YT YouTube2Text

YouTube2Text → Transcripts → New Skills! v1.1 brings /wayfinder, /research, /implement, /to-spec, /to-tickets

New Skills! v1.1 brings /wayfinder, /research, /implement, /to-spec, /to-tickets — Transcript

by Matt Pocock · 3,131 words · 437 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hello friends. First video in a while,
0:01and that is because I've been working on
0:03version 1.1 of my skills repo. It has an
0:07astonishing amount of stuff in there.
0:08There is an entire new approach to
0:11grilling, which probably deserves its
0:13own video, but I'll try and squeeze it
0:14in here. There is a bunch of new changes
0:17to existing skills, including a rename
0:20of two main flow skills. There is just
0:22really way too much for me to summarize
0:24in this intro, so you're just going to
0:26have to watch the video to find out. We
0:27can see the PR is literally ready to
0:29merge now. So, why not? Let's actually
0:31freaking merge this thing. And just like
0:33that, we have our version 1.1 ready to
0:37go. Let's start with the two that are
0:38probably going to be most annoying for
0:40you, and the most like, why did he do
0:42this? And there is a very specific
0:44reason that I did it, which is that two
0:47skills have been renamed. For instance,
0:49two PRD has now been renamed to two
0:53spec. And if we go up one level, then we
0:55go to two issues has been renamed to two
0:59tickets. The reason I've done this is
1:01that this has just been bugging me for a
1:03long time. The thing that we were
1:04creating in two PRD wasn't actually a
1:07PRD. It was a spec. A product
1:10requirements document kind of describes
1:13more things about the actual product
1:15itself, whereas we were allowing things
1:17to leak into the PRD that weren't
1:20necessarily PRDs. So, for a long time
1:22I've wanted to rename it to two spec,
1:25cuz that's what we were creating. We
1:26were just creating a specification.
1:27Specification is a much broader term
1:30that actually entails what we were
1:32building.
1:33for a thing we want to build. That can
1:35be technical, it can be non-technical,
1:37and it can blend the two. It doesn't
1:39really matter. When it came to two
1:40issues, two issues always felt like it
1:42was biased towards GitHub and linear
1:44that use issues, but really we want this
1:47to be tickets. You have a spec, and then
1:49underneath the spec you have the tickets
1:51that are the journey that you
1:53take to actually enact the spec and
1:55create it. This has been annoying me for
1:57a long time, and finally it no longer
1:58annoys me. It brings me joy. Now, one
2:00irritating thing about this rename is
2:02that you will need to probably delete
2:04those skills and re-add them. This means
2:06you'll need to run NPX skills add Map Po
2:09skills. I'm pretty sure that this
2:11skills installer won't pick up the
2:13rename, so it won't try to update to PRD
2:16to turn it into to spec. You get what I
2:17mean. And so, running this command is
2:19the safest way to grab all of these new
2:21skills cuz you get to just pick and
2:22choose which ones you want. And once
2:24you've done that, you should probably go
2:25through a pass through your skills
2:27folder and just check that no bad ones
2:30are still in there. So, you want to make
2:31sure that you're intentionally grabbing
2:33all the right skills. The next change is
2:35that I've fixed a couple of bugs that
2:37people were having with grill me and
2:39grill with docs. Both of them rely on
2:41this kind of central reference grilling
2:43skill that kind of show shows the LLM
2:46how to grill a person. I've sharpened up
2:48this line here saying asking multiple
2:50questions at once is bewildering. Even
2:53with this direction to ask questions one
2:55at a time, it was still occasionally
2:57just going
2:59to have multiple questions at once. So,
3:01I've told it why we don't want multiple
3:04questions at once. We've also added a
3:05confirmation gate on the end. Do not
3:07enact the plan until I confirm we've
3:09reached a shared understanding. Lots of
3:11people on different models were
3:12reporting that the grilling session
3:14would just end and then it would just go
3:16straight into implementation. So, this
3:18is just an extra little gate there.
