← Claude Hacks (10x Productivity)

Claude Code Tips & Tricks Nobody Tells You, 8 Principles to Build Faster & Smarter

Video 3 of 3 · 19:33

Chapters

00:00- Why most people use Claude Code wrong
02:40- Live Demo: Building a Full-Stack Website the right way
05:47- The 8 Claude Code Best Practices (Full Framework)
06:52- Tip 1 & 2: How to plan architecture & use Claude for real-time research
09:19- Tip 3: Break problems into logical, testable components
11:00- Tip 4: Always use a test-first mindset with Claude Code
13:25- Tip 5: Version control — your safety net against hallucinations
15:05- Tips 6, 7 & 8: Sessions, context windows & staying in control
16:52- Full Best Practices Summary
17:41- What's Next

Transcript

Auto generated by YouTube. Click any timestamp to jump to that moment.

Show

0:02are doing well. This is Rajesh
0:03here. Today I'm going to cover
0:06principles of AI coding for product
0:09and I think we are at a very
0:11juncture right AI has evolved so
0:15that now there's lot of power in
0:17hands of product managers to build
0:20they want and we are at AI
0:22managers are at a very central
0:24because we have two very good
0:26right one skill is understanding
0:29requirements we understand the user
0:31thoroughly we understand
0:33market dynamics
0:34have been working on ways and means
0:37improve revenue for a company,
0:39engagement for a company. As a
0:42PM probably we go much much
0:44into improving performance
0:46and couple of important aspects
0:48users journey, right? We have been on
0:51of lot of different things and lot
0:53us have high level understanding of
0:55design, right? Okay. What is a
0:57level architecture? What does it
0:59to build a scalable analytics
1:01What does it take to build a
1:03uh website or gaming engine
1:06caters to millions of people right
1:08a lot of interesting ways and means
1:11which we have been driving the
1:13forward and often times we
1:16the space where probably there's no
1:19our backlog list is so long often I
1:22in a state when I was working with
1:23Walmart or zinga in the past we
1:26huge backlog right we have backlog
1:28100 plus items and my engineering
1:31barely scraps through 10 of them.
1:33cost of that to the company is huge.
1:36thanks to AI, the power is now in
1:39hands to s sort of launch something
1:40launch something in the market,
1:43to gather feedback very very fast
1:45see if you want to put additional
1:47strength to take it to a
1:49huge scale. Right? Having said
1:51you can also now debate that
1:53teams are more than capable to
1:56something huge. Right? As an
1:58I have been running CodePai as
2:01solo founder. Again, PM for almost 15
2:04now. Worked with companies like
2:06Flipkart and Walmart. Obviously,
2:09millions of DAUs and millions
2:11dollars in revenue. But last one year
2:13been pretty fascinating and I've
2:15pretty deep into AI, right? And
2:17used some of these AI coding tools
2:18cloud code or cursor or other tools
2:23sort of create an end toend website
2:25is codepi. And in codepi which
2:29I'm going to show you very
2:30we just cover high level aspects
2:32that you can understand what a
2:34manager can do with the right
2:37of tools and this and in this AI
2:39Right. Great. Let me go through
2:42AI website. When you log into Codep
2:44you got three types of websites you
2:47generate. A simple website is a
2:49prototype that you want to build
2:51that you want to create on the
2:52and showcase to your team.
2:54website is something that
2:56with a basic database. So you need
2:59provide your superbase access token
3:02you're working on a superbase
3:03right now. You may not even
3:05to enter the superbase token in the
3:07but effectively it creates a
3:09stack website. And last one is
3:11store which creates an end
3:12e-commerce store along with
3:14gateway integration with stripe.
3:17So this how it looks. I mean this
3:18to show you guys the art of what
3:20possible when AP really goes deeper
3:22executions and leverages right
3:24of AI coding tools and skills.
3:27I'll go I'll take you through as
3:29discussed earlier we have built
3:31of browse engines event-driven
3:33engines and even the
3:34campaign engines I think
3:36built in a span of 2 weeks single
3:38in two weeks and I was trying to
3:40a landing page for that using
3:42now let me see how that looks what
3:44did is I created a full stack website
3:47then I wrote a simple prompt I
3:49like the initial layout so I need
3:51made couple of modifications and then
3:53I landed in this particular page
3:55now when you see this page
3:58talks about hey this is my assembly
4:00where you can in fact go in it's
4:03a mini amplitude where you have an
4:05you can integrate with your main
4:06and all the events get captured
4:08later you can use them to target
4:10customers right I can request
4:12when I request access I can write
