How I go FAST with AI - 14 tips — Transcript

Full transcript

1. 0:00It's such an exciting time to be a
2. 0:01developer right now. There's so much
3. 0:03learning and experimentation going on as
4. 0:05everybody grapples with these new
5. 0:06superpowers that we've been given in the
6. 0:08form of agentic coding. How do we best
7. 0:10harness these incredibly smart, but also
8. 0:13sometimes incredibly dumb machines to do
9. 0:15our bidding?
10. 0:16Well, probably like all of you guys,
11. 0:18I've been trying to work this out
12. 0:19myself, and I'm still discovering new
13. 0:21techniques, but there are a few things
14. 0:23that have stuck with me now across a
15. 0:24number of different projects that I
16. 0:26think are worth sharing with you guys.
17. 0:28So, that's what we're going to look at
18. 0:29in this video, 14 tips for getting the
19. 0:32most out of agentic development. Things
20. 0:34that I personally have learned over the
21. 0:36past 6 months and use just about every
22. 0:39single day.
23. 0:40Now, although this is a Convex channel,
24. 0:42a lot of these tips I'm going to be
25. 0:43giving here are not Convex specific at
26. 0:45all, so do stick around because they
27. 0:47will still be valid for you.
28. 0:49There's some real gold in here, so stick
29. 0:50around to the end as I'm really keen to
30. 0:52hear what you guys think and also hear
31. 0:54what your tips are because I'm still
32. 0:56learning, too.
33. 0:57But, as this is sure to be quite a long
34. 0:59one, make sure you grab yourself a
35. 1:00lovely cup of tea, drop me a like and
36. 1:02sub, and let's get into it.
37. 1:07Before AI, we used to spend a lot of
38. 1:09time in the planning and preparation
39. 1:11stage of a project. And we did this
40. 1:12because the cost of implementation was
41. 1:14just so high.
42. 1:16Developers' time was and is expensive,
43. 1:19so having them sit there and code
44. 1:20something that in all likelihood will
45. 1:22get thrown away was seen as wasteful and
46. 1:25counterproductive.
47. 1:26Well, you can probably see where I'm
48. 1:28going here with this because everything
49. 1:29has changed with AI.
50. 1:31The cost of implementing something is so
51. 1:33low now that I have to continually
52. 1:35resist the urge to perfect my planning.
53. 1:38That's not to say that planning isn't
54. 1:39useful for some tasks, so but most
55. 1:41things I would rather just have my agent
56. 1:43just knock it out quickly, let me have
57. 1:46it give it a go, and then see whether it
58. 1:47works or not, and then if it doesn't
59. 1:49work, just throw it away.
60. 1:51And if I'm unsure about the change,
61. 1:53whether it's going to work or not,
62. 1:54sometimes I might just push it out live
63. 1:56and just see how it goes.
64. 1:58And if it doesn't work, I might just do
65. 2:00another push or update to fix it or
66. 2:02change it.
67. 2:03So, in my opinion, in general, if the
68. 2:04cost of reversing a change is low, then
69. 2:08just go for it. You can always just undo
70. 2:10it later.
71. 2:13A really cool thing that I've been doing
72. 2:14lately is strongly encouraging my agents
73. 2:17to add a before and after screenshot to
74. 2:20pull request changes that involve front
75. 2:22end or UI in some way.
76. 2:24This serves a few purposes. Firstly, it
77. 2:27acts as a sort of end-to-end test. By
78. 2:29forcing the AI to generate screenshots,
79. 2:31it forces the agents actually run your
80. 2:33application, and make sure that the
81. 2:35project still works, and the changes
82. 2:37look like what is expected.
83. 2:39Secondly, it lets me actually do a quick
84. 2:41smoke test to make sure that the changes
85. 2:43actually look how like I expect them to
86. 2:45look like, because sometimes I'm running
87. 2:48eight agents in parallel, and I can kind
88. 2:49of lose track of which ones I've
89. 2:51actually checked and verified or not.
90. 2:54It might seem a bit wasteful, but I
91. 2:55actually sometimes just iterate using
92. 2:58that screenshot in the PR. I just ask it
93. 3:00for changes to come back, and then it'll
94. 3:01regenerate new screenshot, and then I'll
95. 3:03have more changes, then it'll generate
96. 3:04new screenshot.