3:19Finally, on some situations, it would
3:22just grill itself, which is very, very
3:25odd. Not something I've noticed or seen
3:27in my personal thing, but I I can only
3:29get a small subset of how these skills
3:31are actually used. So, I've basically
3:33tried to use a couple of leading words
3:35to indicate the difference between facts
3:38and decisions. So, sometimes it was
3:40using the previous phrasing here by just
3:42exploring the code base and grilling
3:44itself. This was especially happening
3:45with Fable, actually. And so, I've
3:48decided to make a distinction between
3:50facts, so facts that are things you find
3:53yourself by exploring the code base, and
3:55decisions. So, decisions are needed to
3:57be made by the user. So, just a couple
3:59of sentences, added a couple of things
4:01changed around, and this has made it a
4:02lot more consistent. Definitely getting
4:04a lot fewer complaints about those weird
4:06issues happening. The next thing to say
4:07is that I've added a couple of skills
4:09that really just take the process that
4:12was primarily a planning process, didn't
4:14really hold your hand into
4:16implementation, and turn it into a
4:18proper software development life cycle.
4:20So, many folks ask me, "What is the
4:22flow? What is the main flow you're
4:23supposed to use with the skills?" And
4:26first of all, I mean, this is it. We
4:28have number one, you're supposed to
4:29instead of using plan mode, you get an
4:32agent to grill you. And it uses these
4:34couple of docs to add a glossary, to
4:37kind of understand you better as you go
4:39along, and also add architectural
4:40decision records so you can capture the
4:42non-obvious stuff. The stuff that goes
4:44in the growth for docs then goes into a
4:47spec, as we saw before. That spec kind
4:49of defines the destination, where you're
4:52going. Then, you turn that spec into
4:55individual tickets so you can spread the
4:57development of it out over multiple
4:59agent sessions. That's the purpose of to
5:01tickets. You then implement each one of
5:03those tickets with a implement skill.
5:06And the implement skill is very, very
5:08simple. It just looks like this.
5:09Implement the work described by the user
5:11in the spec or tickets. Use TDD where
5:14possible at pre-agreed seams. That's a
5:16nice one. And run type checking
5:18regularly, single test files regularly,
5:20and full test suite once at the end.
5:22Once done, use code review to review the
5:24work, and then commit your work to the
5:25current branch. I almost didn't make a
5:27skill for this because it's really
5:29simple, right? It's just mostly relying
5:31on the agent's priors, on its, you know,
5:34on the harness, kind of teaching it what
5:35to do. And I didn't honestly think we
5:38needed a skill here, but folks kept
5:40asking me, "What's the flow? What's the
5:41flow here?" And so, I figured just an
5:43implement skill, make it nice and
5:45simple, right? At each stage of the
5:46process, call this skill. So, that means
5:48implement earns its place here because
5:50you know, okay, once we got two tickets,
5:53then we just got to implement each
5:54ticket in a separate coding session.
5:56Implement then itself calls code review.
5:59And code review, I graduated this out of
6:01in progress on version one. So, it's
6:04been around for a little bit, but I have
6:06made some updates to it in this version
6:08two. The theory of the code review skill
6:09is that it reviews code on two axes. So,
6:12it does a sub agent for each one of
6:14these. The first one is the standards
6:16axis. Does the code conform to this
6:19repo's documented coding standards? So,
6:21if you've got a coding standards.md file
6:24somewhere in your repo, then it will
6:26read that and check against those. I
6:28generally think that coding standards
6:30belong outside of your agents.md file.
6:33They're supposed to be somewhere
6:34separate, and the code review point is
6:36where they're most useful. And then,
6:39once you've done the standards, you then
6:40go on to the spec. Does the code
6:42faithfully implement the originating
6:44issue or PR or spec? They both run as
6:46parallel sub agents, and it does a
6:48process here where it walks through each
6:51part. Now, the thing that's cool and new
6:53about this skill is that I've been
6:54reading Martin Fowler's refactoring
6:56again. And what I decided to do is
6:59Martin Fowler names a bunch of different
7:01smells that the agent can identify in
7:04bad code. Refactoring is such an old
7:06book, such a well-cited book, that these
7:09uh kind of smells are deep in the
7:11agent's priors. And so, all you need to
7:13do is kind of invoke the idea Okay,
7:15mysterious name or duplicated code or
7:17feature envy, data clumps, primitive
7:18obsession, repeated switches, divergent
7:20change, speculative generality, message
7:22chains. You see what I mean? Like, these
7:24are all deep in the agent's kind of
7:26knowledge base, and all we got to do is
7:28just really describe them in a sentence.