4:16name email id right and then timeline
4:20maybe I need to integrate this as
4:21and then I asked it to make the form
4:24bit dynamic have couple of addition
4:27elements based on what the
4:29needs. I need smart card targeting
4:33probably I was with mix panel before
4:36I have 100k
4:39per month. I'll do submit request
4:42now as you see when I submit the
4:45actually goes and sits in my
4:47account right so it's all full
4:50completely integrated and if I
4:52through the site you can see the
4:54you can see how to use
4:56SDK quick start guides again core
5:00API references what are the key
5:02cases that you want to run through
5:04what are the best practices use
5:06right welcome emails cart
5:08user activation and again
5:11entire website is created using
5:12PBI. Once I create the website, you
5:15do a oneclick deploy. You can attach
5:17custom domains. You can also view
5:20right? This is the browser
5:21SDK that runs on the site.
5:23someone goes and visits your
5:25we generate basic traffic metrics
5:28how many visitors, which pages did
5:30visit, how many audience came up,
5:32countries are they from, how many
5:34what is the bounce rate, what
5:35the duration. Again all this created
5:39one or two people within a span of
5:42to three weeks right beautiful isn't
5:45what AI can do in this modern world
5:47let's switch some gears let me go
5:50to the presentation and I will just
5:52you so I think we have talked
5:54about couple of what code does
5:57I have now captured is what does
6:00mean for you right I've been
6:02all this for you to sort of show
6:04the art of what is possible now what
6:06this mean to you. What this means
6:08with low teams, with low budgets, you
6:10build very high quality products in
6:12less time. One year back none of
6:14was possible. Now almost everything
6:17possible. But again, it's not so
6:18dory that you write a single
6:20and AI can create everything one
6:22You need to have some basics. You
6:24to have some good understanding.
6:26need to have a strong discipline in
6:28of how to work with AI. And that's
6:30I'm going to cover today. Right?
6:32going to cover what are the
6:34that I observed while doing AI
6:36while building codepai while
6:39couple of these peripheral
6:41that we talked about and how
6:43can benefit from my learnings over
6:46period of last one year. So I split
6:48into eight highle principles. Let's
6:50through each of them individually.
6:52are the foundations? How do I
6:53do this? Right. Um when I come
6:56an idea as product managers, we
6:58understand the user
6:59very thoroughly. We write a
7:02thorough PRD or a document so that
7:04understand what the users are trying
7:07convey and we also sort of when we
7:09the document we internal
7:10some of those workflows and
7:12and why we're doing what we're
7:14The next step given that all of
7:16are product managers you know how to
7:17PRDs is all about how do you now
7:20highle architecture right for
7:22you need some basic level of
7:23of how high level system
7:25what I typically do given I know
7:28basics of high level system designs
7:31go to GPD 502 thinking I upload my PRD
7:34terms of use cases what I want to do
7:36I ask it to create a high very high
7:38architecture in terms of different
7:41interactions how does it work
7:43high level view right high level
7:45design and I do this for two to
7:47days at a stitch I don't just
7:49a single prompt I wait for it to
7:51me something ad hoc and start
7:54immediately I wait I converse
7:56it quite a lot of things it doesn't
7:59in the first iteration so I have
8:01own questions I have my own doubts
8:03happens when we hit 1 million
8:05what happens when we hit 10
8:07events per second that's too
8:09but again just for the benefit of
8:11right what happens When an
8:14user turns up into a signed up
8:16how do we handle this scenario?
8:18scenario typical product manager,
8:21But you need to do that
8:22You need to do that
8:23so that your architecture
8:26is perfectly strong and sound. So
8:28thing that I do, I brainstorm
8:30a lot of thinking model. I prefer
8:33Thinking it depends upon you. Then
8:36do I do? My second step is AIS have
8:39have stale data. I obviously need
8:42data whenever I need an API
8:44So I tend to depend quite a
8:46on perplexity to get all the latest
8:48If I let's say want to
8:50with a anthropic API, an open
8:54API, right? Or some other API, I
8:56or recent API, I jump into
8:59to get all the recent
9:01or summarize the documents and
9:02now I have a highle plan, high
9:05architecture. Before even that I
9:08a high level PR I have a high level
9:10and I also have the latest
9:12on any of the external
9:14that my engine needs to
9:16right what next I break them into
9:20parts LLMs are super super
9:22but they have peanut sized
9:24right the brains are very small
