The new spec-driven workflow is a mess... — Transcript

Full transcript

1. 0:00Sitting alone in a dark room, tinkering
2. 0:02at your code was by far the best thing
3. 0:04about the software developer job. And
4. 0:06naturally, the world has this
5. 0:07unexplained need to destroy everything
6. 0:09that's quiet, satisfying, and can
7. 0:11actually make you happy. So, they came
8. 0:13up with the idea of pair programming,
9. 0:15where instead of one person thinking
10. 0:16carefully and moving forward, you now
11. 0:19have two people negotiating every
12. 0:20insignificant coding decision as if
13. 0:22their life depends on it. Then, when
14. 0:24this didn't gain traction, they came up
15. 0:26with test-driven development. This
16. 0:28usually sounds reasonable until you
17. 0:30realize you are now writing code for
18. 0:32code that doesn't exist yet, testing
19. 0:34behavior you haven't even fully
20. 0:35understood, and then reshaping
21. 0:37everything just to satisfy the test you
22. 0:38wrote 5 minutes ago.
23. 0:39>> [music]
24. 0:40>> Then, we were all forced to do scrum,
25. 0:42which meant slicing your work into
26. 0:43artificial fragments that management can
27. 0:45manage and then wasting most of your day
28. 0:47in daily stand-ups, sprint planning,
29. 0:49backlog refinement, sprint reviews, or
30. 0:51retrospectives. But, I'm afraid the pain
31. 0:54is not going to end here because a new
32. 0:56methodology is emerging, and this one
33. 0:57actually kills the little joy we
34. 0:59developers had left. Spec-driven
35. 1:01development is the new thing every vibe
36. 1:03coder and wannabe product manager is
37. 1:05excited about. Despite our complaints,
38. 1:07the chances are some of us will be
39. 1:09forced to work in this new setup sooner
40. 1:11than we think. So, in this Monday
41. 1:12morning review, we'll look at what
42. 1:14spec-driven development actually is and
43. 1:16how it can impact our future.
44. 1:17>> [music]
45. 1:18>> Then, in the second part of the video,
46. 1:20we'll look at the actual supporting
47. 1:21evidence for this methodology in
48. 1:23real-world projects because to quote a
49. 1:25recent article I read, "The reports of
50. 1:27code's death are greatly exaggerated."
51. 1:29At [music] score, the idea of
52. 1:31spec-driven development is really
53. 1:32simple. Instead of just sitting down and
54. 1:34building something, you first write a
55. 1:36very precise machine-readable document
56. 1:38that describes what you want to build,
57. 1:40why you want to build it, and what
58. 1:41success looks like. You can probably see
59. 1:43that this is already sounding terrible
60. 1:45because the one thing developers hate
61. 1:47writing more than tests is
62. 1:48documentation. Then, crucially, this
63. 1:51spec becomes the source of truth for
64. 1:53your application. In STD, the spec is
65. 1:55apparently alive. In traditional
66. 1:57development, specification is most of
67. 1:59the time an afterthought. You write it,
68. 2:02you feel good about yourself, then you
69. 2:04file it somewhere and let it deprecate
70. 2:05gradually because nobody has the time to
71. 2:07keep documentation in sync with the
72. 2:09actual product. These were the good old
73. 2:12days when writing code was still
74. 2:13considered a valuable skill. Now, thanks
75. 2:16to AI code generation, the barrier to
76. 2:18writing code has collapsed to almost
77. 2:19nothing, and the industry is quickly
78. 2:21getting crowded with vibe coders who are
79. 2:23more than excited to celebrate the death
80. 2:25of the real software engineers. But,
81. 2:27since vibe coding is usually a disaster
82. 2:30waiting to happen, spec-driven
83. 2:31development is now being formalized by a
84. 2:33lot of big companies. One of the biggest
85. 2:35efforts on this front is GitHub SpecKit,
86. 2:37which allows you to focus on product
87. 2:39scenarios and predictable outcomes
88. 2:41instead of vibe coding every piece from
89. 2:43scratch. So, instead of just vibe
90. 2:45prompting your AI assistant and hoping
91. 2:47for the best, you now get to act like a
92. 2:49professional in the CLI and walk through
93. 2:51four very official-looking stages. You
94. 2:54start with a specify step, where you
95. 2:56describe what you want with as many
96. 2:57details as humanly possible.