97. 3:06Oh, and finally, when working in a team,
98. 3:08a quick screenshot is really powerful to
99. 3:10let your teammates know and and see the
100. 3:12the results without having to run the
101. 3:14code themselves, which saves a bunch of
102. 3:16time and energy for for them.
103. 3:21Okay, this is a Comet specific tip, um
104. 3:23but basically, if you're going to read
105. 3:25any of the changes that your agent poops
106. 3:27out, make it the schema changes.
107. 3:30This is for two big reasons.
108. 3:32Firstly, my current attitude with AI, as
109. 3:35I mentioned before, is that I'm just
110. 3:37much more willing to accept what it
111. 3:38gives me without reading it properly if
112. 3:40the cost to reverse it is low.
113. 3:43But once data enters your tables in a
114. 3:46specific shape, it's much harder to
115. 3:49reverse that particular change, because
116. 3:52if you change the shape later,
117. 3:54um then it's going to break that
118. 3:55existing uh contract. To update that
119. 3:58schema later is generally going to
120. 4:00involve a three-part migration, which I
121. 4:02recently covered in depth in this video.
122. 4:05So, because migrations are
123. 4:06time-consuming,
124. 4:08uh depending on the number of rows that
125. 4:09you have in your table, then I consider
126. 4:11this as a high cost to reverse.
127. 4:14Because Cortex is such a nice type
128. 4:15system, you get strong types that kind
129. 4:18of flow out from your schema into your
130. 4:20functions, and then finally down to your
131. 4:22front end. So, I found that one of the
132. 4:24biggest places I can leverage my fleshy
133. 4:26meatbag brain is by thinking about the
134. 4:29schema as much of the rest of the
135. 4:31application structure tends to just kind
136. 4:33of flow and follow on from that. Oh, and
137. 4:35one quick last little tip here while I'm
138. 4:37talking about schema, I've actually
139. 4:39found it really beneficial to enforce
140. 4:41this humans must approve schema changes
141. 4:43rule by adding it to the agents.md file.
142. 4:47By doing this, it forces the agent to
143. 4:49pause before it
144. 4:52commits any
145. 4:53code that's going to require a schema
146. 4:55change.
147. 4:58One way to get the most out of AI is to
148. 5:00treat it as a collaborator, rather than
149. 5:02just a code implementer.
150. 5:05Have it interview you when you're trying
151. 5:07to explore the problem space.
152. 5:09This forces you to consider things that
153. 5:10you might not have considered before,
154. 5:12and gives the LLM additional context
155. 5:14behind the intention of your prompts,
156. 5:16not just the literal request. This
157. 5:18technique has become really popular
158. 5:19lately thanks to the excellent work of
159. 5:21Matt Pocock's Grill Me and Grill Me With
160. 5:24Doc Skills.
161. 5:25If you haven't tried them yet, I
162. 5:26recommend you give them a go. They're
163. 5:28short, but super powerful, and really
164. 5:30open my eyes to the power of this
165. 5:32interviewing technique.
166. 5:33If you don't want to install a skill, or
167. 5:35you can't due to the environment that
168. 5:36the agent is running in, then you can
169. 5:38just simply ask the AI to interview you,
170. 5:41which I found tends to work really well
171. 5:43as well.
172. 5:44And finally, this technique is really
173. 5:46great for non-coders who might just be
174. 5:48starting to explore the AI properly. We
175. 5:51probably all have a few of those in our
176. 5:52lives, but just encourage them to have
177. 5:55their AI interview them. Uh I think that
178. 5:57they'll be quite a surprise by what is
179. 6:00possible.
180. 6:04If you're using AI to build Convex apps,
181. 6:06discriminated unions are one of the best
182. 6:08ways to encourage your agent to create
183. 6:09reliable, maintainable code.
184. 6:12Let's say we have this jobs table. It
185. 6:13can have a few different status values.
186. 6:16But now we want to go ahead and add some
187. 6:18more data that can be set when the value
188. 6:20is one of those statuses.
189. 6:22By the way, this is what we used to do
190. 6:23in SQL land all the time with nullable
191. 6:26columns. With SQL, though, the issue is
192. 6:28that this is not really clear what data
193. 6:31should be set in each of the status
194. 6:32types, if the data should be cleared
195. 6:34when the status changes, etc.
