YT YouTube2Text

YouTube2Text → Transcripts → /wayfinder: Nothing is too big to plan anymore

/wayfinder: Nothing is too big to plan anymore — Transcript

by Matt Pocock · 3,144 words · 446 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I think I figured out a way to plan any
0:02size of work with an agent. The existing
0:05planning tools that I was using and
0:07playing around with, even ones that I've
0:09created, felt too constrained, too tied
0:12to a single session. I felt like I
0:14wasn't able to be ambitious enough, and
0:17because of that, I was kind of
0:18constraining the stuff I was building to
0:20fit AI, which doesn't feel right. This
0:23new approach doesn't have that limit.
0:24You can plan enormous chunks of work,
0:27and it will orchestrate the planning
0:29over multiple types of sessions. It
0:31knows that you can't make your way
0:33cleanly to the destination. You have to
0:35clear the fog of war. It understands
0:38dependent decisions, and it even allows
0:40you to plan in parallel. And the best
0:42thing about this is this is based on
0:44software fundamentals. This is based on
0:46the fundamentals of planning work that I
0:48learned when I was a real developer
0:49before AI. And I've packaged this all up
0:52into a skill that's available right now
0:54on my skills repo called Wayfinder. So,
0:56the way that I was planning work before
0:58was really tied into a single session.
1:00It's to do with the grill me or the
1:02grill with doc skill that's in my skills
1:03repo. That's still a super important
1:06primitive, but it's really just tied
1:08into a single session. Some work is
1:10bigger than what you can fit into the
1:13context window, and especially the smart
1:15zone of the context window of the agent.
1:18And you know that going in, so you'll
1:20often take time ahead of these AI agent
1:23sessions to break it down into smaller
1:25chunks, say, "Oh, I'll just bite off
1:27this little bit. I'll just bite off this
1:28little bit." But then what you'll find
1:30is, "Okay, I'm working towards planning
1:32in this bit of grilling." And then you
1:34reach a question that you can't answer,
1:36or you just find yourself lost in fog.
1:39And all the time you're managing the
1:40smart zone, you're trying not to spend
1:42too many tokens. This has been out there
1:43for a while, and people are freaking
1:45loving this thing. It one-shotted a
1:47prototype, it kept starting again and
1:48again for months. I really hate the
1:50phrase one-shotting, but I think what he
1:52means is it really helped him out. John
1:54here even built his own freaking harness
1:57because he liked the wayfinder approach
1:58so much. It's got these gorgeous little
2:00star map on it that kind of lets you
2:02take tasks as you go. So, it's been out
2:04there for a little while and I'm finally
2:06making the video that people want me to
2:07make. What is wayfinder? How do you best
2:09use it? Well, let's start by looking at
2:11how big work typically gets planned. You
2:14have a start point, a point where you
2:17need to start from, a sort of vague
2:19idea, not really how to get there. And
2:21you're trying to get to some kind of
2:23destination. You know vaguely where you
2:25want to end up, but the steps between
2:27are super foggy. You've no idea how to
2:30get. This is true, by the way, in
2:32engineering, but it's also true in many
2:34walks of life where you're planning
2:35something ambitious. And so, the first
2:37thing you should probably do is have a
2:39grilling session about it. Get the AI to
2:41interview you and figure out the sort of
2:44basic premise of where you're going.
2:45Now, for some work, that's sufficient
2:47and you'll be able to get straight to
2:48your destination. But for a lot of work,
2:50that will still leave you in a lot of
2:51fog. What you might find is based on
2:53that initial grilling session, you need
2:54to do more sessions. So, you might have
2:57a prototyping session, or you might have
2:59another grilling session, or it might
3:01need to go off and do some research as
3:03well. Conceptually, what we're looking
3:04at here is a map. We are creating a map
3:07of how we're getting to our destination.
3:09This is why it's called wayfinder. We
3:11are finding our way to the destination.
3:13And each of these things on the map,
3:15they are tickets. Each ticket requires
3:17its own individual session with the
3:20agents. So, you might have a prototyping
3:22session, a grilling session, and a
3:24research session. And all of those
3:26things are created and managed by
3:28wayfinder. And just a note here, yeah,
3:30this is just a single skill doing all
3:32this and it works with any coding agent.
