes on and on. You can’t presumably consider all of those edge circumstances, however the mannequin might help you establish lots of them.
You may make plan mode much more aggressive within the sheer variety of questions and edge circumstances it asks about by putting in the “grill-me” ability from Matt Pocock.
This planning step is essential. The purpose will not be so that you can simply settle for each suggestion from the AI. In the event you do this, you might be negating the worth of this planning course of. The purpose is so that you can deeply have interaction with the issue and information the mannequin. That is the place your experience comes into play.
You can too ask the mannequin questions again. Within the screenshot beneath, it asks me about “non-contiguous dates.” I’m fairly positive I do know what the mannequin means right here, however I’m going to ask for clarification so we’re on the identical web page.
The planning course of will preserve going even for those who interrupt to ask clarifying questions, and so on.
5. Implement with Autopilot
As soon as the plan is completed, GitHub Copilot will probably immediate you to modify to Autopilot and begin implementing the plan.
Autopilot is a built-in loop. It forces the mannequin to proceed working by guaranteeing that it has truly accomplished what it mentioned it will do—which on this case is finishing each merchandise within the plan.
GitHub Copilot will robotically act as an orchestrator throughout this part. If it must learn recordsdata within the codebase, it is going to use the “Discover” subagent with a small mannequin. If it deems an motion comparatively advanced, it is going to probably select the “Common Function” subagent with a bigger mannequin. Whereas you will get fine-grained management over orchestration in GitHub Copilot with customized brokers and directions, you don’t must do something particular to get some great benefits of subagents and multimodel workflows. This works out of the field, even for those who didn’t know that any of these items existed.
6. Human evaluation and iteration
That is the place you get your dopamine hit. You get to see what the AI has created.
But it surely’s probably that you simply received’t get precisely what you wished. That’s regular and anticipated. The mannequin can’t learn your thoughts, and it’s error-prone. Iterate with the mannequin till you get what you truly need. Whether or not that’s simply code or an improved UI, that is the half the place your style will resolve the standard of the ultimate product.
As an example, right here’s the date picker that GitHub Copilot gave me.
Already I can see it has some points:
Animations are inconsistent
Textual content is unreadable when hovering over a particular date due to coloration distinction
It doesn’t must say “12 YEARS” on the high.
Once I click on “In the present day”, it doesn’t take me to the day if I’m within the month or 12 months view.
Additionally, I don’t love the design. It appears to be like just a little an excessive amount of prefer it was created by AI—as a result of it was!
So right here we’re simply in follow-up mode. I’m going to make use of a CSS framework I created known as Postrboard. I add it as a ability that simply factors to the CSS and tells the agent tips on how to use it. You possibly can be at liberty to put in it your self for those who’d like to make use of it, or you’ll be able to choose another CSS framework on the market that you simply like. Giving the mannequin some design steerage is kind of useful, and infrequently a CSS framework is all you want.
For the date picker, after I click on on the day, it tries to zoom in, however can’t as a result of there may be nothing to zoom to. There ought to be no zoom there.
It doesn’t must say “Zoom Out” on the high
Once I mouse over a month or 12 months that incorporates the chosen day, I can’t learn the hover textual content.
Once I click on “In the present day” it ought to take me to that day view, even when I’m on the month or the 12 months.
The months don’t want numbers beneath them and so they don’t should be in containers
Similar goes for years. And it doesn’t must say “12 years” on the high.”
Discover how conversational that is. Don’t overthink it. Once you’re fixing a bunch of small issues like this, simply give it to the mannequin. In the event you’ve obtained the context, you’ve obtained the immediate.
Crucial factor is to not accept AI output that’s “adequate.” Insist on high quality. Be ruthless about it. That half remains to be your accountability, and understanding what a high quality result’s from one thing that isn’t is the worth that you simply deliver. No AI will ever exchange your human contact and creativity.
Right here’s what my last date picker appears to be like like. Scroll to the top of this submit to see it in motion.
7. Rubber duck the end result
After you’ve iterated and are pleased with what you’ve created, it’s time to do a last evaluation.
Request a Rubber Duck evaluation from GitHub Copilot. You are able to do this simply by asking for it:
In a Rubber Duck evaluation, GitHub Copilot will request a evaluation from a mannequin of a unique AI household. As an example, since I used to be utilizing GPT 5.6 Terra, it requested a evaluation from Sonnet. Totally different fashions had been skilled on totally different knowledge, so that they have totally different blind spots. A Rubber Duck evaluation helps establish potential points that could be missed by a single mannequin.
Be aware that you should utilize this at any level on this workflow. You possibly can rubber duck prototypes. You possibly can rubber duck plans. All of it simply is determined by if you would like a second AI evaluation on one thing.
And if you wish to take this a step additional, you’ll be able to mix rubber duck with Autopilot to get the fashions to work collectively in a loop to enhance the ultimate end result.
“/autopilot rubber duck this date picker implementation. When you’ve the end result, evaluation it fastidiously and make any mandatory changes. Repeat the rubber duck evaluation till each you and the reviewing mannequin agree that the one objects that stay have diminishing returns.”
After this step, you should have an much more refined end result than earlier than and could have probably recognized many further edge circumstances. This step does value extra tokens, however you might be actually battle-hardening the code. Consider it as an funding in your future self who received’t need to cope with these points since you caught them now.
8. Revenue
At this level, you’re able to stage and commit, or transfer on to the following characteristic you need to add together with this pull request.
I’d suggest beginning a brand new chat session for something you do subsequent that doesn’t need to do with this date picker. You possibly can consider chat classes as being topical; for those who begin to diverge an excessive amount of from the principle matter, it’s most likely time for a brand new session.
Right here’s the ultimate end result from my workflow constructing the date picker for this submit.
I notice that this can be a little bit of a contrived instance, however can all of us simply pause for a second and marvel at what we’re in a position to pull off with AI now? Constructing a date picker was once one of many hardest issues you can attempt to do. Simply ask any of the heroes on the market who’ve constructed them.
Issues don’t need to be difficult
This easy workflow might be sufficient for most individuals. The simplicity additionally helps you multitask. It’s simpler to cause about what agent is in what state and what you had been doing final if you preserve issues easy. Your context window is restricted too.
There’s a lot occurring within the AI house proper now. There isn’t a higher restrict on the issues you could construct and experiment with. You possibly can add MCP servers, abilities, directions, and customized brokers. You possibly can arrange workflows and loops, create brokers that immediate brokers, and get up whole digital dev groups.
However remember that no one actually is aware of what they’re doing proper now. We’re all figuring this out as we go. Quite a lot of what’s at this time’s magical incantation for AI might be tomorrow’s anti-pattern.
Simply give attention to getting a repeatable, high-quality end result within the easiest method you could. Be taught the harness and also you’ll be simply nice.
Written by