196. 6:37And if you're confused when you look at
197. 6:38it, then there's a good chance that the
198. 6:39AI is going to be confused as well.
199. 6:41Now, a better way to do this would be to
200. 6:43represent your job as a discriminated
201. 6:44union of states like so.
202. 6:47Now, the entire table is one big state
203. 6:49machine, where each job can be in one of
204. 6:51each of these states.
205. 6:53And the nice thing about modeling your
206. 6:55data like a discriminated union state
207. 6:57machine is that you can do type
208. 6:59narrowing in TypeScript to only get the
209. 7:01fields that exist in that state.
210. 7:03And as extra bonus points, we can use
211. 7:05exhaustive type narrowing to make the
212. 7:08TypeScript compiler yell at us if we
213. 7:09miss out a value.
214. 7:11This is much more clear for you and your
215. 7:13agents, and as it can see exactly what
216. 7:15the valid states are and what fields
217. 7:17should or should not exist at each
218. 7:19point.
219. 7:22When working on a
220. 7:24particular task with an agent, it's
221. 7:25important not to distract it with
222. 7:27something that might not be relevant to
223. 7:29the current goal.
224. 7:30Things like refactoring code or fixing a
225. 7:32bug you spotted or tweaking the UI,
226. 7:35these are all things that can cause your
227. 7:36current agent thread to use up its
228. 7:37valuable context and lose track of what
229. 7:40it was actually doing.
230. 7:41To prevent this, instead have your agent
231. 7:43briefly record this new work in a GitHub
232. 7:46issue to do later.
233. 7:48I call these side quests, and I even
234. 7:50formalize them in the agents.md, and I
235. 7:53find it super useful when working on AI
236. 7:55with in long-running projects.
237. 7:57When you pair this with the post-merge
238. 7:59reflection tip, which I will talk about
239. 8:01in a minute, I I find that the agent's
240. 8:03often creating its own side quests to
241. 8:06improve itself as the project goes
242. 8:08along.
243. 8:12This is another technique I've been
244. 8:13using quite a lot lately, and I haven't
245. 8:15seen talked about too much. It turns out
246. 8:17that the image models that we have
247. 8:19access to in CodeX and Nano Banana and
248. 8:20others are really great not just for
249. 8:23working out what you might look like if
250. 8:25you stacked six cats on your head,
251. 8:27but they're also great for designing
252. 8:29your UI.
253. 8:30I personally like to use CodeX for this
254. 8:32because the image tool is included with
255. 8:34the generous subscription tier, which is
256. 8:35important because I could be generating
257. 8:37a lot of images here.
258. 8:39But just give this a go in one of your
259. 8:40projects. Say, "Use the ImageGen tool to
260. 8:43come up with some ideas for how this app
261. 8:45can be redesigned. Try five different
262. 8:47styles. Show me how it look on mobile,
263. 8:49tablet, and desktop sizes in portrait."
264. 8:52It works really well, and you can end up
265. 8:53with some really incredible-looking
266. 8:54designs.
267. 8:55I find it really interesting that the
268. 8:57image models are really good at design
269. 8:59compared to the LLMs.
270. 9:01Maybe it's because the actual hard bit I
271. 9:03found is that once you have a design,
272. 9:05actually then getting the agent to turn
273. 9:07that design into reality is the tricky
274. 9:10part.
275. 9:11So, this is something that I'm still
276. 9:13kind of a bit of a work in progress
277. 9:14with,
278. 9:15and it takes a little bit of a back and
279. 9:16forth to get this right. Maybe something
280. 9:19like the slash goal style loop here to
281. 9:21have it to continue take screenshots and
282. 9:23iterate until it gets close might help.
283. 9:25Maybe you guys can let me know down
284. 9:27below whether you have found really good
285. 9:29techniques for turning a screenshot into
286. 9:31an actual design.
287. 9:34As they say, two heads are better than
288. 9:36one, and one technique I found myself
289. 9:38doing a lot lately is if I'm not 100%
290. 9:42confident in my agent's plan, be it
291. 9:45Codex or Claude, then I might ask it to
292. 9:48talk to another agent to confer with it
293. 9:51and see what it thinks.
294. 9:53So, for example, if I'm using
295. 9:54GPT-5.6-Soul
296. 9:55and it comes up with a plan, but I'm not
297. 9:57sure about it, then I might just say,
298. 9:59"Please use the Claude CLI, use the
299. 10:01available model to double-check this
300. 10:02plan."
