YT YouTube2Text

YouTube2Text → Transcripts → Spec-Driven Dev Is Back. But Not How You Think • Daniel Terhorst-North & Gojko Adzic • GOTO 2026

Spec-Driven Dev Is Back. But Not How You Think • Daniel Terhorst-North & Gojko Adzic • GOTO 2026 — Transcript

by GOTO Conferences · 7,254 words · 1,144 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:13Well, hello.
0:15Well, hello
0:16what are you doing there? Hey, I'm doing
0:18wonderful. How are you doing?
0:21I'm very well, thank you.
0:23So, one of the things I'm I'm kind of
0:25really curious to
0:26get your opinion on is whether this
0:29whole spec-driven thing is going back to
0:31waterfall and whether it's negating
0:34pretty much everything we've talked
0:35about in the last 30 years or is it kind
0:39of really giving us superpowers to do
0:41fast feedback the way we never had
0:43powers to do.
0:45And so let let let's I guess we should
0:47probably start with some terms. The the
0:49thing that I've
0:51I've seen spec-driven
0:53a lot on the LinkedIn. I don't know why
0:55I'm on LinkedIn. It's just kind of it's
0:56become the new Twitter. It's where
0:58people go and just It's a bit less
1:00racist. It's what? It's a bit less
1:03racist. It's it's a bit less hatey. It's
1:05a bit more respectful.
1:07But certainly [laughter]
1:09there's still a lot of nonsense on
1:10there. Spec-driven then is where someone
1:14puts together a to-do list or a list of
1:18bullets or something.
1:20Um they say, "Right, this is what the
1:21system needs to do." And it might be
1:23prose. It might be a description of a
1:25product. It might be
1:27a bullet list of features.
1:30Everyone agree? Haiku is best.
1:32>> It could be high I mean haiku is
1:33fantastic. I I still want to see an
1:35entirely limerick-driven
1:38GenAI. I think that'd be wonderful. Just
1:40after St. Patrick's Day, I feel like
1:42that's a thing. But no no I I think I
1:44think what I'm seeing is we're using
1:47spec-driven in two very different
1:48contexts and it's the
1:50the context that matters. So,
1:52spec-driven as a one-shot is the thing
1:55that I think you're referring to and I
1:56it's just it's never going to work,
1:58right? I'm going to put it out there.
1:59It's never going to work. You can't
2:01describe a fully functional, fully
2:03complete application
2:06and give it to
2:07a bot or a swarm of bots or whatever and
2:10get a product out the other end. I don't
2:12think that's a thing. Um or at least I
2:14don't think that's the thing that I
2:15would ever want to put into production
2:17with any kind of security, resilience,
2:20availability, compliance, accessibility
2:23guarantees anywhere, right?
2:25But I think the
2:27the idea of the agent doing the coding
2:30bit
2:32in a in an iterative spec-driven world,
2:36I I haven't yet done this, but I don't
2:39see any reason why that shouldn't
2:41be a thing and be the thing that gets
2:43better and better.
2:45Because, you know, I keep hearing this
2:47all the programming's the easy part, the
2:48coding's the easy part. No, I mean it's
2:50it's the easier part. It's the more
2:52deterministic part of a very large
2:54number of parts. It's still not easy or
2:56at least if it were easy, I would see a
2:59lot more I would consider good,
3:01readable, legible, habitable code in the
3:03world and I don't.
3:05So, it's very easy to type things into a
3:08computer. It's very easy to type things
3:10in such a way that a an appropriate
3:13compiler or interpreter can run them and
3:15do stuff.
3:16Um as Martin Fowler said as far back as
3:191996, any idiot can write
3:22code for a computer that a computer can
3:23understand. Programming is about writing
3:25software that humans can understand. And
3:28I think we're a long way off any of the
3:29GenAIs, any of the frontier models
3:31writing software that a that a human can
3:33understand.
3:35But I think if we're going to go in in
3:36little iterative loops with plenty of
3:38feedback, automated feedback,
3:40prompt-based feedback, developing the
3:43spec as we go, that kind of really is
3:45BDD, right?
3:46Yeah, I get the same feeling. So,
3:48there's a couple of things I tried last
3:49few months. There's there's some tooling
3:52around this spec-driven development
3:54that's promoted by Microsoft/GitHub. So,
3:57there's the spec kit, there's the open
3:58spec that's like an open open-source
4:01version of it. There's a Although this
4:02one is open-source as well. So, we're
4:04getting an open-source version of an
4:06open-source tool.
4:07>> [laughter]
4:08>> And then there's like Kiro from Amazon.
4:10There's a bunch of other things and
4:12everybody and their dog seems to be
4:14publishing their own kind of spec-driven
4:16process now, which is a kind of a bunch
4:19of markdown files effectively
4:21implementing what what's it like
4:23the password for this is an RPI loop
4:25like research, plan, and and implement.
