you2idea@video:~$ watch 5Bmp9cz1JN8 [30:37]
// transcript — 465 segments
0:01 Oh, how the times have changed. I still remember the days where I would sit up
0:04 here and tell you guys all about these new technologies, frameworks, and ways
0:08 of building software. And now I'm about to do a whole video about three markdown
0:12 files. In some ways it feels silly, but these markdown files have meaningfully
0:15 changed how I built. I have been using these three skills a ton recently, and
0:19 I've been getting so many questions about them that I figured it was time to
0:22 finally sit down and explain what they do and how I use them. The skills are
0:29 sole loop, full send, and full stop. These aren't skills that I had some
0:32 super crazy clever concept for and decided to force into my work. It's
0:36 actually the opposite. It's behaviors I found myself doing again and again that
0:40 I turned into skills. When I'm working on T3 Code, almost half of my threads
0:45 end up using one of these three skills. And I would bet that the ones that end
0:49 up being merged into main almost always are using at least one of these three.
0:51 These skills have really helped me solidify some of the ways that I've been
0:55 building, and I hope I can help you guys build more effectively once I teach you
0:58 about them. But I got one more thing I want to tell you guys about that's been
1:01 helping me build too. Today's sponsor. Modern AI models have made it way too
1:04 easy to rewrite pretty much all the things that your business probably
1:07 relies on. We've been doing this a ton at my companies in order to trim our
1:10 costs and make things that better fit our needs. But I've noticed there's a
1:13 few things where as good as your custom rewrites can be, they're still going to
1:16 hit edges. What if I told you that one of these things is so annoying that even
1:20 OpenAI and Anthropic have chosen to use a vendor instead of rolling it
1:23 themselves? If I told you the thing was auth, you'd probably call me crazy.
1:26 Sadly, auth has driven me even crazier, and that's why I'm so thankful for
1:29 WorkOS. There's a reason everyone from OpenAI and Anthropic to Cursor are
1:33 relying on them for their auth layer. And it's not just because setting up
1:36 auth is hard. I'll be real, getting a simple sign-in with Google button is
1:39 relatively easy once you've gotten through all of Google's dashboards. The
1:41 problems start to show up with different types of customers. For example,
1:45 enterprises that are going to need a very specific set of things for their
1:49 weird SSO setups. The admin portal in WorkOS will take care of this for you.
1:52 You send the business a link and they can sign up however they want. Then one
1:54 of those customers is going to hit you up and say they want to set up an MCP to
1:57 connect with their agents. Setting up auth for that is nearly impossible.
2:01 Unless you're using WorkOS, then MCP auth is built in. But what if you want
2:03 to go further? What if you want your agents to be able to sign up for the
2:06 service themselves? That's when authmd comes in. This isn't just a product
2:10 WorkOS has. This is a standard that they built with companies like Cloudflare and
2:13 Firecrawl to make a good way for agents to sign up for a service on a user's
2:17 behalf. And as you can guess, there's a one-click setup if you're using authkit
2:21 with WorkOS. Get your auth right for enterprises and for your agents at
2:24 swiyft.link/workos. I think the easiest way to go through these is from the top, but they will
2:29 make a lot more sense when I give you a little bit of an overview on how I
2:33 manage all of my skills. I have a repo on my computer called fleet. This is the
2:37 only repo I find myself opening in VS Code almost ever. The reason is because
2:42 this is where I keep all of the data about my fleet of machines, the
2:46 different computers that I work on. And if you didn't know this already, with T3
2:50 Code, you can connect to many different machines and control them all from one
2:53 app or a phone app as well, which is very nice. But keeping everything synced
2:56 between those machines, like making sure they have auth set up, making sure they
3:00 have my skills, my MCPs, and all of that can be annoying. So I have this one repo
3:04 named fleet where I track all of this. I have folders that are categorized for
3:07 the different types of skills. I have some that I only like in Claude. I have
3:11 some that I only like on this particular machine, which I call command center.
