Plan Before You Prompt: The AI-First Product Loop

In this lesson: plan hard before prompting, run the five-move loop, and ship small behind feature flags.

Derek Peters · 60 min · June 2026

Top 3 takeaways

01

Plan hard before you prompt

Most AI builds go wrong because someone asked the model to invent the plan from a few words. How much AI speed you actually get depends on how well you plan first.

02

Run the five-move loop

Write a vision doc, a PRD, and a roadmap, prompt the agent, and then release and measure. What you learn in production feeds back into the next version, so the work keeps cycling.

03

Ship small behind feature flags

Deploy in small steps with an on-and-off flag so shipping the code and turning it on become two separate, reversible steps. Add instrumentation before you announce anything, since a dark launch with no measurement teaches you nothing.

Derek Peters

Derek Peters

Professor, Gauntlet AI

Professor at Gauntlet AI; a product and program management leader with 30+ years of experience. Currently leads AI development and product structure at Calix. Began in telecommunications and electrical engineering before moving into network engineering, Python, full-stack development, and technical product work, and has led products from ideation through go-to-market. At Gauntlet he supports instruction, grading, and product management.

Connect with Speaker

Lesson notes

A written walkthrough of the lecture, covering the patterns, the code, and the things that trip people up.

Why Planning Beats Typing

For decades, the bottleneck in software was turning decisions into code. AI has largely removed that bottleneck. Teams can now ship multiple features in the time it once took to build one, but only if their planning improves alongside their tooling.

The difference between vibe coding and building a real product is planning. Vibe coding asks the model to invent the plan. Strong teams define the plan first, then use AI to execute it. Most failed AI projects can be traced back to poor planning rather than poor implementation.

The Five-Step Loop

The workflow consists of five steps:

  1. Define the vision.
  2. Create the requirements.
  3. Build the roadmap.
  4. Execute with AI.
  5. Release, measure, and learn.

The final step loops back into the first. Production learnings inform the next vision, creating a continuous cycle of improvement.

For planning, use a frontier reasoning model such as Claude, GPT, or Gemini. For execution, use an agentic coding tool like Claude Code, Codex, or Cursor.

Cadence: A Working Example

Throughout the lesson, Derek uses a fictional product called Cadence, an AI-powered task manager that turns messy thoughts into prioritized action items.

The example is simple enough to understand quickly but complex enough to demonstrate real AI product challenges. Success is measured by a single outcome: whether users actually complete more of their planned tasks.

From Vision to PRD to Roadmap

The vision document defines why the product exists. The PRD defines what will be built. The roadmap defines when work happens.

Instead of asking AI to write these documents from scratch, have it interview you. Use prompts that force the model to ask questions, challenge assumptions, and clarify vague ideas.

The roadmap should prioritize work by risk and value, organized into Now, Next, and Later. Keep all planning artifacts in a version-controlled context file that AI can reference throughout development. Accurate context compounds; stale context creates problems.

Prompting Like a Tech Lead

Treat AI as an engineer, not a search engine. Your role is to provide direction.

Strong prompts contain four elements:

  • Intent
  • Context
  • Rules
  • Acceptance criteria

The most valuable habit is asking for a plan before asking for code. Review the plan, approve it, then move into implementation.

The operating rhythm is simple: plan, review, test, and refine. Small changes are easier to validate than large rewrites, and every line of generated code should still be reviewed by a human.

Transcript

Read transcript

Plan Before You Prompt: The AI-First Product Loop

00:00:09.530Tom Babb: We are live.

00:00:25.610Derek Peters: Right.

00:00:27.050Derek Peters: Right.

00:00:28.700Derek Peters: We'll give, we'll give folks a few moments to join. I appreciate you being here. We'll get started in just a second.

00:00:35.710Derek Peters: I'm gonna post some stuff in the chat, just to get some interaction while we get, as some people are getting in, we're gonna start shortly. I look forward to,

00:00:44.870Derek Peters: to going over this lecture with you. Welcome to night school.

00:01:04.040Derek Peters: All right, in the chat, you can also put up, put where you're calling from. It's always neat to see where everybody's from and all the different places that it's reaching out to.

00:01:12.600Derek Peters: All right, Carl's from Austin, we got some San Francisco area, Nashville.