4:28Kind of my experience is that
4:30a lot of these things are basically
4:33creating a a sourdough starter for a
4:36process where
4:38you're supposed to take that and then
4:40heavily modify it for your use. Some
4:41I've kind of had
4:44a reasonable success teaching Claude
4:46code how to develop the code in a way I
4:48like to. Mhm. And it's producing code
4:50that is
4:51if you take me for a human then humanly
4:54developed. Um it took a lot of work to
4:57get it there and and I've kind of
4:59inspired by some of these frameworks or
5:02elements of of kind of reduced it to the
5:04amount that's useful for me and that's
5:06what I I think
5:08is going to happen is that people will
5:10take these things and
5:12reduce them or or modify them in a way
5:16that's useful for their process where
5:19if you just take Kiro or or take spec
5:21kit, you're effectively
5:24uh implementing Microsoft's process, not
5:27Amazon's process, not your process. I I
5:30think the problem with a bunch of these
5:32things is kind of the generator wall of
5:33text that nobody reads and and we know
5:35that from like last 30 years of work.
5:41I I still remember working with business
5:44analysts who would produce like a
5:45500-page Word document that the
5:47developer is supposed to read and we're
5:49getting to the point where
5:52these tools are becoming those business
5:54analysts producing 500 pages of text
5:56that nobody reads and then approves and
5:58and no wonder the result is kind of not
6:00up to spec. So, I think I I agree with
6:02you in a sense where yeah, we can we can
6:05force through small iterative loops and
6:07that would be amazing and that is
6:08actually kind of BDD and and that's
6:11where kind of all these good ideas
6:14around clarifying things come in. Um
6:17I'm concerned though that
6:19people are going to be overly
6:22enthusiastic about just
6:26browsing through walls of text and
6:28approving them and getting getting
6:29really surprised on the other end. And
6:31you know, you used the word compiler.
6:34I think the danger is where people start
6:36assuming that agents are compilers and
6:40that they can actually transform
6:42a spec into a product in a
6:45deterministic, reproducible way and and
6:47that's not what's happening at the
6:49moment. Maybe it will in the future.
6:51>> [laughter]
6:51>> Well, and I think you need to follow the
6:52money here. I mean,
6:54when you say, you know, there's there's
6:55these products coming out, these tools
6:57coming out, I I get the same the same
7:00flavor as things like the case tools of
7:02the '90s or the model-driven
7:03architecture tools of the 2000s
7:06where there's a buzz around a thing and
7:09the the the thing this time is using
7:11natural language to get a computer to
7:14write code for another computer.
7:16And if we anthropomorphize the middle
7:18computer, we can call it an agent and
7:20use the words like understanding where
7:21there is no understanding cuz there's no
7:23sentience.
7:24But it's it's exciting to then say,
7:26"Hey, enterprise senior purchasing
7:28person, you can buy this tool and this
7:31tool will be your business analyst."
7:33Right? Um assuming that the role of a
7:35business analyst is writing documents
7:38rather than understanding a problem
7:39domain.
7:40Right?
7:41>> [laughter]
7:41>> And coming in with a vast amount of
7:44pre-existing domain knowledge.
7:46Um in the same way that programming is
7:48is typing tokens into a computer rather
7:50than understanding how the stack works
7:53and mechanical sympathy and any of that,
7:55you know, good engineering stuff. So, I
7:57think from the commercial angle, I I I
7:58understand why these tools are coming
8:00into existence and
8:02I predict exactly the same commercial
8:05arc for them as happened with with the
8:07case tools, with the MDA tools, with the
8:09the various other attempts at removing
8:11humans.
8:12And and but but the the frustrating I
8:15>> [clears throat]
8:15>> you know, you I'm going to keep on
8:17banging on about this. I'm going to keep
8:18on I suspect a lot of folks like me are
8:20going to.
8:21>> [snorts]
8:21>> It's a machine, right? It's a very
8:23clever mimicking machine. Yeah, back to
8:25Emily Bender's like lovely phrase,
8:27stochastic parrot.
8:28It it the the the robots have no
8:31understanding. The robots have no
8:33semantic awareness. In particular, the
8:35robots don't care about your stuff cuz
8:36they don't care, right?
8:39So, any attempt at code quality has to
8:41come from not just defining guardrails
8:44cuz that's not hard, right? And I
8:45suspect
8:47the Venn diagram of lots of different
8:49people's quality code quality guardrails
8:51does start to converge on a circle or at
8:54least a circle with some bumps on it,
8:55you know?
8:56Uh I imagine if I were to write down in
8:58markdown files the things I care about
9:00and you were So, I don't think that's
9:02the hard bit. The hard bit is is keeping
9:06the the GenAI, keeping the agents,
9:09keeping the bots on the rails with all
9:11that written. I've been I mean, you
9:12know, you've been doing this in anger
9:14for a while. I'm sort of dipping my toe.
9:19And what I'm finding and it's prob could
9:21well be user error, could be my lack of
9:23understanding of how this stuff works.
9:25I'm giving Claude code in my case very
9:28clear rules that it stashes away as
9:31memories and it has global memories and
9:33it has project memories of specific
9:35things I wanted to do or not do.
9:37And then it immediately just carries on
9:39and does them
9:40>> [laughter]
9:41>> about 10 minutes later.
9:43And I say things like that you So, I'm
9:46I'm desperately trying not to
9:48anthropomorphize it. The model I'm using
9:50is I'm imagining it's a text-based
9:52adventure game. You know, open box, go
9:55north, get key, that kind of thing,
9:57right? Rather than you do this and
9:59having a chassis interface with it.