3:14 And then I have some that I want pretty much everything to have. I'm probably
3:17 going to do a pass and delete a bunch of these cuz I have not been using as many
3:21 of them as I expected. I still love pulling out the Matt Pocock grilling
3:25 skill here and there, but I find that it is mostly just sole loop and false end,
3:30 false stop, and then all my PR-related things like file PR and babysit PR I use
3:35 a lot. Leaving PR comments is another useful one, too. I have a pretty simple
3:40 skill here where I tell my agents to call out what model they are and that
3:44 they are responding on behalf of me if they're using my account to leave
3:47 comments. None of that is what we're here to talk about, though. We're here
3:50 to talk about sole loop. One of the things I've said many times before, and
3:53 I still believe this even if I don't personally follow it as much as I
3:57 should, is that it is a good idea to write these things yourself. This one, I
4:02 let the agents take over a bit. Initially, I did write it by hand, but I
4:07 noticed a couple things that Claude was doing wrong, so I told it to update the
4:11 skill to fix the mistakes. And then I read it for the first time in a while
4:15 last night and realized it got quite bloated. So, I'm going to be going
4:18 through and cleaning this up in the near future, but for now, the thing that
4:22 matters is the core concept. One other important thing to know is the role of
4:27 descriptions. Descriptions are not what they sound like. This shouldn't describe
4:31 what the skill is. This is an indicator to the model on when it should be used.
4:36 So, most of my descriptions are really simple like this. They are effectively
4:40 triggers. I'm describing when the model should do this thing. So, the skill is
4:44 named Soul Loop and the description is use when the user says Soul Loop. So,
4:47 this should not be pulled in unless it was specifically mentioned by me or by
4:51 another skill. So, what is this skill? You own the code and Soul reviews it.
4:54 Fix what's right, push back on what's wrong, and stop when you agree or the
4:58 budget ends. Goal, get the smallest change that meets the intent. Reviews
5:01 make it correct, not bigger. This is some slop that it added. You can tell
5:05 that Opus wrote this. I I'm going to rewrite this myself now. I
5:09 have now rewritten this to be a better description of what it's supposed to be
5:13 used for. This is a workflow for making your code changes more correct before or
5:17 sometimes after filing a pull request. The TLDR is that you should use GPT-4
5:21 Soul to review your code in a loop. Make the most minimal possible changes to
5:25 appease the Soul reviewers. Be careful of scope creep and don't be scared to
5:29 push back. The be careful of scope creep and don't be scared to push back is very
5:32 important. One of the issues with a skill like this is that if you have too
5:35 many agents that have different opinions talking over each other, you could
5:39 quickly end up in a state where your scope blows up and the PR ends up
5:42 significantly larger than it needs to be. This simple addition can help
5:46 prevent a lot of that. I have a budget section saying how much this should go
5:49 cuz otherwise it'll just go forever. I say three rounds total. Round one is
5:52 full. Later rounds check only change code and open findings. Extend to five
5:56 only if there's security permission of data loss or other potential like risky
6:00 things. Stop early if the round returns only nits, decline nits and comments,
6:02 docs, or things that aren't as important. Yada yada. Well, only one
6:06 reviewer at a time. Soul should not start sub agents or its own reviewers,
6:10 otherwise this gets nested real real fast. And then keep track of the things
6:14 that it reports. Then there's a bit about the scope and also a section on
6:18 how to actually call Soul. There's a couple different ways to do this. As you
6:22 can guess, I'm mostly using this skill with Opus 5.5. I found that Opus 5.5 is
6:27 a much better collaborator, overall smarter, writes way better code, just
6:31 more pleasant to work with in general. GPT-61 Soul is more thorough, though. It
6:36 really does feel like the Rottweiler. It grabs onto the problem and doesn't let
6:40 go, and I really like using it for reviews for this reason. I have been
6:44 amazed at how much this skill and this workflow has prevented PRs from ending
6:48 up in an infinite review loop once they're up on GitHub. Having this here
6:52 running on your own machine helps a ton with like the first pass before you put
6:57 it up in a PR, so things like, in our case, macroscope, Grab Tile, Code
7:00 Rabbit, and other code reviewer AI tools, before they get it, is usually
7:04 solid. This is also great because you can take advantage of the subsidization
7:08 that you get for your Codex subs. And since I'm personally not using my Codex
7:12 sub a whole lot recently, this has been a great way to continue taking advantage
7:16 of it even though I've been doing my day-to-day work in Opus. One other
7:19 little thing I do with this that I found is pretty cool is I have a blurb at the