3:34On its map, wayfinder gives you a
3:36frontier of tickets here. In other
3:38words, the decisions that it knows about
3:40so far. And it also keeps track of
3:43everything that's in fog. So, things
3:45that are not quite able to be decided
3:47upon yet because we haven't done the
3:49research, or we don't have a prototype
3:51to look at, or we haven't done enough
3:53conversation, enough grilling. At some
3:55point, all of the fog will be resolved,
3:57and then you'll have finally made enough
3:59decisions to finally get to your
4:00destination. Wayfinder can not only
4:03manage the research, but it could also
4:04do tasks here, too. So, if you need to
4:07set up some configuration, or you need
4:09to go out and talk to someone, and
4:11actually go and run an errand, then
4:12Wayfinder can figure that out for you,
4:14as well. In other words, all of the
4:15complicated stuff that you might need to
4:17do while you're planning something big,
4:19Wayfinder orchestrates it all for you.
4:21It keeps track of everything that's been
4:24done, and it measures the fog of war for
4:26you, keeps track of all the frontier of
4:28things you can decide right now. How
4:30does it keep track of it? Well, it does
4:31it in your issue tracker. In my public
4:34course video manager repo, here are all
4:36of the Wayfinder maps that I've done
4:38recently. And you notice that if we look
4:40at this one, there are This is the big
4:43old map here, and underneath it are 12
4:47subtasks, or sub-issues, and these are
4:49the decision tickets. So, we can zoom
4:51down here, and we can understand all of
4:54the decisions that have been made. As
4:55decisions get made, then obviously, they
4:58get resolved inside the ticket. So, in
5:01this one, this is a sub-issue, close the
5:03clips during publish race, and we
5:06resolved it with a discussion a couple
5:08of weeks ago. That resolution also gets
5:10written back up to the parent map. So,
5:12if we look back up here, we can see that
5:15a small version of that also gets
5:17written in the map. Wayfinder is keeping
5:19track of all the decisions that have
5:21been made, all the prototypes that have
5:23been created, all the tasks that have
5:24been done. And by the way, even though
5:26I'm using GitHub for this, my skills are
5:28issue tracker agnostic. So, you can use
5:30it with any issue tracker you like. You
5:32just need to do a little bit of
5:33configuration via
5:35setup.mappercode.skills. Use it with
5:36Linear, use it with Jira, use it with
5:38literally whatever you like. The very
5:40first thing you'll need to decide when
5:41you kick off a new Wayfinder session is
5:43the destination. For instance, in this
5:45one, I was adding a command palette with
5:48a bunch of new actions into my
5:50application. And what I ended up wanting
5:52was a buildable spec. So, I wanted a
5:55specification for this command K command
5:58palette in the CVM diagram window. So, I
6:01started it off like this. I invoked the
6:02Wayfinder skill, and then I gave it a
6:04description of what I wanted. I'd like
6:06the ability in the CVM to add an icon
6:08picker. Not only that, I want the
6:10ability to search other diagrams. I want
6:11the ability to copy things from the
6:13diagram and save them as, you know, big
6:14old chunk of work. It went through and
6:17explored the repo, and it invoked the
6:20grilling skill, and it grilled me about
6:22what I wanted. It first asked me what
6:25done looks like, whether I wanted a
6:26spec, and it recommended a spec. That's
6:28good. And then it asked me a few initial
6:30questions before then going and creating
6:33some tickets and the first map. And it
6:35created the other tickets as sub-issues.
6:37So, we kicked off with seven tickets
6:40immediately. However, only three of
6:41those tickets were takable right now.
6:44So, figure out where icon names come
6:46from, component storage schema, and
6:49palette information architecture and
6:50grid keyboard. I don't remember that
6:51one. And so, what I did was I then
6:53worked through each of those tickets in
6:56a new session. The way I did that was I
6:57just called Wayfinder on that ticket
7:00name. I did it in a slightly fancier way
7:02where I actually have a handoff skill
7:04that automatically wrote me a prompt and
7:06spawned a Claude sub-agent. But what it
7:08was essentially doing is just calling
7:10the Wayfinder skill on this map and on
7:14the specific ticket, wherever it was.
7:16Yeah, here it is. Here's your ticket.
7:17Translucent SVG geometry to path
7:20builder. And it just mentions the full
7:22ticket name. So, this is how you work
7:24through a Wayfinder map. You do an
7:26initial Wayfinder prompt just to chart
7:28the map and figure out the next tickets.