10:01And one of the rules I have is do not
10:03speak in the first person.
10:04Right, I don't want the robot having
10:05[clears throat] a personality.
10:07But yeah, I I'm I'm clearly failing at
10:10these guardrails, right? At defining the
10:12guardrails and and having it stick to
10:13the guardrails. And when I say, "You're
10:15not sticking to the guardrails." it
10:16goes,
10:17"Catch." Or something equally chirpy.
10:20Like, I just deleted your stuff.
10:22>> on that, here's my thought on that and
10:24I've I've kind of hit the same bump and
10:26I've actually I think I have a good
10:27solution for that.
10:29So, Uh genius.
10:32Yeah,
10:32>> [laughter]
10:32>> so so that
10:33you know, given that this is kind of
10:35supposed to be a chat, I'll give you a
10:36slightly convoluted and a convoluted
10:38kind of way to get there because I I
10:40want to give you the thinking model
10:41behind this.
10:42>> All right, all right, please say. So, I
10:43I had a couple of these very very
10:45domain-specific rules around writing
10:47JavaScript.
10:49There's a specific class called workflow
10:52and you're not supposed to mock uh
10:53certain elements of the workflow. You're
10:55supposed to use them directly in tests
10:56and when you
10:57>> All right. production. So, this is very
10:59very project domain-specific. Really
11:01project domain-specific.
11:02And I I I I I wrote markdown files and
11:07markdown files and exception to markdown
11:08files and got Claude to write its own
11:11markdown files
11:12and then occasionally it still made
11:13mistakes.
11:15And I was thinking about this like
11:17I'm using ESLint to lint things,
11:21but it's linting things that are like
11:23use tabs, not spaces, lint this, lint
11:25that. If I had a tool like ESLint that
11:28could do my domain-specific stuff,
11:31then I wouldn't have to rely on these
11:33markdown files. I could even unit test
11:34that this works well.
11:36And I I
11:39was kind of thinking about well, how
11:40would I do this? And then I I asked
11:42Claude like
11:45write a custom ESLint rule
11:48that detects
11:50a workflow class being mocked and give
11:53yourself an error of saying, "Never mock
11:56this method, use the real method
11:57instead."
11:58And then I got it to write a unit test
12:00for this and I got it to run the unit
12:02test so I can see that ESLint is with a
12:04positive and negative doing these things
12:05correctly.
12:07And from then on, I have a fully
12:09deterministic, properly automated check
12:12for my domain rules.
12:14And
12:15once the first one worked, I really kind
12:17of that started going off the rails
12:18because now I have hundreds of
12:20domain-specific rules for very very
12:23particular things that are particular to
12:25my project. I have custom HTML linting
12:28tools. I have custom CSS linting rules.
12:30I have all these things.
12:31>> Okay, okay, okay, okay,
12:33you're giving away far too much here.
12:34What you have is a software quality
12:38product, commercial software quality
12:40quality product for JavaScript that
12:42you're going to now start selling into
12:44the enterprise. What I have here What I
12:45have here is What I have here is I think
12:48this thing where if you look at the, you
12:50know, Simon Wardley's how how things
12:53transform from like a totally bespoke
12:55thing to a
12:57commoditized product, some things that
12:59kept
13:00bumping up and you identify the rules.
13:02Once you identify the rules, you kind of
13:04transform them into a hard not
13:07hard-coded, but into a coded
13:08deterministic check that's specific to
13:10your project. And really, if I was if I
13:13had that superpower 10 years ago, I
13:15would have been doing it for all my
13:16projects, but
13:18there was never enough time to like
13:20write custom linting rules. You you
13:24communicating to developers, "You
13:25shouldn't do this, you shouldn't do
13:26this, you shouldn't do this, you
13:27shouldn't do this." Or you kept it in,
13:29you know, some kind of classic knowledge
13:31in in There was never explicit. And I
13:33think
13:34one of the things I'm I'm kind of
13:35thinking about now is like I had a
13:38conversation with with with
13:40a friend last week and the metaphor we
13:43came up with is imagine you had to
13:45onboard a new developer to your team
13:46every hour because that's what's
13:48happening when you're launching an
13:49agent.
13:51What would you have to do? Well, all the
13:52rules like they shouldn't be in Word
13:54documents. You know, people should get
13:56feedback from the system in a good way.
13:58If they do something, the bumpers should
14:01be in the system automated. So, if
14:02you're coding it the wrong way, Yeah.
14:05the system should kind of fix you. And
14:07And the other piece of the puzzle came
14:08from this conference talk I I saw by
14:12somebody from Google doing the MCP for
14:15for Chrome
14:16where they realized that the error
14:18messages for humans
14:20were
14:22not as good as like the error messages
14:24needed for agents because the agents
14:26were getting stuck in a loop. They said,
14:28"Oh, you know, access denied." And they
14:29would just try 10 times instead of a
14:32human basically saying, "Access denied,
14:33let me check what's wrong with the
14:34access keys." Mhm. And they said what
14:37the what what one of the key lessons was
14:39that error messages have to instruct
14:42how to fix the problem. So, "Access
14:44denied." is not good enough. "Access
14:46denied, check your access keys." or
14:48"Access denied,
14:50um check your configuration." or
14:51something like that. So, all of these
14:53custom linting errors have proper here's
14:56how to fix me.