00:01:18.090Derek Peters: Go Rocky Talk!

00:01:20.410Derek Peters: Orlando?

00:01:22.350Derek Peters: Lincoln, not… oh, gosh, Philippines.

00:01:29.970Derek Peters: Dubai. All around the world!

00:01:35.370Derek Peters: Got Maryland, got Virginia, nice. I'm from Tennessee, so I think I saw somebody in from Nashville.

00:01:43.880Derek Peters: I do have a little draw, so I'm gonna try…

00:01:47.280Derek Peters: To not let it get the best of me.

00:01:51.770Derek Peters: Alright.

00:01:57.750Derek Peters: Brooklyn Alright, while you guys are answering that question…

00:02:03.710Derek Peters: We'll go ahead and get started. So, it looks like we've got a good crowd with us. I want to start by introducing myself. My name is Derek Peters.

00:02:13.900Derek Peters: I work with the product engineering team here at Gauntlet.

00:02:18.090Derek Peters: I have 15 years in engineering, planning, and go-to-market, experience.

00:02:24.190Derek Peters: and I'm from Tennessee, and one of my biggest hobbies is I love to go fly fishing.

00:02:32.390Derek Peters: So, today, and what brings you here today is, for tonight's night school.

00:02:39.050Derek Peters: I want to talk about… kind of take a step back from coding, developing, and…

00:02:45.020Derek Peters: Everything that we're willing as engineers to just jump right into the mix.

00:02:50.580Derek Peters: And just take a step back and realize that you're not just building it, you're getting it launched.

00:02:56.660Derek Peters: And… What does that mean?

00:03:00.800Derek Peters: So that's exactly what we're going to talk about tonight.

00:03:04.730Derek Peters: So… Let's go ahead and go to the next slide.

00:03:11.070Derek Peters: So, here's the shape of the next hour.

00:03:15.480Derek Peters: We open with why building products just changed, and the methodology that we followed.

00:03:23.140Derek Peters: Then, we'll take a look at 3 important pieces.

00:03:26.820Derek Peters: All in one real product.

00:03:29.910Derek Peters: So, we're gonna touch base on planning.

00:03:32.760Derek Peters: So that's turning an idea into a vision document, a product requirements document, also referred to as a PRD, and then a roadmap

00:03:43.160Derek Peters: the AI could actually build from.

00:03:47.380Derek Peters: And then we'll talk about prompting.

00:03:51.290Derek Peters: you are only as good as your prompt, so I feel like now, like, it's a mix between prompt engineering and engineering. Prompting is very important.

00:04:01.430Derek Peters: As you steer where your product is going.

00:04:04.970Derek Peters: So you're driving the build straight out of those documents. And then we'll talk about… we'll touch base on shipping.

00:04:12.850Derek Peters: Release. How do we measure it and close the loop?

00:04:17.180Derek Peters: And then, we'll also have some time for Q&A. We're also monitoring… oh, I just saw somebody from Denver, nice.

00:04:26.570Derek Peters: We're… ADD moment. But by the end, you'll see exactly which

00:04:31.620Derek Peters: Of these three are often skipped, and why they need to go together.

00:04:38.320Derek Peters: And I'd bet money it was one that you may have bookend and just forgot about.

00:04:44.590Derek Peters: So let's dive in.

00:04:47.170Derek Peters: So we're gonna start here, because it reframes everything else. So, for decades.

00:04:55.030Derek Peters: The bottleneck in building software was typing.

00:05:00.880Derek Peters: was coding, was executing. Turning a decision into working code, took weeks.

00:05:08.470Derek Peters: Well, now, with all the great technologies that are out there, That's gone.

00:05:14.090Derek Peters: So look at these two bars.

00:05:16.280Derek Peters: It's the same timeframe. It's the same week, or sprint, or however you operate. It's the same team.

00:05:24.250Derek Peters: So, the old way.

00:05:26.180Derek Peters: That whole week brought you one shift feature, or that whole sprint brought you one shipped feature.

00:05:32.880Derek Peters: The new way, the typing collapses.

00:05:36.770Derek Peters: So, in the very same week you ship, you can ship, let's say, 3 features.

00:05:42.360Derek Peters: three times the amount of work. In some cases, four and five. It depends on how good you are at planning.