7:30And then for each ticket, you say
7:32Wayfinder with the ticket URL. So, you
7:35use Wayfinder for both, both for
7:36charting the map initially and then
7:38walking through each ticket. As you can
7:40probably see from this diagram, tickets
7:42can have different types, and there are
7:44four types. And these ticket types are
7:46actually brought into the issue tracker
7:48themselves. So, we actually have
7:49Wayfinder research, which is a ticket
7:52type. Research tickets are where the
7:54agent needs to go off and find some
7:55information and bring it back. And it
7:57usually kicks it off immediately. So,
7:59you don't actually need to watch it. It
8:01does it in a sub agent and then reports
8:02back. Prototype tickets, which are the
8:04next type here, create a prototype,
8:07which is so unbelievably in valuable for
8:11really seeing things come to life as
8:13you're planning. I've done a whole extra
8:14video on this, on how important
8:16prototypes are, and it reuses the
8:18prototype skill from that video. Some
8:20folks look at Wayfinder, and they think,
8:22"God, that's a lot of planning. Doesn't
8:23that look like waterfall?" And the
8:25prototypes are the way that you prevent
8:27it from becoming waterfall, huge amounts
8:30of low-fidelity upfront planning. A
8:32prototype is a high-fidelity way to get
8:35feedback on what you're actually
8:36building. And the fact that Wayfinder
8:38encourages you to build so many
8:39prototypes means that the output is
8:41unbelievably good. So, so far, we've got
8:43research, prototype. Obviously, there
8:45are grilling ones as well. So, grilling
8:47sessions. And this is just where you
8:49need a discussion over maybe an
8:51implementation detail, over a particular
8:53aspect of the plan. And the final type
8:55of ticket are tasks. These are things
8:57that need to be done in the real world,
9:00stuff that the agent can't quite do
9:01itself, or possibly sometimes the stuff
9:04agent can do itself, but is scheduled
9:06behind other work. One really cool thing
9:08about Wayfinder is the way that it
9:10establishes blocking relationships
9:11between tickets, because some decisions
9:13can only be made once other decisions
9:15are made. And so, what you end up with
9:18is here, we've got 14 out of 17 done on
9:21this map. So, a lot of work done, but
9:23we've still not built the skill that
9:25this whole map is built around. And once
9:27we build the skill, then we actually
9:29need to revisit some other stuff based
9:31on how the skill works and how it
9:33actually improves things. And And what
9:35you're doing a lot of the time when
9:36you're working through a Wayfinder map
9:38is going, "Okay, I've resolved that
9:39ticket. Let's see how this opens up new
9:42tickets. What has the frontier moved
9:44to?" So, then once the map is complete,
9:46what do you then go and do with it?
9:48Well, this one, because its detonation
9:50was a spec, the Wayfinder map is
9:53probably a little bit too dense to
9:55create a spec. So, what I like to do is
9:57create a spec from the map. This was the
10:00spec that I created from it, and you can
10:02see it's basically the same setup as
10:05I've had before. I literally just called
10:07to spec on the Wayfinder map, and it
10:09pulled in this enormous
10:12document with basically all of the
10:15decisions that have been pulled from the
10:16Wayfinder map into this GitHub issue.
10:19The initial draft was actually too large
10:21for GitHub's character limit, so
10:23>> [laughter]
10:24>> that kind of tells you how big it was.
10:25And from there, I turned it into tickets
10:28using my usual approach, which is to
10:30spec and then to tickets. In other
10:31words, Wayfinder fits in just in exactly
10:34the same place that Grill with Docs does
10:36in my usual approach. So, instead of
10:38doing Grill with Docs and then doing to
10:40spec into tickets, you're spending a lot
10:42more time in Wayfinder creating this
10:45enormous map, and then taking that map,
10:47turning it to spec, turning it to
10:49tickets, and then implementing each
10:51ticket, and then running code review at
10:52the end. The really cool thing about the
10:54Wayfinder setup is that the specs that
10:56it creates are so dense, and they all
11:00link back to the original decision
11:02tickets. So, you can actually go, and
11:04the agent can go and view the primary
11:06source if it's confused about anything.
11:08That was always a kind of weakness with
11:10Grill with Docs, which is that you were
11:12really relying on the spec to be the
11:14source of truth, but the spec is always
11:17just a summary of what was actually said
11:19in the meeting. Whereas now, with
11:21Wayfinder, you've actually got access to
11:23that primary source, which is amazing.
11:24So, that is Wayfinder. It's a way of
11:27mapping huge chunks of work by planning
11:30things out really in detail ahead of
11:32time. It can handle prototyping, can
11:34handle research, can handle arbitrary
11:36tasks, can handle discussions, too.