301. 10:03And vice versa, if I'm using Claude, I
302. 10:04might say, "Please use [music] the Codex
303. 10:06CLI, use GPT-5.6-Soul-X-High
304. 10:09to do to double-check this plan."
305. 10:12I tend to find that this results in some
306. 10:13new things being caught or new
307. 10:15perspectives on the problem, and gives
308. 10:17me just a little bit of extra confidence
309. 10:19that what we're going to be doing next
310. 10:21is the right thing to be doing.
311. 10:26I used to hate being forced to comply
312. 10:27with strict linting rules and
313. 10:29requirements in somebody else's code
314. 10:31base.
315. 10:32I always felt like it it just kind of
316. 10:34slowed me down and forced me into a into
317. 10:36a way of coding that just wasn't mine.
318. 10:38Well, now that AI's writing all the
319. 10:40code, I love linting. In fact, I love
320. 10:43all kinds of static code analysis these
321. 10:45days, be it type checking, linting, code
322. 10:47base coverage, etc.
323. 10:49The more of these deterministic tools
324. 10:51and processes I can apply to my code
325. 10:52base, and the more obnoxious I can be
326. 10:54with that, the more confidence that I
327. 10:56have that what the AI is writing is
328. 10:59going to be good to go and not result in
329. 11:01a big ball of unmaintainable spaghetti.
330. 11:04On a long-term personal projects I'm
331. 11:05working on, I have a TypeScript ESLint
332. 11:08Prettier follow, not to mention all the
333. 11:10unit test, integration test, end-to-end
334. 11:12test, behavior tests I also run.
335. 11:14For linting, as well, be it ESLint Biome
336. 11:16or something else, don't be afraid to
337. 11:18have your agent create new linting rules
338. 11:20to statically enforce whatever rules you
339. 11:22want to see encoded in your project
340. 11:24structure.
341. 11:25I used to find writing linting plugins a
342. 11:28bit of a headache and too much effort to
343. 11:29go to, so I never really did it, but now
344. 11:31that we've got agents that can just kick
345. 11:33it out very quickly, you might as well.
346. 11:35And follow, by the way, is a really cool
347. 11:37Rust-based static analysis tool that
348. 11:40comes preloaded with a whole heap of
349. 11:42best practice code quality rules out of
350. 11:44the box.
351. 11:45And finally, I have my agent apply a
352. 11:47number of these pre these checks as
353. 11:50pre-commit hooks using Husky.
354. 11:55This is a Convex specific tip, but make
355. 11:57sure you use the Convex plugins for
356. 11:59Codex and Clojure code if you're going
357. 12:00to be doing any sort of serious Convex
358. 12:02work.
359. 12:03We've been putting a lot of work into
360. 12:04them lately, and we think that the the
361. 12:06Codex and the Clojure code are going to
362. 12:07be increasingly become like the entry
363. 12:09point for folks when working on Convex
364. 12:11projects.
365. 12:12The plugin format is iterating rapidly
366. 12:14inside these tools, but the Clojure code
367. 12:17plugin is got some really nice
368. 12:18functionality that allows you to listen
369. 12:21for runtime or OCC errors in the
370. 12:23background, and then it automatically
371. 12:25informs agent for ways that it can fix
372. 12:27or make improvements.
373. 12:29So, make sure you get them installed.
374. 12:33Code review bots have become a bigger
375. 12:35and bigger part of my coding loop as I
376. 12:36rely on them more and more to catch
377. 12:38things that agents miss while
378. 12:39implementing.
379. 12:41I have a couple of them just follow me
380. 12:42around on various projects, and it
381. 12:44seems like they do catch different
382. 12:45things, so more bots equals betterer for
383. 12:49now.
384. 12:50By the way, I did do a couple of videos
385. 12:51on comparing code review bots on Convex
386. 12:53code a few months back if you're
387. 12:55interested in seeing how the various
388. 12:57bots stack up.
389. 13:00Storybook is for some reason a
390. 13:01controversial tool with developers. I
391. 13:04have no idea why, because I personally
392. 13:05love it, and I've used extensively on
393. 13:07many different projects
394. 13:08before and after AI.