7:24 end for when the PR is filed or updated that calls out that it should have this
7:27 section added to the bottom, reviewed with Soul loop, N rounds with model. I'm
7:32 going to have uh another piece here, Soul verdict, approved. This is because I've noticed a
7:38 few times Opus almost wants to hide the fact that Soul wasn't happy with the
7:43 changes. Now, this will call that out directly. This is why it's important to
7:46 actually like edit these files yourself. On that note, I'm going to ship these
7:53 changes. Uh updated the Soul Loop skill. I'll just say fix formatting and push it
7:58 across the fleet. The thing I was going to say that I forgot is how you actually
8:03 give Claude Code access to Soul. There's a few different ways to do this. The
8:07 easiest way to do this in something like Claude Code is to set up CLI proxy in
8:11 order to have all of your auth in one place. I've been doing this for a while
8:14 now and have customized the hell out of my dashboard to see like how fast I'm
8:17 burning through my rate limits, what traffic's going where. I even went as
8:21 far as setting this up to dump all of my requests and responses into a ClickHouse
8:25 DB so that I can query them to get useful info. CLI proxy is one of my
8:29 favorite things I've done to my setup. I highly recommend forking and setting
8:32 something like this up yourself. And once you have that set up, you can use
8:36 all of the models that are accessible in it in whatever tools you want. Be
8:40 careful using the Claude models other places though. If you're using a Claude
8:44 Code sub, Anthropic is really upset if the system prompt isn't the Claude Code
8:48 system prompt. So, you should only use this with Claude Code if you're using
8:52 your Claude subs. I use my Codex subs with Codex. I use my Claude subs with
8:56 Claude, but I also use my Codex subs in other things like Claude through this
8:59 because OpenAI doesn't care where you use your tokens effectively. Anthropic
9:03 cares a lot. OpenAI doesn't. So, if you do set this up, be careful with that.
9:06 The reason I'm bringing it up here is it makes it very easy to set up Claude Code
9:11 to let you use OpenAI models in it. And if you're worried about getting banned
9:14 for this, don't worry. Tibo actually quote tweeted this saying that it is a
9:18 blessed solution and if this does get you blocked, then he owes you a reset.
9:23 So, this is allowed as far as they have officially said. So, if you really want
9:26 to stick with Claude Code but get these reviews in and also let it do its thing
9:30 its normal way inside of Claude Code with sub agents, this is a good solution
9:34 for that. And this was how I was doing it for a while up until the newest
9:38 update shipped on T3Code nightly, which is Orchestrator V2. Quick TLDR of
9:43 Orchestrator V2 is that it lets your agents and threads call other agents
9:46 harnesses threads and whatever else within T3Code. So, I'm going to ask in
9:49 this thread where I made a PR doing some changes to how we log data about the
9:53 SQLite database connections, and I'm going to tell it to have Muse, wonderful
9:58 GPT-61 Sol, and Gemini 4 Argon review your changes in delegates. Address any
10:04 useful feedback they give. And despite the fact that I'm using Opus and Claude
10:07 Code right now in T3Code, which by the way uses your actual Claude Code with
10:11 your actual Claude Code sub, this will now spin up with T3's built-in
10:16 orchestration all of these other agents using their native harnesses. So, I'm
10:21 not stuffing 61 Sol into Claude Code here. I've taught Claude Code how to
10:27 call Muse, Codex, and Gemini models all within T3Code itself. And now you see
10:30 all of these sub-agents spinning up. They all get their own threads
10:35 internally that allow them to review these changes and to communicate with
10:38 the main thread. This makes setting things up much easier. The T3Code
10:42 nightly has this. The stable release is coming very soon. Being able to have
10:45 your agent spin up other sub-agents like this to get feedback is great. This
10:50 isn't just cuz like using all of your subs is great. This is because the
10:54 different models have different training. They act differently. This is
10:58 similar to like if you get three different people with different
11:01 experiences to review a PR, the likelihood that they find unique things
11:04 goes up. So, if you take advantage of your other subs this way, it's very
11:08 useful. I personally find that OpenAI models are the most thorough and find
11:12 things better than others do. Claude models can find like really weird,
11:16 potentially dangerous things that other stuff will miss. And then Muse and
11:20 Gemini both can find things here and there, but I find that Gemini
11:23 hallucinates more than actually finds. So, if you don't want to make this a
11:26 scale or you just want to test the flow, what Morg said here is a great way to
11:30 start. Delegate to N sub-agents, compare the answers, refine your dispatching,
11:34 and then repeat however many times. And we now have our verdict from Soul here.