14:58And I think that that itself is almost
15:00at that. You you want some meta rules.
15:02You want rules for your rules, right?
15:04>> Yeah, yeah, yeah. And And And then And
15:06then kind of what's really wonderful is
15:08like this is wonderfully useful for for
15:10humans as well. It's just good. Yeah,
15:12yeah. It's It's And And I can switch
15:15from Claude to some other system
15:17tomorrow.
15:19I don't have to worry about whether they
15:20read markdown the same way. I have this
15:22in ESLint. It can't be ignored because
15:24the CI system runs linting rules
15:26separately from Claude and and Claude,
15:29you know, Claude will need to take
15:30feedback from the CI system. It can't
15:32push things out that don't pass CI. You
15:35can do pre-commit hooks. You can do all
15:36sorts of things. So, I think the
15:38guardrails
15:40as they stabilize should be pushed to
15:42existing tools, not not to And And this
15:44saves tokens so you're not wasting time
15:47on this kind of meta textual analysis.
15:49It makes it more kind of project-based.
15:52It makes it oriented towards humans as
15:54well. And I I think that's kind of where
15:56these things need to go. And And I think
15:58it's quite empowering to be able to
16:00build up these bumpers on the road. I I
16:02I like this kind of metaphor from
16:05product-led growth where they say you
16:06have kind of this lane where you need
16:07to, but you put bumpers so that users
16:10can stay in the lane.
16:12And these things are bumpers, really. I
16:13mean, if Claude goes there, you just
16:15bump it a bit back and you tell it where
16:16to go. And instead of keeping it in
16:19these markdown files, I think as soon as
16:21you figure out what the rule is,
16:24Yeah. you get Claude to automate it in a
16:26different way. And I got it to write
16:28custom CLI tools.
16:30I got it to use existing kind of tools.
16:33I got it to do grep with with like
16:35bizarre grep rules to check for CSS
16:39violations and a bunch of other things,
16:42but I think that's that's the way
16:43really. As soon as you figure out these
16:45things, push them into get Claude to
16:48push it or write write deterministic
16:50rules for itself using something else. I
16:53like that. So, for me, Getting Claude to
16:55do this for you right now, that's that's
16:57cheating.
16:58So, [snorts]
16:59So, So, I I there's two things I want to
17:01tease out there that I found really
17:02really I I'm going to take away at
17:04least. I found really useful.
17:06So, there there there's a class of rule,
17:08if you like, there's a class of
17:09heuristic that is about your code.
17:12And there is something that you could it
17:15if there were if there were a linter for
17:17it, you could write a linting rule.
17:19And if there isn't, you could probably
17:20write a linter for it and then write
17:22rules for the linter, you know? Yeah,
17:24exactly. Can Can you push that kind of
17:26linting into that domain? Something that
17:27I'm realizing, and again, this might be
17:30a learning thing for me or it might be
17:31just where the tools are going.
17:33I found
17:35rather than using Claude to write core
17:38stuff,
17:39you know, I tend to still write core
17:41stuff. Maybe I'm not good enough with
17:43Claude yet or whatever on or the
17:44tooling. I know we're both talking about
17:46Claude cuz we're both using it, but um
17:47other other LLMs are available. Other
17:49frontier models are available.
17:51But I find what what I found is rather
17:54than saying, "I'm going to use Claude
17:55for this."
17:56it's more I'm using it for lots and lots
17:59of little itch-scratching things that I
18:02would never have got around to
18:03otherwise. So, it's like I'm not
18:06designing I'm not using it to build a
18:08car, but I'm using it to make the
18:10headrest a bit more comfortable or to
18:13give me a little button on the side that
18:14automatically moves things backwards or
18:17to have memorized seating position for
18:20different drivers or just stuff that I
18:22could probably get around to, but I
18:24haven't got time for, Yeah. Yeah. It
18:26seems to really fit neatly into that
18:29niche of if you're going to look at the
18:31Neil Nicholson the purpose alignment
18:33thing, things that are non-critical,
18:37right? They're non-differentiating, but
18:39they're quality of life things. And I I
18:41I feel like that's a really nice right
18:43now at their level of maturity, a really
18:45nice kind of niche for them. Absolutely.
18:48And I think you know, when when you when
18:49you look at something like this, like I
18:50always remember in in discussions like
18:52this, there was
18:54You probably worked on billing systems
18:55that reconcile banking information. I
18:58don't know
18:59>> [laughter]
19:00>> What gives you that idea?
19:02Realistically, realistically, you know,
19:04like money shouldn't be lost between
19:06here and here. Right. I But But somehow,
19:11you know, we need to build
19:12reconciliation systems to to make sure
19:14that the money is not lost or or where
19:16it's lost, to to find how much is lost.
19:18No, there's an entire You need to
19:20understand there's an entire domain
19:22language around this.