97. 2:59>> [music]
98. 2:59>> Here, you're expected to define user
99. 3:01journeys, business requirements, success
100. 3:03criteria, and pretty much everything
101. 3:05else used to figure out while building
102. 3:07the actual thing. Then, in the planning
103. 3:09step, you pretend you still care about
104. 3:11architecture. This is where you define
105. 3:13your tech stack. If you happen to know
106. 3:15anything about tech or stacks, you write
107. 3:17up your dependencies, and you design
108. 3:19your system. This is interesting because
109. 3:21the entire premise of AI coding is that
110. 3:23you don't have to think about these
111. 3:25things too much anymore, but now you're
112. 3:26formalizing them even more than before.
113. 3:29Then, [music] the task step is where you
114. 3:30break everything into neat little chunks
115. 3:32of granular work items, dependencies,
116. 3:34and acceptance criteria. Basically,
117. 3:37these are the old Jira tickets now
118. 3:38defined in plain text in a markdown
119. 3:40file. Then, finally, in the
120. 3:42implementation phase, you sit back,
121. 3:44relax, and let the AI one-shot your
122. 3:46million-dollar startup idea. One of the
123. 3:48biggest red flags, in my opinion, is
124. 3:50that this model expects us to front-load
125. 3:52our entire thinking and planning in the
126. 3:54specification phase. In my experience,
127. 3:56you usually learn about problems and
128. 3:58edge cases while actively working on
129. 4:00them. With this model, we are expected
130. 4:02to understand [music] it completely
131. 4:03before writing a single line of code,
132. 4:05and this is by far the biggest fiction
133. 4:07in software development.
134. 4:08>> [music]
135. 4:09>> However, it somehow gets worse thanks to
136. 4:11BIMAD, or the breakthrough method of
137. 4:13agile AI-driven development. This is a
138. 4:15newly proposed methodology where,
139. 4:17instead of writing code, you are
140. 4:19orchestrating a team of AI agents with
141. 4:21various skills and responsibilities, all
142. 4:23communicating between them through
143. 4:24various markdown files. Your job is to
144. 4:27sit in the middle and make sure these
145. 4:29agents don't completely lose the plot,
146. 4:31all while building new billion-dollar
147. 4:33startups every 12 hours. The natural
148. 4:35conclusion of this trajectory is that
149. 4:37you are now expected to manage a system
150. 4:39that writes software for you based on
151. 4:41documents that try to approximate
152. 4:42reality before reality has had a chance
153. 4:44to contradict you. This is where tools
154. 4:47like Hero AI come in, which basically
155. 4:49functions as a digital babysitter for
156. 4:51your fleet of AI hallucinations. But, if
157. 4:54you are not ready yet to change your job
158. 4:55description to prompt architect or
159. 4:57system orchestrator, the good news is
160. 4:59that a lot of the things I mentioned are
161. 5:01really far-fetched, and the software
162. 5:02engineering job is not as dead as some
163. 5:04would like you to believe. A few days
164. 5:06ago, I read a really good article where
165. 5:08the author starts with a simple but
166. 5:10uncomfortable observation.