11:38 No important product code findings, all seven tests passed. Here are other
11:42 issues that it cares about. You get the idea. I have found that when I invoke
11:46 this skill, that the PR that goes up has fewer issues, is more likely to be
11:50 correct, and ends up getting through review faster, and being merged faster.
11:54 It does take longer for when I send the prompt for the PR to open, but it takes
11:57 a lot less time for that PR to get closed. And most importantly, is the
12:01 thing I value a lot. I don't want to throw a bunch of slop at my team. And I
12:04 found that this skill reduces the likelihood that slop goes from a thread
12:08 in T3 Code, or whatever else you're using, up into a pull request. If you
12:12 have seen PRs your agents opened and felt guilty, this should help quite a
12:16 bit. Now it's time to talk about those other two skills, full stop and full
12:19 send. I think this will all make more sense if we start with full send, but
12:23 before we can do that, I want to talk a bit about how I have been doing things
12:27 in T3 Code. And this is kind of becoming a T3 Code atom. I'm sorry, it's free and
12:31 open source, so I hope you're not too mad at me for building the things I like
12:34 into my products. If you haven't used T3 Code before, one of the important
12:39 mindset changes to understand is this idea of the sidebar being an inbox. When
12:44 your work is done, it should go away. Settled is similar to archive. When your
12:48 thing has merged, or you're throwing it away, or you've solved the problem you
12:53 have, you should not ever be scared to settle it. And ideally, you have a small
12:56 number of threads at any given time. I'll show you what I mean with this
13:00 particular thread that has a PR open that I just reviewed and seems fine to
13:03 me. Now this PR has been merged on GitHub in approximately 60 seconds at
13:08 most, probably less. It now gets marked as merged, and this one doesn't auto
13:11 settle cuz I actually turned it off cuz this was a long-running one. But if I
13:15 didn't have that flag off, which is very annoying timing-wise, this would have
13:19 just disappeared. And that's your goal with T3 Code, all threads should be
13:23 working or gone. So I specifically really want to optimize to get my PRs merged because I
13:30 want these threads to go away. So, once again, showing these workflows cuz I
13:34 want to get this idea across of my goal being every thread should get merged and
13:39 closed as reasonably quickly as possible. I'm not saying merge a bunch
13:43 of slop. I'm saying that you should try to optimize for not having too much
13:48 stuff that is waiting on you. And you should make it as easy as possible to
13:52 merge with confidence so that you can get these threads gone. So, hopefully at
13:55 the end of the day, you have a bunch of stuff in the little optional working
13:59 section and not a whole lot of stuff that you owe responses to. So, with all
14:02 that said, it's probably time to talk about my full send skill. The point of
14:06 this skill is to really lean into this idea of not taking up space in my brain
14:11 unless it has to. Remember, when PRs merge, the thread gets auto-archived. I
14:15 also have things set up with this optional working section in T3 Code.
14:20 This is a beta feature. I really like it and I might even make it the default in
14:23 the near future. It looks like they just made it the default in Claude Code,
14:26 funny enough. The point of the section is that if a thread is currently
14:29 working, it doesn't need your attention. It should just go hide until it's ready
14:33 for you. This also means that I'm not going to see what the threads are doing
14:36 if they're doing something like waiting, for example, like this one has 10
14:40 sub-agents it's currently running. This thread might output something that it
14:44 wants me to see and I won't see it because it's stuck under this working
14:47 section. I think that's ideal. If it does ask a question with like the
14:51 question tool, it will get yanked out and thrown into your inbox up here. But
14:55 I decided to build more of my workflow around all of this. I wanted the ability to agree with the
15:01 model, to know what it's going to do roughly, to know it's going to verify
15:05 it, to test it, check it. And if it's a small fix that I know what it's going to
15:09 do and how it's going to do it, or maybe it already made the fix and it has to do
15:11 one more small thing before it's mergeable. I don't want to have to go
15:15 back to that thread. I don't want it to interrupt me again to force me to
15:21 re-remember all of this just so that I can go click the one button merge.