7:30And what I found is that leads the word
7:33or leads the agent to repeat that word
7:35back to you and say, "Yes, I found some
7:37message chains. I need to remove them. I
7:40found a middle man situation. I need to
7:42uh fix that." So, I tested this for a
7:43couple of weeks and it was outrageously
7:45useful. It was really, really nice at
7:47improving the quality of my code and
7:49it's really cheap to add here, just kind
7:51of like 10 lines. But, let's go and talk
7:53about the one that I'm really, really
7:55excited about, which is a whole new
7:58change to the way that we kick off and
8:01shape specs. So, the pre-spec bit. In
8:04other words, it goes here where it may,
8:07in some situations, replace Grill with
8:09Docs. And it's called Wayfinder. I will
8:11make an entire post uh entire video
8:14about Wayfinder, but uh suffice to say
8:16is that I would love for you, in
8:18situations where you're thinking about
8:20using Grill with Docs, instead to
8:22default to Wayfinder instead. What
8:24Wayfinder does is it's designed for
8:26situations where you have a ton of stuff
8:29that you want to plan, but and too big
8:32for one agent session. In other words,
8:33you're going to blow out of the smart
8:35zone of the agent or you might even blow
8:37out of the context window of the agent.
8:39You need to split it into multiple parts
8:41in order to figure out where you're
8:42going. A loose idea has arrived, too big
8:44for one agent session and wrapped in
8:46fog. The way from here to the
8:47destination isn't visible yet. This
8:49skill charts the way as a shared map on
8:52the repo's issue tracker, then works its
8:54tickets one at a time until the route is
8:56clear. These maps are saved in GitHub
8:58issues. For instance, this is one on the
9:00Sandcastle repo where we're doing a
9:02spike to think about maybe pulling in
9:04the AI SDK as a dependency, a big, big
9:07change. And so, you can see there are no
9:09decisions that have um been made so far
9:12and all of the decisions that need to be
9:14made are saved in sub-issues. And these
9:16sub-issues have blocking relationships.
9:19So, we can see that no decision can be
9:21made here before we make this key
9:24decision at the start. Each one of these
9:25decisions is scoped to be the size of an
9:28agent session. And we can see that
9:30they're labeled as different types here.
9:32So, for instance, this one is labeled as
9:34a research task. So, this is really an
9:36AFK task for the agent to go off, do
9:39some research, and then come back. This
9:41one I think is a research as well. This
9:43one is a research task, and this one is
9:45a grilling task. So, this one needs a
9:48grilling session to be done here. I
9:50think these are all grillers. We can see
9:52these defined in the ticket types down
9:54here. So, we have research, we have
9:55grilling, we also have prototype as
9:58well. So, this is something I've been
10:00really advocating for recently is doing
10:03more prototyping before you get to a
10:05spec. The idea is you raise the fidelity
10:07of the discussion by making a cheap,
10:08rough, concrete artifact to react to. An
10:11outline, rough tech stub, UI logic code
10:13via the prototype skill. We'll get to
10:15that in a minute. Links to the prototype
10:17as an asset. And it says use when how
10:19should it look or how should it behave
10:21is a key question. And this is essential
10:24for almost anything that touches
10:25front-end code. So, I would definitely
10:27be recommending using Wayfinder for
10:29anything that touches the front-end. The
10:31final one here is just tasks. So, config
10:34that needs to be set up, um provisioning
10:36access, you know, moving data into the
10:38shape, you know, all the sort of boring
10:40stuff that doesn't need a grilling
10:42decision and can't really be automated
10:44by AI. What you end up with is after all
10:46of these tickets are closed, all of that
10:48information gets saved onto the map with
10:51the original tickets as kind of primary
10:53sources for what was captured. And you
10:55can then take this map and just turn it
10:58into a spec in the regular way. What I
11:00found that instead of having the kind of
11:01anxiety of managing my session with
11:04grill with docs having to hand off,
11:06worry about the smart zone. With
11:07Wayfinder, it's kind of all managed for
11:09me. I just get to close a session, open
11:12up the next Wayfinder ticket. It's all
11:14saved in GitHub, so it's collaborative.
11:16You can share it across your team. And
11:17once the map is done, once it's
11:19complete, you just go to to spec, and
11:21you're good to go. To support the
11:22Wayfinder skill, we have a new research
11:24skill, which is very small, very handy
11:27for when you just need to do a research
11:29session, or it kind of influences the
11:31model in researching in the right way as
11:33well, or at least the way that I like.