395. 13:11It allows you to experiment with lots of
396. 13:12different ideas in isolation from your
397. 13:14main app.
398. 13:16And when you pair it with a
399. 13:16discriminated union-based schema that I
400. 13:18talked about earlier, it can be a super
401. 13:20powerful way to explore, design, and
402. 13:22debug the various states that your front
403. 13:24end can be in.
404. 13:26Maybe people used to object to it
405. 13:27because of the extra verbosity and
406. 13:30overhead of having to structure your
407. 13:31components in a way that's compatible
408. 13:33with Storybook.
409. 13:34I don't know, but in the age of AI, it's
410. 13:36super fast to create stories and
411. 13:39structure your code in a way that's
412. 13:40compatible. So, I always pull it into my
413. 13:42projects that has any be any sizable UI
414. 13:45now.
415. 13:48A big part of any project, be it
416. 13:50human-centric or AI-centric, is a
417. 13:52process of actually improving the
418. 13:54process itself.
419. 13:56That's why I like to to improve try to
420. 13:58improve the process by having a
421. 13:59post-merge reflection pass.
422. 14:02I make the agent write on the PR and in
423. 14:04the thread whether there's anything that
424. 14:06could be improved with the process
425. 14:07itself. Like, is there any missing
426. 14:09tests, any remaining work that should be
427. 14:11followed up, any docs or tooling needed?
428. 14:14Then periodically, I have an agent go
429. 14:15through all of these reflections and
430. 14:17propose a series of changes so that
431. 14:20future development and PRs go more
432. 14:21smoothly.
433. 14:22And this sort of like closing the loop
434. 14:24with something I remember doing when we
435. 14:26used to work in an an Agile environment.
436. 14:28>> [music]
437. 14:28>> You know, we would always dedicate some
438. 14:30time at the end of a sprint to reflect
439. 14:31upon the sprint process itself and see
440. 14:33if there's anything we could improve on.
441. 14:35Not sure if there's something that Agile
442. 14:36practitioners still do or not, but I
443. 14:38feel like in the age of identical
444. 14:40coding, like I feel more like a project
445. 14:42manager than a coder and and thus I
446. 14:44think it's reasonable to introduce a a
447. 14:46few of these techniques that I used to
448. 14:47use when I was working on teams
449. 14:49previously.
450. 14:52This is a bonus tip as I haven't
451. 14:54actually tried these ones out myself,
452. 14:56but a colleague mentioned that if you
453. 14:57are looking for bugs in code, it's
454. 14:59actually better to start off with the
455. 15:01position of pessimism than an optimism
456. 15:04one.
457. 15:04For example, rather than saying, "Please
458. 15:06look through this code for bugs."
459. 15:08Instead say, "There are some bugs in the
460. 15:10code. Please find them."
461. 15:12So, I'm being more pessimistic by
462. 15:14telling it there are bugs in the code
463. 15:16even though I'm not really sure. And
464. 15:17apparently, this results in finding many
465. 15:19more issues than the optimistic route.
466. 15:22Apparently as well, you can find more
467. 15:24issues if you tell that your agent that
468. 15:26another agent, not it, wrote the code.
469. 15:29For example, if Fable is reviewing a
470. 15:31code, uh tell it that GPT wrote it, and
471. 15:34often it will find more issues that way.
472. 15:37Again, I haven't tried this a lot myself
473. 15:39yet, but it's really interesting hack,
474. 15:41and I will definitely be trying it out
475. 15:42more in the future.
476. 15:46Well, that's about it for now. I think
477. 15:48I've covered some of the most important
478. 15:50key learnings I've discovered so far
479. 15:52while working with AI.
480. 15:53I will no doubt have many, many more
481. 15:55over the months and years to go on, so
482. 15:57expect more videos like this in the
483. 15:58future.
484. 15:59And if you did like this one, make sure
485. 16:01you do drop me a like and sub, uh so not
486. 16:03to miss out on the next one.
487. 16:05And by the way, if you have tips
488. 16:07yourself, then please do share them in
489. 16:08the comments down below. I'm always
490. 16:10looking on ways to improve my own
491. 16:12process, and keen to hear about your
492. 16:15techniques.
493. 16:16But, I think I'll just leave it here for
494. 16:17now.
495. 16:18Thanks for watching. Until next time.
496. 16:21Cheerio.