15:24 Ideally, once the model and I have gone back and forth and I have confidence
15:28 that it knows what I want and roughly how to do it, it will not bother me
15:33 again unless it gets stuck. And that's why I made this full send skill. Again,
15:37 the description is just a trigger. I say trigger if the user requests full send.
15:41 If the user triggered this, it's because they're not watching the thread and they
15:44 want you to complete the work end-to-end autonomously. Your updates and your
15:48 final message go unread until long after you stop. You own the work from its
15:52 current state to a merged PR. If the user did not explicitly request this in
15:54 some clear way, do not assume full autonomy. If they did, everything else
15:58 below applies. This is an escape in case it randomly pulls in the skill. I call
16:01 out explicitly what I'm granting permission for because I found that
16:06 certain models, OpenAI models, get antsy about merging unless they have in their
16:11 context user gave explicit permission to. So, I have Theo granted you
16:15 permission to merge the PR when the merge gates pass and to force push the
16:19 PR's own branch with force with lease after rebase. This is one of those
16:23 things I added cuz again, my skills are malleable. If I notice things not
16:28 working how I want, I ask the model, "Why did you do that?" And it says, "I
16:32 thought this because of this." And then I go update the skill to prevent the
16:36 thing I didn't like. So, these bullet points here are all because OpenAI
16:40 models were doing things I didn't like. Then I describe the flow. Finish the
16:45 change, prove it, file the PR, babysit the PR, merge, and report. The prove it
16:50 section is important. I want the model to have found some way to verify that
16:54 its changes worked. I also mention my other skills to make sure it properly
16:58 babysits and watches pull requests. This is again cuz my babysit PR skill was
17:02 leading to OpenAI models not taking advantage of T3 Code's built-in watchers
17:06 and doing stupid instead. So, I added another section here. In T3 Code,
17:11 you should use the watch pull request feature. If you stop before the merge,
17:15 call unwatch pull request first. Again, bug fixes in my skills. Here I have my
17:19 proof section. Proof is evidence that you observed yourself in the session.
17:22 Command outputs, reproduction, screenshots. Use the most direct proof
17:26 that the change allows. Reproduce the original failure, apply the fix, then
17:30 watch the same reproduction pass. Run the checks that the repo CI runs. Add
17:34 one focused test that fails without the fix when the behavior has none and the
17:39 repo has a test setup. Enforce UI and runtime behavior. Run the app and look
17:43 at the change. Green mocked tests don't prove the user experience, yada yada. I
17:46 also call out my file upload skill so that it can upload screenshots or videos
17:51 when it proves the thing is fixed. I then have an explicit merge gate. This
17:55 is what I'm expecting before it can merge. The proof covers every change.
17:58 The required checks are all green. All review bots have finished and the
18:01 findings are fixed or dismissed with a written review. No human review is still
18:05 requesting changes and the branch is up-to-date with the default and GitHub
18:08 reports it is mergeable. Then we have the most important section.
18:12 As I mentioned before, the point of this skill and the point of this flow is to
18:16 make it so that when I'm done with the thread, I don't have to look at it again
18:19 and I trust it to get the change in without having to look at it one last
18:23 time. Asking Theo is an important section for this reason. The model needs
18:27 to know how to get a hold of me when I'm ignoring the thread. I told it
18:31 explicitly to use the question tools that it has available to it. A question
18:34 in a normal message goes unseen and the thread looks done. To ask Theo
18:38 something, use the harness's question tool so the thread shows that it's
18:41 waiting. Ask user question in Claude code, request user input async or normal
18:44 request user input in Codex, or the available question tools and things like
18:48 antigravity. Put all of the context in the question. The PR link, what blocks
18:51 you, and the options that you are recommending first. Ask only when you
18:55 can't continue alone. Cuz I noticed after adding the section, it was asking
18:59 all the time. So I put this here as like a gate saying, "Don't ask unless." And
19:04 one last, take the safe path if you don't have access to the things that you
19:07 need. You know how nice it is having all of this data in one database on my network so my
19:12 agents can like query against it. It's so cool. So, in the limited window that
19:16 I am currently querying for my usage here, I had 32 times where I said full send,
19:23 and of those, 27 merged without me having to look at them again. Do you
19:26 know how cool that is? I know it seems kind of insane to not look, but as I've
19:30 talked about in other videos, you should build good verification systems so your
19:34 agents can check if the thing works, but you should also build good safety nets.
