Am I holding this right?

Oh look, another session about AI! Your social media feeds are already flooded with agentic this and spec-driven that; one-shot rewrites and everyone-is-a-programmer-now.

I want to share some models and metaphors that are helping me make sense of this new world: how riding a fixie isn’t really cycling, why Claude is just a mercenary contractor, how test-first is the new TDD.

Beyond the hype, we should still care about iterative development, bounded contexts, decent tests, intention-revealing names. And AI is not going to replace junior developers any time soon, instead it is something... other. I don’t use genAI to go faster or produce more, at least not primarily. Instead it is helping me do 'adjacent' things, allowing me to focus where I want to rather than where I otherwise would have to.

My goal is to give you a more nuanced take on generative AI, to help you cut through all the noise and get actual work done.

Transcript

1. Am I holding this right? Daniel Terhorst-North Craft Conference 2026
2. Hello, my name is... I've been writing code for *cough* years Started using Claude in February 2026 (about 100 AI years ago) Before this I was very much in the Stochastic Parrot camp Spoiler: I still am! LLMs can be 'just a token extruder' and still have enormous utility I am not... an enterprise a big team even a small team! I'm just a boy, standing in front of an LLM, asking it to do some dumb stuff
3. Am I holding this right? Everyone was telling me how they were: producing code faster producing more code coding all night! (ok, using agents) I wasn't sure I had any of these problems! I have written lots of software but I don't care about code I care about the impact that the code has I care about how easy the code is to change
4. OK let's see what this puppy can do!
5. Fit the first Can you convert this jq script into nushell for me? Why yes, yes I can, and you look fabulous today! OK this was easy, but I was impressed with the details It wrote an idiomatic nu script It ran through a bunch of test files, identified and fixed its own bugs I had become LLM-curious
6. Fit the second Enough with the sucking up, I need an auto-centring autocue Sure! I want you to use TDD I can do TDD I want you to use describe - it I can use describe-it With an active verb phrase, and describe the behaviour not the code I can do that And it did!
7. Fit the second Enough with the sucking up, I need an auto-centring autocue Sure! I want you to use TDD I can do TDD <- liar liar pants on fire I want you to use describe - it I can use describe-it With an active verb phrase, and describe the behaviour not the code I can do that And it did! (sort of)
8. Fits the third, fourth and fifth Then I suddenly had a burning platform: I wanted to teach my classes online but I only had hand-written 'slides' I have over 40 teaching topics Each class covers about 10 topics Each class has a different 10 topics So I would either have 40 tiny decks or a ton of duplication or... I could use Markdown! Claude had introduced me to Marp which seemed like a good fit but first I had to make sense of all of my training material
9. My training material
10. And some more
11. Sidebar: Three Criteria for Paradigm Shift 1. An existential threat 2. No other options 3. Information 'I make friends with a company and then wait for the first two' — Eliyahu Goldratt
12. I needed a plan! Claude, I'm going to need you to analyze over 20 workshop PDFs with only a mind map and a TBR crash course! synthesize related content and bring it together review for missing topics, identify tangents and sidebars design two 1-day classes that work as a 2-day class structure the topics, write the teaching arc, add bookends, split PDFs by topic Not gonna lie, I was genuinely amazed at the results
13. Going sideways But now I needed some new tooling! Something to compose topics into a single workshop deck A HUD to present the deck for a busy and distracted trainer A few other odds and ends, like cool text morphing!
14. Meet the tools We love cool text morphing! deck-compose: aggregate items into a schedule deck-present: bells-and-whistles session and topic timers, sliders and scrubbers deck deck-compose deck-present excalidraw-build
15. Each in its own repository npm install @goalwards/deck deck deck-compose deck-present excalidraw-build
16. Notes from the field
17. The good... Frontier models like Opus are amazing at synthesis and understanding This does not make them 'intelligent', just extremely useful They excel at Family Fortunes-style questions (en-us: Family Feud) 'We asked 100 web developers how to centre an <img> tag on a page...' while having zero understanding of what a page, a tag or even a web developer is! Plot twist: most of programming is just recreating existing code! James Noble and Justin Biddle called this Postmodern Programming in 2002(!) Ivan Moore and Nat Pryce ran a scrapheap challenge at OOPSLA in 2005 Agentic coding is just Scrapheap Programming-as-a-Service!
18. The bad... There are no rules, even WHEN YOU USE CAPS! Everything is a heuristic Your LLM will never consistently: use the tools you specify rather than fallback defaults write half-decent code, however carefully you articulate this LLMs are terrible at self-regulation or self-correction You can mitigate this with explicit self-checking LLMs are suckers for sunk cost fallacy It will keep going until you rein it in or run out of gas It is great at kaizen but terrible at kaikaku There is no such thing as 'You are an expert software tester...'
19. The ... different Claude prefers test-first and one-shot chunks to 'strict' TDD and [shocker!] I'm OK with that It's like riding a fixie! bigger steps 'feel weird' at first but then you do not want to go back! Adjacency is my new superpower I don't go faster with Claude, I go broader I have an army of squirrels to satisfy my ADHD* Sometimes the squirrels discover a whole new product! Welcome to the age of Adjacent Engineering
20. Wrapping up
21. An LLM contains multitudes Context but no memory Kaizen but no kaikaku Understanding but no awareness Ability but no agency (ironically!) Intelligence but no wisdom or discernment Personality but no personhood The genAI is out of the bottle!
22. Things you will forget even though I told you Everything is a heuristic, there are no rules If you want a rule, move it outside the LLM (h/t Gojko Adzic) It is equally confident wrong as right 'Are you gaslighting me right now?' I think you mean 'gaslighting' Claude isn't a person! Even Claude knows this, but it will keep 'forgetting' Adjacent Engineering will be the real win here Teach your teams (and yourself!) to think wider rather than faster We don't need more software, we need better ideas LLMs let us have ideas faster!
23. So am I holding this right? I can't be sure, but it certainly is an adventure!
24. So am I holding this right? I can't be sure, but it certainly is an adventure! https://goalwards.co daniel@goalwards.co linkedin.com/in/tastapod @tastapod@mas.to @tastapod.com