00:05:50.650Derek Peters: And hopefully this will help you become a better planner and separate vibe coding from being a product

00:06:00.690Derek Peters: Have some product sense, and kind of go into… That development phase.

00:06:08.370Derek Peters: So, it's being ruthless and planning before, and then shipping after. So that's the line between what I mentioned before, VOB coding, let's say VOB coding a slick demo.

00:06:20.300Derek Peters: And actually building a product That users are going to utilize.

00:06:32.600Derek Peters: So here's the methodology that we're going to use tonight.

00:06:36.370Derek Peters: I call it the AI-first product loop.

00:06:39.450Derek Peters: And it's just 5 moves. You frame the vision, that's your product vision document, You define the requirements.

00:06:54.150Derek Peters: And that's your PRD. You sequence the work, that's your roadmap build.

00:06:59.860Derek Peters: And then you go into building. That's where you finally prompt the AI and work on shipping those incremental releases.

00:07:08.510Derek Peters: And at that point, you're in… you release and you measure. But the most important part of this whole diagram is the arrow that loops back.

00:07:19.260Derek Peters: So… So everything that you learn in production feeds to the next vision.

00:07:27.120Derek Peters: So notice the mapping along the bottom frame.

00:07:30.330Derek Peters: Define the sequence.

00:07:33.220Derek Peters: well, yeah, define and sync what's all that lives inside of planning. The first 3 to 5 moves

00:07:40.270Derek Peters: Spent before you write a single prompt?

00:07:42.970Derek Peters: And… So, prompting is one move, shipping is one move.

00:07:48.000Derek Peters: plus the loop home, so it kind of goes in full circle. Keep that picture in your head, because every slide that we go through, everything that we discuss tonight.

00:07:57.270Derek Peters: is going to be… it's going to come up, and it's going to frame together nicely. And it's a loop, it's not a line.

00:08:05.060Derek Peters: So, Learn feeds straight back into frame, so your second version is always sharper than your first.

00:08:11.510Derek Peters: And then pretend you could get, you know, the whole thing at once. This assumes you won't.

00:08:18.080Derek Peters: And the builds are… as we get into the building piece, we'll kind of get into that cadence, so…

00:08:25.570Derek Peters: Let's talk about…

00:08:28.430Derek Peters: making it real. So, I'm gonna use, you know, a product that I'm gonna reference back to. It's a fictional product, but it is a real product that could be created. So, tonight's product is Cadence.

00:08:43.049Derek Peters: It's a task tracker with AI built in.

00:08:47.220Derek Peters: So, think of it as, like, you pasted all this chaos in your head.

00:08:51.490Derek Peters: The half-thoughts, the… everything that you want to accomplish.

00:08:55.740Derek Peters: All the to-dos that you have to do across notes and chats and everything that you have to do in, you know, your day-to-day work environment.

00:09:05.880Derek Peters: And cadence turns it into a clean, prioritized list in seconds.

00:09:09.880Derek Peters: Now, I picked it for a couple reasons. Everyone lives this pain every day. Your tasks are spread across notes, messages, maybe some post-it notes.

00:09:19.560Derek Peters: And your own memory, and let's say.

00:09:22.950Derek Peters: Yeah, they exist here, and sometimes you forget them because you get sidetracked, or you see some squirrels.

00:09:28.900Derek Peters: But…

00:09:33.280Derek Peters: It's hard to keep up with.

00:09:35.290Derek Peters: So I picked that because it's a real pain point for a lot of us.

00:09:39.300Derek Peters: And two, there's some real depth into this. You're asking a model to read a messy human input, and then hand back something structured and trustworthy, which is exactly where AI products get hard.

00:09:53.520Derek Peters: And there's an end state on the right. That's a brain dump in, prioritize task list out.

00:10:01.510Derek Peters: And…

00:10:02.790Derek Peters: as I'm planning this, or taking, like… and I'm gonna take you through the steps that I would go through in planning this.

00:10:11.230Derek Peters: But we'll trace the first feature from ideation to ship. And because every prompt and every result is right here on the slides.

00:10:21.980Derek Peters: the experience that I want you to get is understanding the workflow, and… and the reason why I didn't want to make it a live exercise is

00:10:29.950Derek Peters: It's difficult to follow watching others type, so…

00:10:33.840Derek Peters: Let's talk, you know, before we touch, like, a single tool, or talk about a single tool.