19:37 So, if it does fail, you can catch it before it hits a large percentage of
19:40 your users and fix it before that happens. In T3 Code, we have a nightly
19:45 that goes out every 6 hours, and we get feedback immediately if it breaks
19:48 something. So, I can fix it before the next stable release goes out. And I know
19:51 this whole workload might seem kind of insane, but here is one of the lead devs
19:56 on the Cloud Code desktop app publicly saying that they might change the
20:00 default to grouping by status to better align with this type of workflow. More
20:04 and more, we are seeing devs that are taking advantage of AI, trying to think
20:09 less about how do I get the best code for me to go look at, and more about how
20:15 do we build workflows where the agent can verify itself and ship confidently,
20:19 while also being able to undo broken stuff when it does happen. Cuz in the
20:22 end, no matter how good you, your agents, your verification systems, or
20:26 your developers are, mistakes will happen. You need to make sure you can
20:29 undo those mistakes with as little damage as possible. That said, of the 32
20:33 times I've used this in the last couple of days, 27 merged themselves, and I
20:38 know which threads I used this on, none of them caused problems. Once you've
20:43 done enough of this type of like agentic coding, you know what the models can and
20:46 can't do, what they can and can't verify, what needs an extra look, and
20:50 what doesn't. And once you've built that intuition, it's relatively simple to
20:53 know this thing can probably merge itself, this thing probably needs
20:56 another look. But I love this workflow way too much to not use something like
21:01 it, even when the changes are a bit hairy. And this is where my final skill
21:06 comes in. Full stop. If you've been paying attention, you might have an idea
21:10 of what full stop is and how it differs from full sent. I'm going to be real.
21:15 This skill is pretty much entirely slop. It should probably be called full slop.
21:19 The reason this skill exists is because I liked the full sent workflow so much
21:25 that on the scarier changes I missed it. I didn't want them to merge, but I
21:29 wanted them to do everything else. So by the time it bugs me, it's ready to go.
21:35 So full stop is Claude rebuilding my full sent skill with one big change. Own
21:40 the work until the PR is ready for him to merge. Keep fixing, verifying, and
21:44 reviewing. Never merge it. This is again, let the model operate without me
21:49 paying attention to it, but don't bother me until you are very high confidence
21:53 the thing is ready to go. Do everything in your power to verify your changes, to
21:58 talk to other agents and other models to get verification of your changes, put up
22:02 the PR, baby sit it, deal with all the CI failures, deal with all of the review
22:06 bots leaving comments, deal with all of the other people on my team reviewing
22:10 and leaving comments, and do not bug me unless there is something hard blocking
22:14 you that you need before or you think it's ready to merge. I love this skill.
22:18 This has quickly become my most used skill to the point where when I was
22:22 starting stream today, people were sending me PRs they wanted me to look at
22:27 and pretty much all of them I opened up and just immediately opened with figure
22:31 it out full stop. This one is actually funny. Somebody reported a bug to me in
22:36 Twitter DMs, sent me the two PRs they filed to fix it. I didn't feel like
22:39 going through it all and dealing with it. So I copy pasted his message and
22:42 said, "Friend sent this. Triage and takeover. Full stop." I do also have a
22:46 takeover skill. This one's relatively simple. I wouldn't overthink this. This
22:50 is very much like a I'm a maintainer and I have I have to deal with
22:53 skill. Take over existing work. Use when the user asks to take over a PR branch
22:57 or agent thread or request to take over. Take responsibility for finishing the
23:00 original goal while preserving the original contributor's credit. The
23:03 reason I made this is when I saw PRs that were fixing things that mattered,
23:07 but I couldn't necessarily trust or I didn't want to just like push a bunch of
23:10 things on top of their branch or PR, I made this skill so I can link a PR to T3
23:15 code or whatever agent I'm using within it and say, "Take this over." Full stop.