19:24So, So, in in banking in in settle So,
19:26settlements in banking for those who
19:28don't know, basically lots and lots and
19:30lots of counterparties all trade with
19:32each other and there might be millions
19:34of trades in a day worth trillions of
19:36dollars and at the end of the day we've
19:37got to figure out who owes who what,
19:39right? And we could just make everyone
19:41make all those payments and that would
19:42just take forever. Or we add up all the
19:44money Goiko spent, all the money Daniel
19:46spent and we figure out that actually
19:48Goiko just owes Daniel a few million
19:49euros or whatever. And that's called
19:51netting. So, settlements and netting.
19:52And as you say, this is generally across
19:54lots of independent stovepipe systems
19:56that all have slightly different
19:57rounding rules or currency rules or
19:59storage rules or
20:00>> A leaky leaky pipe with the water
20:01leaking.
20:02>> pipe. And and the the
20:04>> [laughter]
20:04>> and the the language the domain terms
20:06for this is breaks and fails.
20:09And breaks and fails are your two types
20:11of things going wrong. One is where the
20:14trade didn't happen at one end or the
20:16other, but the other counterparty thinks
20:18it did. And the other is where is the
20:21leaky pipe. Is logically it's all
20:23happened. Semantically it's all
20:25happened, but somewhere we're missing 84
20:28euro cents, right? And we don't know
20:30where they've gone.
20:32>> [laughter]
20:33>> Or you know, 84 million euros. Or 84
20:36million euros, you know, I mean, yeah,
20:38between friends, right?
20:39>> Yeah, let me say 84 million, so
20:42So so so you know, I I remember us
20:44working with this client where the
20:46business users were insisting that they
20:49can't do any automated testing
20:52for reasons and and needed to do like
20:54human review and things like that. And
20:56actually, I mean,
20:58some of it made sense because some of it
21:00needs like human human you know, like
21:04domain knowledge reasoning, whatever.
21:06But 99% of a test
21:10had nothing to do with that evaluation
21:12moment.
21:1399% of the time spent in a test was
21:16loading the data, cleaning up the data,
21:19unloading the data, re-restarting the
21:21the thing, restarting the Those were all
21:23kind of it's not quality of life, but
21:26small chunks, small buttons that that
21:28actually got the thing moving. And these
21:31were all deterministic parts. I remember
21:32kind of when we built a data loader from
21:34Excel
21:36that allowed them to prepare a test case
21:38without opening five different
21:40applications and kind of registering
21:43different things and approving and and
21:45pushing them through the pipeline, just
21:46you know, get everything through. They
21:48were amazed. And and I think that's
21:50where again, there's a lot of benefit
21:52for this even if you don't want to get
21:54it to do the the like the core core bit.
21:56There's so much stuff around it
21:59that
22:00needs to work,
22:02but is not
22:03the differentiating factor. It's not
22:06the the kind of human making this
22:08evaluation. It's not your key domain
22:11knowledge. It's just like loading Excel
22:13into five databases.
22:15Well, I mean, I can give you an example
22:17of where I I was clawed one time. But
22:20actually, before I do that, I just want
22:21to pick out again something that you
22:22mentioned. I think it's I really want to
22:23kind of pull it out and give it a name.
22:25I feel like you're talking about the
22:26difference between semantic validation
22:28versus like syntactic or structural
22:30validation. I mean, you say it's just
22:32validation. I'm talking about how you
22:33have you have kind of you know, stuff
22:36that
22:38uh is is deterministic and and and is a
22:40side quest or a you know, a supporting
22:43task that
22:45basically just can't be done in in a
22:47reasonably good way and we can farm it
22:49out to machines. Yeah. Yeah.
22:53And and well, I I as opposed to stuff
22:55that needs a human in terms of domain
22:57knowledge or experience or context or
22:58whatever. So, so quick sidebar, I was at
23:01a large American bank many years ago
23:04and they had a one of their core again,
23:06but perhaps attached to the settlements
23:07process was a system that would
23:10it it used to take I don't know,
23:11millions and millions of trades and put
23:12them into lots of buckets, couple
23:15hundred buckets. It's all it's one of
23:16the banks in America, is it?
23:18It could well be one of the largest
23:20banks in America, yes.
23:21>> [laughter]
23:23>> And and they had a large complicated I
23:26think it was Oracle PL/SQL thing
23:29that would go through, you know,
23:30bajillions of rows. Okay.
23:33That's
23:34>> [clears throat]
23:34>> And and and for each row each record it
23:37would put it into one of these buckets.
23:39And this process had gradually over time
23:41had created more and more rules. And the
23:43rules were literally if, you know, this
23:45this field has this value in it and this
23:47field matches this and da da da da da da
23:48da da da da da dozen of these things.
23:50And so many many hundreds of checks.
23:53And if it also failed, suddenly it was
23:55like a catch-all bucket at the bottom.
23:56And it would do this big if else if else
23:58if else for each of several bajillion
24:00records.
24:01>> easily hide like if name equals Goiko,
24:03then you know,
24:04>> Easily. Absolutely. You know, let give
24:06me the rounding error into my offshore
24:07account. And this thing, you know, had
24:10gradually crept up, you know, from 20
24:12minutes to an hour to two hours to and
24:14it was now it had gone from nine hours
24:16to And the problem is going from nine
24:18hours to 10 hours is a lot more serious
24:19than going from seven hours to eight
24:21hours because going from nine hours to
24:2210 hours means that if I start at the
24:25close today, it won't be finished by the
24:28open tomorrow. And suddenly it was this
24:30big deal.