00:10:39.960Derek Peters: The most important choice of the night.

00:10:45.040Derek Peters: There's two ways to build with AI.

00:10:48.090Derek Peters: And they start with the same prompt.

00:10:53.380Derek Peters: But they end completely different places.

00:10:56.930Derek Peters: So, on the left is VOB coding.

00:11:00.260Derek Peters: You open chat, you type, build me a task app.

00:11:03.610Derek Peters: And chase whatever comes back.

00:11:06.320Derek Peters: Your prompt after prompt, You're burning tokens.

00:11:12.220Derek Peters: It's intoxicating, it's draining, and while it is fast.

00:11:17.740Derek Peters: really all you're getting is a… is a demo. There wasn't a whole lot of thought, a whole lot of planning behind it. And the problem is, you end up somewhere that you can't explain. And this is where I find engineers

00:11:31.000Derek Peters: struggle sometimes, because it is so easy to… like, with all the tools that are out there, it's so easy to just get in there to start typing, and not understand the why.

00:11:42.440Derek Peters: So, you gotta be the master of the ship as you're coding.

00:11:46.580Derek Peters: As you're prompting, as you're planning. Planning, to me, is the most important aspect of engineering.

00:11:53.940Derek Peters: And you still have to plan. You can use all the resources and tools out there to kind of help bounce ideas off

00:12:00.760Derek Peters: And we'll get into that tonight.

00:12:03.490Derek Peters: On the right, though, we… you know, is what we actually teach here at Gauntlet.

00:12:10.870Derek Peters: Plan, and then prompt.

00:12:13.420Derek Peters: you decide… The, you know, you decide what's… what's the best first, you know, what's best going forward.

00:12:22.470Derek Peters: And then… You flesh out the vision, you flesh out the requirements, you create a roadmap.

00:12:30.340Derek Peters: You create your architecture diagram, and then…

00:12:34.420Derek Peters: Once you have all those rich.

00:12:37.530Derek Peters: documents going in, then AI can help you

00:12:42.270Derek Peters: Build from those documents, and now you can steer.

00:12:47.070Derek Peters: And be a product manager, and let…

00:12:51.720Derek Peters: The agents be your development team.

00:12:55.720Derek Peters: Now, here is probably one of the most common things that people miss.

00:13:01.270Derek Peters: It's… it's not slower.

00:13:04.350Derek Peters: Same model, same speed, But between Vive Coding and planning, Let's see, is prompting…

00:13:19.730Derek Peters: Yeah.

00:13:20.850Derek Peters: You can even… you can even bounce some ideas off it and have it create better prompts.

00:13:28.680Derek Peters: Absolutely.

00:13:30.450Derek Peters: It's the workflow wrapped around it, so… The biggest miss…

00:13:37.800Derek Peters: that I usually, or typically see, it, it, well…

00:13:45.730Derek Peters: Let's… let's get into talking about it.

00:13:51.590Derek Peters: And I also want to talk about what we actually use per stage.

00:13:55.620Derek Peters: And I want to talk about tooling, for example, because it is important.

00:14:00.120Derek Peters: And it's usually the first thing that people want to, like, dive into and ask.

00:14:07.780Derek Peters: And it's the workflow that matters.

00:14:10.410Derek Peters: Far more than, let's say, the brand of tooling that we use.

00:14:14.960Derek Peters: So, for planning, you won't… A good frontier reasoning model.

00:14:21.340Derek Peters: So, like, Claude, Opus, GPT, you can even use Sonnet for just bouncing ideas.

00:14:28.460Derek Peters: Gemini, you know, whatever you got handy, those are all great tools to kind of help start for planning.

00:14:36.210Derek Peters: And I run it in Cloud Desktop, or sometimes I'll bounce it off just to… I'm always in the mindset, trust but verify I'm not bounce it off a GPT with a product so that my vision and my requirements ride along in every message.

00:14:53.600Derek Peters: And then for prompting, this is where you move into an agenic coding tool.

00:14:59.990Derek Peters: And there's really two camps right now. So you have Anthropic's Cloud Code, you have Codex, and you do have Cursor.