9:26can't go beyond 200k context window
9:30claim they can go till 1 million
9:31the moment you go beyond 200k the
9:33quality deteriorates drastically
9:35how do you manage is this don't try
9:38build the entire system in a single
9:41It's not built for that. We're
9:43there yet. LLMs are not there yet.
9:45how what do you do? You break
9:48highle component into multiple
9:50and sub components. Typical
9:53typical product principles
9:55terms of problem solving right one
9:57solving. Now let's take an
9:59I'm building an analytics
10:00What does it mean for me? I need
10:02collect data from customer
10:03I need to then store the
10:06somewhere for different purposes.
10:08can be for logs just to replay just
10:12look at the past data and validate if
10:14work correctly or not. Another
10:17streaming the live events onto the
10:19website. Another for probably
10:22all these events and
10:24my campaigns. Another for
10:27analytics I want to roll up at an
10:28level, at a day level, at 7-day
10:31things like those right there. So
10:33are multiple use cases. So break
10:34into individual components. First
10:36probably is data collection,
10:38Only build that component, then
10:41the data storage component, then
10:43the analytics component, so on and
10:45forth, right? Break it into as many
10:48usable, feasible, testable
10:52as possible. And then you
10:54them one after the other and then
10:56connect and then you make a full
10:57Right? That is the third
10:59What is the fourth principle?
11:02are very good at goal-driven path
11:06Right? Now, what do I mean by
11:08If I say build me this analytics
11:10it'll definitely libert quite a
11:12It will take quite a lot of time.
11:15will take quite a lot of tokens. It's
11:17token efficient. Then what do you
11:19You probably go in with a test first
11:22right? You always already had
11:24use case document. You have your
11:26definition. You know which
11:28component you're working on.
11:30Now ask the LLM to create lot of
11:32cases based on what you want to
11:35Customer data acquisition or
11:38data capture. That's your
11:40statement, right? Write lot of
11:41cases, right? Test cases. One of the
11:44that I quickly want to show you
11:45how I do it, right? So this was the
11:48think I showed you analytics in terms
11:49how it captured in codebook AI or
11:51campaigns. There's an SDK
11:54we released recently onto the
11:55If you want you can subscribe
11:57you know you can start using. You
11:59have to drop in a mail and we'll
12:00you an API key. But again, so this
12:02an SDK that we have recently built.
12:05for this SDK when I was developing
12:07I needed to test it, right? I this
12:09a end to end mini amplitude kind of
12:12SDK which captures all the events and
12:14you can trigger campaigns based on
12:16Now what did I do? So this is a
12:17example of how I tested my SDK,
12:20I created this particular
12:22with lot of examples with lot
12:25examples. Right?
12:27you can see, I've got edge cases, I
12:30test cases, I got trackers, I got
12:34are lot of things that you know
12:36case that I created even before
12:39it. So ask LLM to create all
12:43test cases. Ask it to be
12:45at the same time don't ask
12:47to overgineer. Right? LLM LLM tends
12:50overengineer quite a lot. So you have
12:52be very very very about some of those
12:55you prompted prompted to say hey
12:58comprehensive test cases but
12:59overengineer at all. So the fourth
13:02is test first mindset okay so we
13:06brainstorming with thinking
13:07GBD 5.2 Two, research latest
13:10using complexity. Break
13:12into logical parts. As many
13:15components as possible, test
13:17mindset. Create all the test cases
13:19before you even start coding the
13:22SDK.
13:24now the fifth part, LLM tend to
13:26quite a lot, right? It it
13:29should not touch the code that I
13:31ask it to code, but it tends to do
13:33quite often, right? And there are
13:36approaches to stop it. One is you
13:38always tend to plan on cloud code.
13:41do shift. You do plan mode. You
13:43quite a lot during plan mode.
13:45look at the plan, read the plan,
13:46the plan. Once you're all good,
13:49do auto approve all the changes.
13:51Always do manual monitoring of
13:54Right? So it'll so when you
13:57that option, you can see what
13:59the changes it is making and
14:00dynamically. But even then there
14:02a lot of changes that are possible.
14:04are a lot of code messing that
14:06be possible. So what do you do for
14:08You get push, right? Git is a
14:12beautiful tool for you to
14:14It's free to use. Push into
14:17as much as possible so that you have
14:19many backups as possible. Again, just
14:22product managers, we advise to a
14:24of product managers to communicate
14:26a lot. Overcommunication is always
14:29Under commmunication is very very
14:30Similarly, try to push to get as