167. 5:11Specifications written in English only
168. 5:14feel precise [music] until you try to
169. 5:15implement them. You think you've written
170. 5:17a perfectly precise spec until you
171. 5:19realize Live Sync doesn't explain what
172. 5:21happens when two users in different time
173. 5:23zones delete the same paragraph during a
174. 5:25Wi-Fi flicker. These are the moments
175. 5:27when the prompt experts are usually
176. 5:29realizing that their one-sentence
177. 5:31requirements, which worked well when
178. 5:32building a to-do app, isn't enough to
179. 5:34cover all the edge cases of real-world
180. 5:36projects. Code emerged because the human
181. 5:39brain can only juggle so much until it
182. 5:41runs into cognitive overload. Code helps
183. 5:43humans compress complexity into
184. 5:45something manageable. As D extra noted,
185. 5:48the point of abstraction isn't to be
186. 5:49vague, it is to create a level where we
187. 5:51can finally be absolutely precise. The
188. 5:54industry may try to rebrand us as
189. 5:55digital babysitters for AI
190. 5:57hallucinations, but the task of
191. 5:59understanding the problem, building the
192. 6:01abstraction, and ensuring it actually
193. 6:03works well in the wild still remains. A
194. 6:05vibe might get you 80% of the way there
195. 6:08in 10 seconds, but that last 20% [music]
196. 6:10is where real engineering lives. And
197. 6:13here's the really interesting thing.
198. 6:14More studies are coming out showing that
199. 6:16those last 20% are now harder and more
200. 6:19time-consuming because increased use of
201. 6:21AI correlates with higher rates of
202. 6:22software instability, including more
203. 6:24frequent rollbacks and patches. Code
204. 6:26still requires human validation,
205. 6:28customization, and debugging,
206. 6:30particularly for edge cases and
207. 6:32business-specific logic. But, the most
208. 6:34concerning aspect of all this is skill
209. 6:36development. Last week, I posted a video
210. 6:38about the recent updates in the web dev
211. 6:40world, and the number of people saying
212. 6:42that these updates are now irrelevant
213. 6:43was pretty surprising. Studies show that
214. 6:46over-reliance on AI weakens
215. 6:47understanding, especially in debugging,
216. 6:49and results show that developers using
217. 6:51AI performed worse in knowledge
218. 6:53assessments, suggesting that outsourcing
219. 6:55cognitive effort reduces long-term
220. 6:57competence. We all love a good graph,
221. 7:00and the latest data from the 2026 Dora
222. 7:02AI report has given us a particularly
223. 7:04fascinating one. This describes what
224. 7:06researchers are calling the U-shaped
225. 7:08productivity curve. At the start of the
226. 7:10curve, you have the vibe peak, where
227. 7:12beginners and prompt architects enjoy an
228. 7:1480% speed boost because they're building
229. 7:17simple CRUD apps and to-do lists where
230. 7:19the stakes are low and the edge cases
231. 7:20are nonexistent. But, as soon as you
232. 7:22move toward a real-world system, you
233. 7:24fall off a cliff into the complexity
234. 7:26trough, and productivity actually drops
235. 7:29below baseline human speed. Digital
236. 7:31babysitting turns out to be more
237. 7:32cognitively taxing than actual
238. 7:34engineering because you can spend hours
239. 7:36debugging a hallucination that an AI
240. 7:38one-shotted into your codebase in a few
241. 7:40seconds. So, that 20% of real
242. 7:43engineering is becoming the new
243. 7:44bottleneck where all your saved time
244. 7:46goes to die, proving that while AI can
245. 7:48generate code at the speed of light, it
246. 7:50still debugs at the speed of a
247. 7:51frustrated human trying to read a
248. 7:53stranger's mind. We'll mix things up
249. 7:55this week, and instead of the usual
250. 7:57awesome trivia, I have an awesome
251. 7:59recommendation. Don't Starve is one of
252. 8:01the games I've been playing for years
253. 8:02when I'm looking for a bit of escapism.
254. 8:05It also teaches you to craft items, cook
255. 8:07food, and survive the elements, which
256. 8:09will come in handy once we all lose our
257. 8:10jobs and the world economy goes to sh-
258. 8:12If you like this video, you should
259. 8:14consider joining our community where I'm
260. 8:16posting more dedicated weekly content.
261. 8:18Please don't forget to smash all the
262. 8:20buttons, and until next time, thank you
263. 8:22for watching.