11:35Spins up a background agent to do the
11:36research so you keep working while it
11:38reads. Investigate the question against
11:40primary sources, write the findings to a
11:41simple markdown file, and save it where
11:43the repo already keeps such notes match
11:45the existing convention. So this is
11:47useful too if you need to do any
11:48research, you can just invoke the
11:50research skill and you're good to go.
11:51The next one, of course, is the
11:53prototype, which I've kind of shown off
11:55a little bit before. I don't think I've
11:56done a full video on it. This is now
11:58modeling, folks, so that Wayfinder can
12:01invoke it itself, and it essentially
12:03gives you a choice between logic or
12:05state. So it's either a logic prototype
12:08or a UI prototype, and they react quite
12:11differently. The final change is
12:13something that people have been asking
12:14for for a while, and I finally decided
12:16to pull the trigger on it, which is
12:18before in my TDD skill, it would
12:20recommend a set of steps for you to
12:22follow, and that was a little bit
12:24awkward sometimes. The steps were like,
12:26it would confirm what tests it wanted to
12:28write with you, and then you would, you
12:31know, walk it through, walk it through,
12:33and it didn't fit with most people's
12:36idea of how TDD should work, which is
12:38you should be able to pass an AFK agent
12:41the TDD skill, and it should just work.
12:43And so this TDD skill is now reference
12:46material only. So it doesn't specify any
12:48particular steps apart from just the
12:51order in which you should write tests
12:52in, so to do red green refactor. So it's
12:55just says red before green, one slice at
12:56a time, and it also splits away
12:59refactoring as as not part of the loop.
13:02So it's no longer a red green refactor
13:04loop, it's more just red green. I tend
13:07to think that putting the refactoring in
13:10the code review part is a lot more
13:12productive because then you don't
13:14overload the implementation. So that is
13:17all of the changes that have come in on
13:19the skills. It is a lot of changes, and
13:22if you're nervous about missing any of
13:24the updates, then I recommend that you
13:25clear out all of your skills and do NPX
13:28skills update and grab all of the new
13:30ones. If you've made updates to your
13:31skills in the meantime, then you can
13:33just point your clanker at my repo and
13:35just say pull down all of the good new
13:36stuff, especially pointing at the
13:38release notes. The thing I think this
13:40release will be remembered for is to
13:42spec and to tickets changing because
13:43that is just a little bit of friction,
13:45but I think good friction cuz it names
13:47it properly. And I hope to be the start
13:50of you getting obsessed with Wayfinder.
13:51I'm using Wayfinder for literally
13:53everything, even non-coding stuff. I've
13:55actually been planning my next course
13:57with Wayfinder and it's really, really
13:59good. In fact, why don't I just show you
14:01that course now? This is the AI coding
14:03crash course. This is going to be
14:05different from the cohorts that I
14:07usually run. It's going to be much, much
14:09cheaper and it's going to be self-paced
14:11so you can purchase it anytime. You get
14:13help from the Discord kind of in the
14:14usual way, but it's not going to be
14:16gated like a normal cohort. It is going
14:18to be the perfect intro for anyone who's
14:21looking to get into AI coding, whether
14:23you are a developer or whether you are
14:26not a developer. So, for senior
14:27engineers, it's going to be a conversion
14:29course. For folks who are new to
14:31development, it's going to be the way
14:33that you can actually get productive
14:34using these crazy new tools. I've not
14:36announced a price yet. I am going to
14:38just be adding sign-ups in here and
14:40it'll be available once I finish filming
14:42it, maybe in about
14:44August time, I think. But folks, thank
14:46you so much for watching. It's always a
14:47pleasure sharing these skills updates
14:49with you. It is really cool to see the
14:51usage just absolutely grow and grow.
14:54Everyone tells me I shouldn't show the
14:55star counts, but it's up to like 160K
14:57stars now. Uh 7 million downloads on
15:00skills.sh. It is just bonkers. So, thank
15:03you so much for enjoying the skills. I
15:04hope they are helping you ship more and
15:07ship more productively and I'll see you
15:09very soon.

About this transcript

This page contains the full transcript of New Skills! v1.1 brings /wayfinder, /research, /implement, /to-spec, /to-tickets by Matt Pocock, generated from the public captions YouTube serves with the video. The transcript has 3,131 words across 437 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.