24:31And it was a Friday afternoon and we'd
24:33basically finished what we needed to do.
24:35And I said to a couple of the folks, I
24:36said, "I don't understand this domain
24:38model at all. My my PL/SQL is pretty
24:40good, right?
24:42Uh I'm going to try Well, I looked at
24:43your stored proc and it's basically the
24:45same thing over and over and over again.
24:47Guess what, right? Copy paste stuff.
24:49And I feel like that is not what stored
24:52procs do. Stored procs do set-based
24:55operations, right? Relational databases.
24:57Whereas a language like Python or Java
25:00or whatever is really really good at
25:02sequences of things.
25:03So, I'm going to I'm going to just as a
25:05as a Friday afternoon thing, we'll
25:07timebox it to two hours.
25:09We will pull out all of the records,
25:12apply the same logic, again, guess what,
25:14using graph and set and or can whatever
25:16to pull out the logic bits and turn it
25:18into some Python code.
25:20Turn it into the equivalent Python code
25:22and run it.
25:23And it turned out that the the Python
25:26code you know, we tested it against the
25:28last several days worth of data and the
25:30buckets that it all ended up in. So, it
25:32wasn't test-driven or anything. It
25:33literally it was just a spike. And this
25:35spike got the right records in the right
25:38buckets for all of the data
25:40except that it was now taking something
25:43like 11 seconds
25:45instead of nine hours.
25:47>> [laughter]
25:48>> Because it was the right tool. And
25:50again, it was it turns out it was a
25:51two-hour fix, right? With without any
25:52robots. No one in that team had enough
25:56breadth across more than one technology
25:58and no one outside of that team cared
26:00enough about going in and helping them
26:02to do this.
26:04Um I then went away. Someone whose
26:06Python was much better than mine made it
26:07like eight seconds.
26:09And then someone else migrated that to
26:12Java and then that ran in like three
26:14seconds.
26:15And the brilliant thing with a
26:17three-second job as opposed to a
26:18seven-hour job is that you can then put
26:20a web app in front of it cuz it's only
26:22reading. It's only a read lock.
26:24And you can put a web app in front of it
26:26and now hand it to the audit teams
26:29because it was an audit job and they
26:31could now run this thing real time
26:34intraday
26:35whenever they liked. It was no longer
26:37part of the nightly run book. So, I
26:39accidentally [laughter]
26:41took this job out of the run book and
26:42turned it into an app. I mean,
26:43ironically you know, you can run Java
26:46Java stored procs in Oracle. Like you I
26:48think that that was that was impossible
26:51for a long time. So, I don't know when
26:52you did that, but
26:54I mean, this is still not where I would
26:56want to have held my I'd love to know
26:58about the version control or source
27:00control of Java code embedded in Oracle.
27:03It's where you know, we we we did we did
27:05like something
27:07something where the code was in in
27:09version control, but there was a
27:10publishing process that was updating the
27:12database. So, it's like it's a
27:14deployment thing. It's uh but Yeah, but
27:16you know, I mean, the the yeah, but the
27:18the the the point really of all that is
27:20A, you don't need robots to do this. But
27:23you know, I I I again, I'm looking at
27:25LLM as you know, it's a tool, right?
27:27It's neither necessary nor sufficient.
27:29It can be useful in the right context.
27:31And if you put it in the right context
27:32and you use it well, it can be a massive
27:34multiplier. Most cases I'm seeing most
27:37content most environments I'm seeing,
27:39they're not using it well. They're using
27:40it as a blanket edict. Yeah, you have to
27:42rub AI into this otherwise you don't get
27:44your bonus.
27:45And they're not being given the support.
27:46They're not being given the learning
27:47tools. And they're not being given that
27:49tinker time. I mean, the thing about
27:50being independent, I'm both independent,
27:53is you can decide Goiko as Goiko's boss
27:56can decide to give Goiko a couple of
27:59days just to muck around with this and
28:01see what happens.
28:02Right? And I would argue that everyone's
28:05boss should be giving them a couple of
28:06days to muck around with this and see
28:07what happens cuz that's how learning
28:08works,
28:10And they're not. They're being told,
28:11"Whatever you're delivering right now,
28:13carry on delivering that and rub some
28:15co-pilot into it or whatever.
28:17Um other you know, otherwise otherwise
28:19there's going to be trouble. And also
28:21double your throughput."
28:23And
28:24>> [laughter]
28:24>> And none of this makes sense.
28:27Mhm.
28:28So, I I I kind of you know, there's
28:30there's a couple of things there. I
28:31think people generally don't
28:35accept [clears throat]
28:36like how do you put this?
28:38I think
28:39our industry is kind of obsessed with
28:41automation. I mean, by by by by the
28:43definition, we're obsessed with
28:45automation. And I think people think
28:47that automation makes things better, but
28:48it doesn't. Automation makes things
28:50faster.