00:15:11.050Derek Peters: But… it's a whole new system. It's a terminal command line interface, so it's a CLI,

00:15:18.740Derek Peters: And an extension that lives inside Cursor and VS Code.

00:15:24.810Derek Peters: Cursor itself is the editor, a lot of people build in it.

00:15:28.980Derek Peters: But, you know, use a strong model. So Claude Sonnet or Opus is good, or GPT 5.5,

00:15:36.380Derek Peters: for real building. So, like, whenever you're actually… you get into your building, those three LLMs are really strong for, helping you code and turn those ideas into something that real users can use.

00:15:51.790Derek Peters: So, let's go ahead and kick this off with planning. So, this is… this is the first and most important step, in my mind.

00:16:04.970Derek Peters: So… I'm going to… Ask you to do something that feels… feels backwards, And let's talk about speed.

00:16:15.010Derek Peters: Because in the world of AI-first development.

00:16:18.520Derek Peters: It's go faster, go faster, ship, submit, go faster, go faster.

00:16:24.790Derek Peters: You're still going fast, but take a step back and slow down.

00:16:30.150Derek Peters: The single biggest predictor of whatever an AI build

00:16:35.340Derek Peters: Succeeds is how much thinking happened before the chat window ever opened.

00:16:41.410Derek Peters: When a build goes sideways, 9 times out of 10, it's because someone asked the model to invent a plan, or to do the planning for them based off of a couple words.

00:16:55.250Derek Peters: I'm looking at the chat.

00:16:59.360Derek Peters: So, Mark asked, how do you choose between the three agents? One front-end or back-end? Well, so…

00:17:06.869Derek Peters: Usually, I'll start with, like, Claude Code.

00:17:09.530Derek Peters: And, like, when I'm in my development phase.

00:17:13.030Derek Peters: Of what I want to accomplish, and it does a pretty good job with both back-end and front-end.

00:17:20.130Derek Peters: But we'll definitely get into that in the shipping piece.

00:17:24.609Derek Peters: So, so, for the next few minutes.

00:17:29.100Derek Peters: We'll take a look at,

00:17:31.570Derek Peters: You know, 3 short documents that…

00:17:34.840Derek Peters: will help you be better planners along the way. That's the vision document, requirements document, and a roadmap.

00:17:42.690Derek Peters: And…

00:17:44.170Derek Peters: Only then do you want to prompt. No code yet. This is just thinking out loud on paper and know what the objection is and what you're trying to accomplish.

00:17:55.750Derek Peters: So, you start with a vision document.

00:17:59.280Derek Peters: So, before features, before Roadmap, You write out your vision document.

00:18:08.240Derek Peters: One page on why this exists, and who's it for?

00:18:12.940Derek Peters: What problem are we trying to solve?

00:18:17.090Derek Peters: Who's gonna benefit from it?

00:18:19.730Derek Peters: And here's the cadence on the right.

00:18:22.620Derek Peters: So, the vision.

00:18:24.050Derek Peters: Turn the chaos in your head into a clear, prioritized plan.

00:18:28.980Derek Peters: So the user, the core problem. Tasks live everywhere, and…

00:18:35.130Derek Peters: Prioritizing can sometimes be a constant manual work.

00:18:40.300Derek Peters: And the one metric that matters?

00:18:42.670Derek Peters: Plan task… planned tasks are actually completed. That's all that matters, is what you plan gets completed.

00:18:50.860Derek Peters: And then… and principles, so which, you know, let's say, which, for a tool that you live in all day.

00:18:59.560Derek Peters: means calm, trustworthy priority, and you always stay in control. Now, how do you write this with AI?

00:19:07.800Derek Peters: You do not ask to invent your vision.

00:19:11.090Derek Peters: One thing that I like to do, and I have an example here, is…

00:19:18.140Derek Peters: I ask it to interview me.

00:19:20.510Derek Peters: if I'm getting blocked in my mind, like… and look…

00:19:25.390Derek Peters: Kind of look how the prompt is built, like, act as a product coach.

00:19:29.000Derek Peters: One question at a time.

00:19:31.040Derek Peters: Don't invent the answers.

00:19:33.650Derek Peters: And push back… You know, have it push back when you're vague, like, when you're vague.