23:21 Now, what this will do is it will take a look at that PR, review it, read the
23:24 description, see why they made it, decide if the code is in a good enough
23:28 state to continue or if it should be scrapped and treated as a reference
23:32 rather than actual like starting points, make a new PR with all of the changes if
23:35 that's what makes sense, contribute to the existing one if that's what makes
23:38 sense, credit to the original authors. This is very important to me cuz I don't
23:42 want people feeling like I am just closing their PRs and cloning their
23:46 ideas. I want the people who make those original PRs to get credit. It's just
23:50 too much work to try and get them to make the PR perfect. I would much rather
23:56 have somebody on my team that I know can get this through to the end and can hit
24:00 the merge button most importantly. I'd rather have them own the PR and get the
24:04 thing through. So, this skill allows me to do that without killing the credit
24:09 for those contributors. I also call it my modernize skill in this one, which is
24:13 a really, really simple deal with this PR being out of date, rebase on main,
24:17 check to make sure nothing else has merged in this time that addresses the
24:21 same stuff so that we know if this is necessary or not, and close it if it's
24:24 no longer necessary. What you might be noticing here is this weird like graph
24:28 of skills that is happening. I started with my modernize skill, which is a flow
24:32 I found myself doing a lot. I was saying the exact same thing to models like,
24:35 "Make sure the PR is up to date, close it if it doesn't make sense anymore cuz
24:39 other things that are similar merged." So, I turned that into a skill. I kept
24:43 telling my threads to take over PRs and branches from other things, realized
24:46 that should be a skill cuz I was running into the same problems consistently.
24:49 Then I wanted to take more advantage of using sole sub-agents to review, which I
24:53 did by hand a lot, so I made the sole loop skill. I realized the quality of
24:56 the code coming out was better than I expected, so I introduced the full send
25:00 skill to see it all the way to the end. Then I missed that flow a whole bunch,
25:03 so I made the full stop skill to get as much of it as I could without having to
25:07 be bothered. Before I forget, I'm going to fix this description. Change the
25:11 description to use when Theo explicitly asks for full stop. This is actually how
25:16 I update my skills across the fleet. I think there's a script in the repo
25:18 somewhere that gets run when I do this. I don't know. I don't care. I would much
25:22 rather just say what I want, and the models are smart enough to do this. One
25:25 more thing before I forget cuz I saw people saying this in the comments, and
25:29 I need you guys to understand why I don't like it. Wouldn't it be better to
25:33 just invoke it directly? I saw other people saying, "Why don't you turn off
25:36 skill invocation?" For those who don't know, in the skill standard, there's an
25:41 option for disable model invocation. What this means is the model can't pull
25:45 the skill in itself. You have to manually invoke the skill. So, this
25:50 would be like a {slash} soul loop or {slash} full send. These are the {slash}
25:56 commands or like the dollar sign syntax for it to pull skills in. Those are
26:00 great if you are typing yourself, and you want to manually make sure the skill
26:04 only comes in when I specifically do the thing like that. I don't want to have to
26:08 remember to {slash} command the different skills. It's also worth noting
26:11 that certain harnesses, mostly cloud code, don't let you manually invoke
26:15 multiple skills at the same time. So, if you disable the model's ability to
26:20 invoke, you will not be able to trigger multiple at once in certain cases. It's
26:24 incredibly annoying, and I'm going to be real. I don't remember the last time I
26:27 invoked a skill with {slash} other than the grilling skill, because I do have
26:31 the Matt Pocock grilling skill as manually invoked cuz that's how they
26:35 provide it. I don't manually trigger my skills. I write descriptions that
26:39 describe the trigger phrase, and then I can do something like soul loop
26:46 takeover, full stop, or even press my speech-to-text button say,
26:50 "Full stop takeover soul loop." And it will get it right like that. I don't
26:55 need to do a bunch of special syntax just to invoke the different skills I
26:58 need. But when do you invoke multiple skills? I don't know, the five times I
27:02 just showed it. Here's one that I did literally as I was starting stream today
27:05 cuz somebody linked an issue that they had fixed with this PR for opening
27:10 explorer in Windows. So I took the PR that they linked, I copy pasted it, said
27:14 take over and get this landed. Full send. Now I have invoked two different
27:20 skills and it got this fixed merged and it closed the old PR cuz it wasn't
27:23 needed anymore. And I bet if we go to the one that it merged, which by the way
27:28 I've never even opened, it merged and it gave credit to the original author. So
27:33 this combination of skills allowed me to take a PR I hadn't looked at, make a new
27:37 PR based on that, credit the original user, verify the fix has worked. I think
27:42 it actually used a Windows PowerShell like VM to test this, and then merge it.