14:33as possible. Now, how do you do
14:34You can either ask cloud itself do
14:36or you can create a create a concept
14:40hooks right whenever it writes
14:43a code file you create a hook post
14:45usage do a good push right test it
14:48do a good push so there are a lot of
14:51in tools I'm not going to
14:52all those today that will become
14:54will make the conversation too long
14:57take away for you do get push quite
15:00so that your hard work is not
15:02and LM don't ruin your time. What
15:05the sixth point? Restart when stuck.
15:08happens quite often, right? Often
15:10get stuck in a huge context window
15:12pileup mess and it tends to shift you
15:15multiple different directions which
15:17never wanted to go. So what do you
15:19then? Simply restart. Restart always
15:22right? Laptop doesn't work,
15:24it. Code doesn't work, restart
15:26Right? But jokes apart, that
15:29quite a lot. 30% of the time, I
15:31to clear my conversation. and I
15:32to restart the session and it works
15:35well. The last two points are
15:37about how do you use the context
15:40more efficiently. I tend to
15:42lot of MD files. So whenever I
15:45any major decision within the
15:47I capture those in MD
15:50For large code bases like
15:52I think I have got roughly
15:5525 to 30 plus MD files with all
15:58decisions captured over there.
16:00whenever I start my clot again, the
16:03session need not read the entire
16:05which is way too exhaustive. It
16:08looks at this MD files and figures
16:10the critical decisions and moves on
16:12there. It's a very beautiful way of
16:14your overall context windows.
16:17again, don't let this mess up your
16:19Don't let this mess up your Git
16:23So always add a g ignore
16:25so that your github repository
16:28clean and precise.
16:31last point context is the king.
16:33try to dump too much information
16:35the context window and expect LLMs
16:38make magic. Llms only amplify the
16:41sets that you have. So you know
16:43you need to learn a lot. You need to
16:45on top of your game and use LLM for
16:48the required highquality context to
16:50it work beautifully well. make it
16:52for you and beautifully well right
16:55overall what does it mean for you
16:57are the key learning lessons you
16:59lot of time in brainstorming with
17:01models like GBD52
17:03perplexity as much as possible
17:05real-time information break a big
17:08into multiple small components
17:10on one component at a time test
17:13mindset create lot of test case
17:15so that your LLM has a goal to
17:17get pushed quite aggressively so
17:21you don't lose your hard work due
17:22LM calcinations. Restart when stuck,
17:25happens quite often, 20% of the
17:27Markdown files as much as possible
17:30manage your context window more
17:31and capture your important
17:33Context is the king. Don't
17:36garbage. Fill LLMs with only the
17:39that you need. Right? I
17:42that's all from my side. Again,
17:44one more. So based on all these
17:46at codepai we were able to
17:48four plus production grade
17:50including a full-fledged loable
17:53of a system analytics engine
17:55testing engine browser
17:57engine and quite a lot to come
17:59all this done with simple one
18:02founder with multiple with couple
18:04interns and for managing all this
18:07we have lot lot of MD files
18:09files so it's all fun if you
18:12this session I think I would
18:14encourage you to explore the
18:16starting in March first week.
18:19I'm still figuring out the cohort
18:21and details and further things
18:23those. But at a very high level,
18:25is this will be super super
18:27friendly. Any product manager,
18:29founder who wants to get onto this
18:31use AI to expedite their
18:33use AI to build something usable,
18:36production ready, this
18:38this cohort will be super
18:40for you. So we I'll be personally
18:42you to use claude code teach
18:46some of the design first principles
18:48terms of component making create a
18:50project you can pick your
18:52and I'll guide you in terms of
18:54to build what to build when you get
18:56you know how to unlock and all
18:57things and how to build guardrails
18:59AI coding right so that AI builds
19:02very very costefficient token
19:05system for you to create these
19:07right and also build guardrail
19:10that you don't expose your API keys
19:12security codes outside and you have
19:15of these patterns well baked in
19:17you hit the production. So, we're
19:19to teach all this. It's all going
19:20be about AI coding and it's all going
19:22be how you can leverage AI coding to
19:24your skills to the next level. Hope
19:25see you guys soon and if you have any
19:27drop a note to me on Rajes at
19:29redcodep codepub.a. Call it chase.

I Built PRDs 10x Faster with Claude (PM Hack)

Want the next one in your inbox?

Join 1,000+ Product Managers getting one deep dive every Friday.