00:19:40.540Derek Peters: That's a planner's prompt, not a wish.

00:19:43.380Derek Peters: And the card beside it is…

00:19:46.590Derek Peters: what it came back with. A PVD.

00:19:49.910Derek Peters: So… let's say, through that conversation. So, act as a product coach. Interview me one question at a time.

00:19:57.050Derek Peters: about cadence.

00:19:59.250Derek Peters: the user, the core problem, the metric that matters, and the three principles. Then draft a one-page PVD. Don't invent answers if I'm vague.

00:20:09.090Derek Peters: Push back until it's sharp.

00:20:12.300Derek Peters: And then have it interview you.

00:20:14.470Derek Peters: And go through, and then that way you can kind of get all the stuff that's in your head kind of out there.

00:20:19.690Derek Peters: And you also give it direction, and now it knows exactly what role it has to play, and in return, it's going to help you flesh out that vision.

00:20:28.810Derek Peters: I do that all the time, like, if I'm struggling to,

00:20:34.080Derek Peters: to kind of get what… what I'm thinking about. I… I… instead of just…

00:20:39.960Derek Peters: typing, and then have it respond, and then keep going back, I have it interview me. I give it a… I give it a title, and have it interview me.

00:20:51.860Derek Peters: It really depends, so this is for Mark.

00:20:55.400Derek Peters: I do like the kind of the interview flow, where, like.

00:21:00.280Derek Peters: we take it one bit at a time. The LLMs have gotten to a point now where you could probably do everything all at once and answer all those questions, but for the way that I process information, it's good for me to get it back, you know, kind of turn by turn.

00:21:16.240Derek Peters: And… answer, and then have it clarify, answer again.

00:21:21.940Derek Peters: But you could have it go to ask question user flow. My preference is turn by turn, just because I know how I work and operate, and that works best for me. So whatever works best for you, but that's a great question.

00:21:37.780Derek Peters: So let's talk about, turn a vision into a PRD.

00:21:43.640Derek Peters: So this is where you're… this is where you're defining and kind of fine-tuning. So the vision is a way… well, I'm sorry, the vision is the why, the PRD is the what. So anytime that you're developing, anytime that I'm thinking about what I'm going to do, I always want to…

00:22:01.620Derek Peters: I always want to ask myself, why am I doing this?

00:22:06.460Derek Peters: And then get into… what? What are we doing?

00:22:11.750Derek Peters: That satisfies the why.

00:22:13.950Derek Peters: You can even go back to, like, the user story logic. As a user, I want to do this, so that I can do that.

00:22:22.410Derek Peters: Because that way you have the who, The what and the why.

00:22:28.300Derek Peters: But… And, and you have one ver… you know, let's see…

00:22:36.100Derek Peters: You have the vision, which is the why, the PRD, which is the Y, and only for version 1, you have the PRD. So, the prompt on this slide, and I want you to… I want to call this out. Notice how much instruction I'm packing into my prompt.

00:22:50.930Derek Peters: Because that's the whole difference between a vague ask, that's the difference between vibe coding, and a useful one. That's really going to drill down, it's actually going to save you some time, too, from tokens, the clearer that you can be with the LLM and interact with it. The lot…

00:23:06.510Derek Peters: you're going to get a more richer response. So I'll point the model at the pvd.md.

00:23:12.290Derek Peters: As the single source of truth.

00:23:16.010Derek Peters: And… I name the exact sections that I want.

00:23:22.700Derek Peters: and then… well, and I… what I want to do is I name the… not only do I name the exact sections that I want back.

00:23:30.660Derek Peters: I have 3 to… you know, give me 3… 3 to 5 user stories.

00:23:35.960Derek Peters: Give me in scope versus out of scope, and measurable success.

00:23:40.410Derek Peters: Because that's what you need to do as a product planner.

00:23:46.680Derek Peters: I'll tell it to cut anything that's non-essential.

00:23:50.120Derek Peters: And park it under out of scope.

00:23:52.430Derek Peters: And I saw the comment in the… yes, that's… I always do that because…

00:23:58.030Derek Peters: Sometimes we tend to take on more than we can chew, and knowing clearly on the front end what's out of scope and what's in a scope can help as you develop. And then you can set some stretch goals.

00:24:12.780Derek Peters: And then I sequence with a roadmap.