27:47 All without ever having to look at it. And since my skill specifies to make
27:52 sure that this works and is safe before merging and if not, call Theo and make
27:56 sure, I'm pretty confident if I never had to see this thread again, that we're
28:00 good. This is also very good to know. Thank you, Alex. When you invoke a skill
28:05 manually in Claude code, it injects the prompt as a argument to the skill
28:10 because skills in Claude code can be a bit more programmatic like you can pass
28:13 them information. So when you do invoke a skill manually in Claude code, it's
28:17 going to put your whole prompt twice in the context. Not ideal.
28:22 Don't manually invoke, it sucks. And there you go. I think that showcases why I'm enjoying
28:26 these skills so much and how I've been thinking about my skills. I already know
28:31 how many of y'all want me to include all of this in some repo somewhere. I I am
28:35 leaning towards doing that. You know what? it, it's time. New project and it looks like we are
28:44 going with T-Stack. I will not have this ready particularly soon, but you should
28:49 definitely keep an eye out for a T-stack video in the future. So, if you haven't
28:51 already subscribed, you should consider that. Helps us a lot, and it keeps you
28:55 on top of all of these cool developments in the AI dev world. Educational content
28:59 is not the best performing stuff on YouTube. I appreciate you immensely for
29:03 getting this far into the video. I am finally caving and doing my first ever
29:08 course. If you don't have a full-time job as a software dev, and you're not
29:13 being paid real ends money, you probably shouldn't do this. But, if you can get
29:17 your educational budget to cover it, or you can convince your boss to cover the
29:20 cost for you, Mega Dev is going to be really really cool. I've spent the last
29:23 few months working on the first chapter for this course. I'm the opener, John
29:26 Lindquist, the creator of Egghead, the closer, Kenzie Dots and Angie are
29:30 filling up the middle. This isn't going to be your usual course on like build an
29:34 app from scratch type stuff. This is how we use AI to really ship at actual
29:41 companies doing real-world work. And we tried to make it not so specific to how
29:46 things work today, rather more general how we think about AI and how we use
29:51 these tools and technologies to navigate in this new world. Very excited about
29:54 what we're cooking there. Tell your boss about it so you can get it covered. But,
29:56 please don't buy this in order to get a job. Buy it because your job covers it,
30:01 and you want to level up your skills. Maybe get your team to sign up, too. I
30:03 think that's all I have to say on this one. I hope you guys got the lesson I
30:07 was trying to teach here, which is not you should go copy my markdown files cuz
30:11 I'm super smart, rather that you see how I find solutions to the problems I have,
30:14 and I'm building a more and more abstracted workflow where I am involved
30:19 at a different layer than I used to be. I'm becoming a manager of myself in a
30:22 way. And in order to get the most out of your agents, it's important to analyze
30:26 your flow and figure out what you do and don't trust them with, and start
30:29 encoding that in stuff like skills, so that you can ship more effectively. All
30:32 that said, I guess T-stack is coming soon. So, keep an eye out for that. Sub
30:35 if you haven't already, and until next time,
$

These 3 skills changed my life

[AI agents and automation][community building and audience growth][product development and MVP][productivity and workflows]
// description

As I've shipped more and more, I've realized these three skills I wrote are the cornerstones of my workflow and how I ship in 2026 Thank you to WorkOS for Sponsoring! Check them out here: https://soydev.link/workos CHECK OUT MEGA https://soydev.link/mega Oh also I did make tstack... https://github.com/t3dotgg/tstack Want to sponsor a video? Learn more here: https://soydev.link/sponsor-me Check out my Twitch, Twitter, Discord more at https://t3.gg S/O @Ph4seon3 for the awesome edit 🙏 #ai #

now: 0:00
// tags
[AI agents and automation][community building and audience growth][product development and MVP][productivity and workflows]