11:38Let's jump into an FAQ now of frequently
11:41asked questions that I get when people
11:43ask me about Wayfinder. The first one is
11:45"This is way too much process. This way
11:49too heavy for the kind of work that I
11:50do. When should I actually use it?"
11:52Well, the answer to this is if you think
11:54the work that you're doing can be
11:55completable and planable in a single
11:57session, then plan it in a single
11:59session. If you kind of already know the
12:00way to your destination, then there's no
12:02need to use Wayfinder because you can
12:04just path your way there in a single
12:06session and just figure it out.
12:08Wayfinder is for the cases where you
12:10have the fog of war. You'll no idea
12:12quite where to go and you just need to
12:15start and then see where you get to. By
12:17the way, I've actually been using
12:18Wayfinder for non-coding tasks. So, I've
12:20been meaning to put up a garden office
12:21in my garden and I've been using
12:24Wayfinder for that. So, it's
12:25commissioning a site survey, figuring
12:28out all that stuff, figuring out who to
12:29contact, doing all the research, finding
12:31the different firms that could build it.
12:32It's awesome. Another response people
12:33have to Wayfinder is this is SDD. This
12:36is spectrum development and I don't want
12:37to do spectrum development. I don't want
12:39to spend all this time putting together
12:41a spec. This seems bananas. Well, the
12:43way I think of specs is really just a
12:46destination for a multi-session piece of
12:49work. In other words, we have a huge
12:51task down here, let's say task number
12:52four, that we're trying to schedule over
12:55multiple agent sessions cuz it's just
12:57too big. And what we want to do is we
13:00need a spec so that we can when we get
13:02to the end figure out where we were
13:04going. That's all a spec is in this
13:06context. It's just a destination
13:07document to handle this multi-session
13:10work. And then each session is done in
13:12an implementation ticket. Also, this is
13:14people get confused when they first use
13:16Wayfinder cuz they go, "Right, it's
13:17creating some tickets. Aren't we
13:19supposed to do the tickets later?" These
13:21are kind of implementation tickets
13:23versus decision tickets. So, in
13:24Wayfinder you have decision tickets.
13:26These are implementation tickets. So,
13:28the difference between my approach and
13:29most other approaches is that people,
13:31when they get to the end of this, they
13:33will keep that spec around somewhere.
13:36For me, I close the issue containing the
13:38spec, and the spec is gone. It's gone
13:41from my repository. I rarely, if ever,
13:43refer to it again. Once the spec is
13:45present in the code, then you can just
13:47delete the spec. Whereas, people who do
13:49spec-driven development go back to the
13:51spec and edit it and modify it. There
13:54are lots of approaches to spectrum
13:55development, so I'm probably annoying
13:56someone with that, but what I'm
13:58essentially trying to say is that these
14:00specs are non-persistent. With that,
14:02folks, I recommend you go off and you
14:04chart your own awesome foggy idea. I
14:08have found Wayfinder just so liberating
14:10in that it just lets me get started, and
14:13it handles all of that difficult
14:15decision for me. I've been using it to
14:16plan courses, been using it to do
14:18engineering work, been using it to build
14:20a garden office. It is just awesome. The
14:22cool thing about it is that the
14:23destination is totally up to you.
14:25Whether you want it to create a spec
14:27that you then run through an AFK agent,
14:29which is what I do, or if you just want
14:31to it to implement the work for you in
14:33tasks, then it totally can. There is no
14:36more fun feeling than starting a new
14:38Wayfinder session, and knowing that
14:40you're going to see something awesome,
14:41but not quite knowing how you're going
14:42to get there. If you're into this stuff,
14:44and if you want to keep up with my
14:45skills, then you should check out this
14:47seven-lesson free course that I've put
14:49out on AI skills for real engineers,
14:51which is on my AI Hero site. I'll add
14:53the link below. This lets you build up a
14:55repeatable workflow that you can ship
14:56great work, and it's all built on solid
14:59software fundamentals. Thanks so much
15:00for hanging out. It is always fun
15:02filming these sessions, and I'm so glad
15:04that people are enjoying Wayfinder so
15:05much. So, cheers, pals. I will see you
15:07in the next one.

About this transcript

This page contains the full transcript of /wayfinder: Nothing is too big to plan anymore by Matt Pocock, generated from the public captions YouTube serves with the video. The transcript has 3,144 words across 446 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 choose a caption track available on the video.

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, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.