00:24:15.450Derek Peters: So, this is…

00:24:17.060Derek Peters: This is what I refer to, you know, in the planning phase, this is what I like to call sequence. And you're turning that PRD into a roadmap.

00:24:25.130Derek Peters: It's… And people often get confused, like, this is not a wish list, this is, like, hard must-haves.

00:24:34.700Derek Peters: You might have nice-to-haves on the back end somewhere, kind of listed as your stretch goals.

00:24:40.880Derek Peters: But this… this is not a wish list sorted by what sounds fun.

00:24:44.810Derek Peters: And this is where a lot of people, when they're planning, they miss out. Like, sequence by risk and value.

00:24:52.220Derek Peters: And notice the prompt does the thinking for the model. So, sequence by risk times value.

00:24:58.830Derek Peters: Not by excitement, Because you're only in the infancy of this planning.

00:25:05.540Derek Peters: And put only the riskiest assumptions work in now?

00:25:09.940Derek Peters: and demand… you could demand, like, a one line of rationale per item if you need help making a decision on where to move forward. But here's the roadmapap that it gave back. And just so you all know, like, I went through this exercise, and…

00:25:27.510Derek Peters: condensed the output, but what you're seeing is the output. Now, next, later.

00:25:33.990Derek Peters: So…

00:25:35.650Derek Peters: And that's a really great tool to follow, because that way it keeps you on track, and it keeps you… keeps a nice sequence as you go into developing.

00:25:46.080Derek Peters: Yes, when I do this, they live in a GET repo.

00:25:50.060Derek Peters: And, they're usually README files, MD files, that I put into a docs folder.

00:25:56.620Derek Peters: But you could put them in,

00:25:59.270Derek Peters: Notion or whatever other tool that you would like to utilize, but I mostly keep them in my repo. That way, I can always have everything in one place.

00:26:15.380Derek Peters: The next question is, how do you communicate with the dev teams? The way that I communicate to the dev teams is, during stand-ups, or let's say there's

00:26:25.100Derek Peters: like, a once-a-week product, like, a planning session to, like, for plan… like, you have, like, a product planning session where you decide, like, what tasks are going to get worked out.

00:26:35.940Derek Peters: that is typically, like, an hour long, and then we follow up with, like, daily stand-ups. Just like a scaled, Agile or,

00:26:45.610Derek Peters: kind of a scrum framework, like, none of that really changes. What changes is how much you plan for your sprint in this new world of…

00:26:54.100Derek Peters: technology.

00:26:57.820Derek Peters: Yeah, they would appear into, into the, like, whatever repo is for, like, the product or the feature that you're working on. Typically we keep them in the… you would want to keep it in the repo themselves, like, of the product repo.

00:27:11.930Derek Peters: And then they can access those. Now, those were some great questions. Thank you so much.

00:27:18.910Derek Peters: Okay.

00:27:21.270Derek Peters: And the last planning move, and it… this is where it kind of compounds. You've got 3 documents now.

00:27:30.850Derek Peters: Don't paste them into the chat and lose them.

00:27:33.910Derek Peters: put them into a durable context file that AI reads every single time, And here's… Here's cadence.

00:27:45.150Derek Peters: This is the product that we're working on. So it points to the vision document, and points at version 1 of the requirements document. Because remember, this is… the reason why this is V1 is this is the first iteration of this product.

00:27:59.970Derek Peters: This is an… it's an ideation, that's where we started. Now we got something kind of fleshed out.

00:28:08.770Derek Peters: Yes, we're definitely going to get into that, Christian.

00:28:11.600Derek Peters: Not so much the architecture, because this is more product planning for tonight, but but yeah, we'll talk about prompting.

00:28:20.130Derek Peters: That that'd be next.

00:28:23.890Derek Peters: And…

00:28:25.910Derek Peters: So the PVD points to version 1 of the PRD, because that PRD is a living document, it's going to change as more requirements come about, and as you get… as you continue to progress your product.

00:28:40.440Derek Peters: and one rule that really matters for the… you'll also, like, let's say in that, PRD, as one thing, one thing that might live in there is, like.

00:28:52.260Derek Peters: We'll have the definition of done, and then,

00:28:56.380Derek Peters: maybe, like, whatever AI or whatever features that you're planning to, to integrate.