28:52Well, I see Does it make it better or
28:54worse? It just makes it faster. I want
28:55to unpack this this this obsession
28:57thing. I There's a There's a certain
29:00level of executive certain level in
29:01organization that's obsessed with
29:03cheaper. Yeah. And the association is
29:05that automation means cheaper
29:08>> [clears throat]
29:09>> and we can get rid of a bunch of
29:10headcount.
29:11I think the
29:12There has I think we we we we like we
29:14like automation, but it's you're
29:16absolutely right. It
29:18It doesn't just make things faster. And
29:20this is the really important thing with
29:21automation. Automation is a trade-off.
29:24It makes [clears throat] things faster
29:26and it makes them the same.
29:29And when I say the same, that means if
29:32you got the description or the
29:33definition of the thing wrong,
29:35it will continue to be wrong every time.
29:38If you got it absolutely right and the
29:40world moves on and it becomes wrong, it
29:43will continue to remain wrong every
29:45time. So, as soon as you automate
29:47something, you lock it in, right? You
29:49lock it into whatever. If the thing
29:51you're locking it into is stable, is
29:54stable across time, is stable across the
29:56portfolio, whatever, brilliant. So, I
29:58build automation, path to live
30:00automation is a fantastic idea once your
30:03product is reasonably stable.
30:05You know, if the tech if the tech stack
30:06in your product is is volatile, is
30:08changing all the time,
30:10any effort you spend automating, you're
30:12going to throw away tomorrow to automate
30:14something else.
30:16So, that ties back really nice with with
30:17the kind of, you know, tooling thing we
30:18were talking about. You know, you start
30:20with your rules, you kind of clarify
30:22them in markdown. Once they become
30:24stable enough,
30:26you automate them using a linter, or you
30:28automate them and and you were talking
30:29about this kind of semantic and and
30:31syntactic and things like that. And I
30:32think, yeah, that was for syntactic
30:34stuff, this is kind of trivially easy.
30:36For semantic stuff, I I I think we still
30:39have a way to go to to, you know, have
30:41tools that semantically validate things.
30:43But, I think it's not it's not beyond
30:46implementing something that checks like
30:48semantically this needs to be in this
30:50way. We just need to kind of figure out
30:52what the rules are. And
30:54I I I think
30:55quite often on a semantic level,
30:58people don't have stable rules. These
31:00things evolve. They're kind of broad
31:01guidelines uh that are in people's heads
31:04that evolve. And And that's the
31:05interesting part, like how would you
31:07express that in a meaningful way?
31:10Yeah, I think And something else I I'd
31:12actually wanted to ask your advice on
31:14really, ask your Again, you've been
31:15around this You've been sort of
31:17tinkering with this more than I have.
31:19So, we've been talking about the rules,
31:21the the heuristics, the the guardrails
31:24for the product, right? Whether that be
31:26the code base, the whatever it is.
31:29I'm the the the thing I keep on tripping
31:31over at the moment is the rules, the
31:34heuristics for the
31:36bot itself.
31:38So, I'm saying to Claude, this is how I
31:40want you to operate.
31:43This is how I want you to do things.
31:45Phrases that don't don't refer to
31:47yourself in the in the first person.
31:49Don't you know, don't come back with
31:50chirpy, you know, whatever. Act like a
31:53computer cuz you're a computer, sort of
31:55thing. And it says, "Okay." And then it
31:58does start acting like it for about 5
31:59minutes. Then it says,
32:00"I think that's a great idea." And I'm
32:02like,
32:03you have a memory to not use the first
32:05person. "Oh, good catch."
32:09>> [laughter]
32:09>> Uh
32:10and so, I like I I'd love to figure out
32:13a way to
32:14to keep that stuff honest cuz it's it's
32:18clearly not. So, I I I do some of this
32:21through a pre-command and post-command
32:24tools. So, for example, I I wanted to
32:28never use a directory change in a
32:31command in the same kind of CD something
32:34and then something else because then it
32:36can't check what's been pre-approved in
32:38the other command and it keeps asking me
32:40for stupid stuff to approve that I've
32:42already approved. So, I've I've got it
32:44to build a pre-tool use hook that
32:47detects and and
32:49and detects if CD is at the start and
32:51and basically gives it a
32:53message saying don't combine CD and this
32:55thing, do something else. So, it would
32:57be reasonably simple to get it to write
33:00a command that says, you know, if you
33:02catch any of these words like great or I
33:05or whatever,
33:06just, you know, don't don't reply
33:10in like
33:11anthropomorphized
33:13>> Yeah. Yeah. And and yeah, it it's it's
33:15kind of So, I think anything anything
33:18that you can define in a reasonably good
33:20way, I think you can get it to write a
33:22tool for itself to to kind of
33:23self-correct. And and I would punish it
33:27by by doing this. You know, waste more
33:30tokens, but tokens are not the problem,
33:32your time is. And and
33:34the more it can review itself, the less
33:36you have to review it. And I think
33:38again, that, you know,
33:40all these tools now support hooks.
33:42Claude supports hooks for lots and lots
33:43of different things. Get it to kind of
33:47respect its hook get it to respect its
33:49its commands its instructions by hooks.
33:52But then again, they're bytecode in
33:54Claude, so it's not beyond them to kind
33:56of bytecode something that stops
33:58respecting hooks, but
34:00we'll see.
