WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=Yn8h5Ip-L9c

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/Yn8h5Ip-L9c.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:00.000 --> 00:00:02.800
Everyone's obsessed with automating work with engineers.

s2
00:00:02.800 --> 00:00:07.320
And the way you do it best is to figure out where you need human review.

s3
00:00:07.320 --> 00:00:09.440
Let's say you have a process that looks like this.

s4
00:00:09.440 --> 00:00:17.160
Imagine totally pre-AI where maybe you plan some work and then you ship some work and then maybe you review some work.

s5
00:00:17.160 --> 00:00:22.600
And all of these checkpoints, all of the moments where you check the work, these are all human.

s6
00:00:22.600 --> 00:00:27.360
A lot of people's dream, I think, is to have these all translated over to AI at once.

s7
00:00:27.360 --> 00:00:29.920
But for me, that doesn't really make sense.

s8
00:00:29.920 --> 00:00:33.840
On very large pieces of work, you probably want to align with the AI first.

s9
00:00:33.840 --> 00:00:37.080
In other words, you want to be involved at the initial planning stage.

s10
00:00:37.080 --> 00:00:40.600
That's what my skills like Grill Me and Grill with Docs and Wayfinder all do.

s11
00:00:40.600 --> 00:00:45.280
And then you probably also want some human review right at the end before you ship it to production.

s12
00:00:45.280 --> 00:00:53.920
However, there are some situations where you don't need that much human review at the end because maybe it's just a refactor, just an internal documentation rename.

s13
00:00:53.920 --> 00:01:00.320
Maybe the human can just align early on, basically consider that the plan is done, and then the AI can ship it all the way to prod.

s14
00:01:00.320 --> 00:01:02.000
That is good for some cases.

s15
00:01:02.000 --> 00:01:09.920
There are also some cases where you can just specify the work that needs doing in like a single prompt, not bother really aligning with the AI, not bother grilling,

s16
00:01:09.920 --> 00:01:14.120
and just let it ship through the process, and then just review it before it goes live.

s17
00:01:14.120 --> 00:01:24.720
That can be really effective for very simple changes, very small diffs, stuff that's not going to make a big splash or a big impact, you know, just changing the color of a button or changing the name of a title or something.

s18
00:01:24.720 --> 00:01:31.800
Even things like fix a small bug here that's just a couple of lines change, you should probably shift that checkpoint to the right.

s19
00:01:31.800 --> 00:01:33.120
And that's the right language to use here.

s20
00:01:33.120 --> 00:01:40.160
You shift checkpoints left for big work, for stuff you need to align on first, and you can shift it to the right for small work.

s21
00:01:40.160 --> 00:01:46.400
The reason for that is that the smaller the diff, the easier it is to totally regenerate the code and hold it all in context.

s22
00:01:46.400 --> 00:01:53.680
So, if you have really massive work, you want to align beforehand because fixing it is really expensive when you get to a later stage.

s23
00:01:53.680 --> 00:01:54.440
So, that's the theory.

s24
00:01:54.440 --> 00:01:56.760
For big work, you shift the review left.

s25
00:01:56.760 --> 00:02:00.520
For small work, you shift the review right.