34:02Fantastic. So, okay, so we we started we
34:04started on on spec driven. We we
34:07discovered a bifurcation there. We're
34:09like spec driven meaning one shot versus
34:11spec driven meaning iterative.
34:13Then we got on through heuristics into
34:15rules into linting
34:18checks, which that's I I I really like
34:21that. That's that's my takeaway. That's
34:22my back pocket takeaway.
34:24And then we were thinking, okay, what
34:25would the lint
34:27uh analog be for a bunch of other use
34:29cases,
34:31up to and including pre and post hooks
34:34for for Claude itself.
34:36Uh um the idea that that
34:38perhaps right now with the level of
34:40maturity we've got, we should be using
34:42it much more for the the quality of life
34:44things than the
34:46than the core domain itself, but who
34:48knows? I think it changed it enough to
34:50to do like core domain the way I want to
34:51do it, but it took a few months to get
34:53it to do that. Yeah.
34:55>> So,
34:56um and and you know, the nice thing
34:57about it being a robot is that once you
34:59do it for one project, you just kind of
35:02scale it to other projects as well. And
35:04that's what makes it really powerful
35:05because I have I'm I'm currently working
35:07on two products, and I'm kind of
35:09effectively because it evolved on one
35:11product, I was able to just apply it on
35:13the other product without
35:14having to show it. But, I I do think it
35:17does take
35:18it does take quite a quite an investment
35:20in time to to train it to work properly.
35:22And okay, so so you said we're wrapping
35:24up. Where where would I look for that?
35:26So, I'm I'm potentially I'm advising
35:29some folks on on exactly this journey.
35:31You know, some higher-up has said adopt
35:33blah.
35:34And rather saying, you know,
35:36push back or adopt it blindly, how do we
35:38do this in a sensible way that's going
35:40to be sustainable, realistic, work for
35:42us?
35:43What would a good place cuz there's so
35:45much noise out there. There's so little
35:47signal.
35:48>> is publishing their own frameworks. I
35:49think Yeah. I I I really think about
35:51them as sourdough starters. So, any of
35:54them is going to be equally bad for for
35:56what you need to do.
35:58But,
35:58>> [laughter]
35:58>> I I if you take something that works for
36:00somebody else and quickly iterate on on
36:03removing the parts that don't work for
36:04you,
36:05and and and noticing really really
36:07noticing where
36:12you're hitting your head or where the
36:15machine is constantly making the same
36:16mistakes, and then you're quickly
36:18evolving rules around that for your your
36:20stuff. I I think that's the way to go.
36:22And And from that perspective, it really
36:24doesn't matter if you adopt like one of
36:27these B mod or or or was it Spark or any
36:31of these other kind of published
36:32frameworks, or just like start yourself
36:35with automating bits and pieces of your
36:37process. I think that's also a good
36:38thing where
36:40spotting parts in your process where
36:42you're wasting time.
36:44Yeah.
36:45And automating parts of that. So, at the
36:47moment, I'm And I'm not saying copy
36:49this, I just want to kind of give you an
36:50example of what I have is I have a
36:54command that helps me write like a
36:57commander's intent template and reminds
37:00me to do goals, anti-goals, constraints,
37:03and things like that. Just because it
37:05can do some code analysis and and and
37:07figure out similar things in the past.
37:08Anti-goals are probably going to be the
37:10same for kind of a bunch of things we
37:11do.
37:12Then I have a command that does
37:15from that impact analysis on an idea
37:18that goes through the code base and
37:19figures out all the places that need to
37:21be changed without changing them and and
37:24figuring out what what we're missing in
37:26terms of functionality, demo pages, UI
37:29design, APIs, and and just documenting
37:32these things. Then I have another
37:33command that from that like creates an
37:37iterative implementation plan where a
37:40user can actually see observe results at
37:43every every step and do change no more
37:46than three files at every step. So, I
37:48can review it quickly through Git diff.
37:50And then I I basically just let it run.
37:53So, I think this whole kind of a set
37:54that the mental model for this is the
37:56RPI loop like research, plan, and and
37:58invest implement, but people do it in in
38:01different ways. Some people like seeing
38:02pseudo code. Some people like seeing
38:05Like, for example, with this product I'm
38:07working on now, I get it to do static
38:10HTML pages that I can review.
38:12Like key moments
38:14of the application, how the application
38:16changes, and then I can play on the
38:17static pages. Once that's
38:19to my satisfaction, I just let it
38:21implement.
38:23Um so, diff- different products,
38:24different things will need different
38:26approaches, but I think kind of figuring
38:28out what your process is and kind of
38:29codifying it is really important. And
38:32then starting with any of these kind of
38:34frameworks
38:36is is perfectly fine as long as you
38:38approach it as as like we are very
38:40quickly going to kind of modify it to
38:42our stuff.
38:43That sounds like some
38:45solid gold advice to leave our listeners
38:47with.

About this transcript

This page contains the full transcript of Spec-Driven Dev Is Back. But Not How You Think • Daniel Terhorst-North & Gojko Adzic • GOTO 2026 by GOTO Conferences, generated from the public captions YouTube serves with the video. The transcript has 7,254 words across 1,144 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.