00:29:02.410Derek Peters: But the beauty of this is, I never have to re-explain what… Cadence is, again.

00:29:09.890Derek Peters: And every prompt inherits the vision and the scope automatically.

00:29:15.090Derek Peters: it… it gets treated like code, especially, and this is the beauty of having it in your repo, into a, like, a separate Docs version, or into a… I'm sorry, into a separate docs folder.

00:29:27.220Derek Peters: And you can version it, and then you can tell it, like, which version to go. Like, you could have multiple

00:29:32.700Derek Peters: Iterations of it as you continue.

00:29:35.330Derek Peters: But you can version it, you can review it, you can keep it current.

00:29:39.350Derek Peters: A stale context file can be dangerous.

00:29:43.070Derek Peters: So, make sure that it's always pointing to the right documentation.

00:29:49.020Derek Peters: Because the model will trust it completely.

00:29:52.980Derek Peters: And… You being kind of the captain of that ship, and kind of making sure that it's on track?

00:30:00.070Derek Peters: Definitely, I would recommend using version control.

00:30:03.390Derek Peters: But this is all V1, because this is a new idea.

00:30:12.150Derek Peters: So let's talk about prompting.

00:30:18.440Derek Peters: Yeah, you can convert anything to an ELD.

00:30:22.730Derek Peters: there was a question in here that I saw. Do you manage memory separate from PRD or PVD? I typically just use version control, so it knows which one to point to.

00:30:35.230Derek Peters: And then I'll archive. I don't know if you can elaborate on that question, Mark.

00:30:40.370Derek Peters: I'll be more than happy to answer it.

00:30:42.300Derek Peters: But let's get into prompting. So prompting, this is the part that…

00:30:48.930Derek Peters: I hope a lot of you get out of this.

00:30:52.720Derek Peters: But we wrote 3 documents before getting here.

00:30:56.070Derek Peters: Three of the five moves in the loop that we talked about, you know, going back to the methodology.

00:31:02.660Derek Peters: And now those documents pay off enormously.

00:31:09.370Derek Peters: So the mental model that I want you to adopt

00:31:14.060Derek Peters: You are not a user of the model.

00:31:18.540Derek Peters: You are the tech lead. You are a product manager of that model. So clear intent goes in, reviewable work comes out.

00:31:28.160Derek Peters: And you put your name on all of it. Everything in…

00:31:31.870Derek Peters: And this next section is about getting the signal in and keeping the loop tight enough that you never lose control of it.

00:31:42.270Derek Peters: Oh, yeah, yeah, you can do that. Sometimes you can also cache some of that, but yes. No, that's good.

00:31:53.060Derek Peters: Thank you, Mark.

00:31:54.500Derek Peters: Alright, so let's get into it here. So, a re… a weak prompt…

00:31:59.050Derek Peters: And we've all started here when AI first came out, and we're like, oh my gosh, what can we do with this? And we just, like, asked it all kinds of crazy questions, but it's… has evolved to a powerful tool that we all use today. But a weak prompt says, build

00:32:15.810Derek Peters: let's say, build the AI thing that makes tasks.

00:32:19.380Derek Peters: Or build me a task management system.

00:32:24.650Derek Peters: A strong prompt says three things in one… Breath.

00:32:30.730Derek Peters: what do you want? The constraints it has, you know, the constraints it has to respect.

00:32:37.410Derek Peters: And you'll know both, or you'll…

00:32:41.390Derek Peters: you'll both know it's done. Or…

00:32:44.920Derek Peters: How you'll… like, how do you know it's done?

00:32:47.220Derek Peters: It will tell you that. And the highest leverage habit

00:32:51.670Derek Peters: There is, ask for a plan before any code.

00:32:55.200Derek Peters: The anatomy is… On the slide here.

00:32:59.610Derek Peters: Intent, context, rules, and acceptance.

00:33:03.890Derek Peters: So, what and why in one sentence.

00:33:06.770Derek Peters: Files, stack, patterns to follow.

00:33:09.960Derek Peters: straight from the PRD, or a context.md, or a context.md.

00:33:15.440Derek Peters: and accept the PRD's own success metrics.

00:33:20.930Derek Peters: You make the model show you its plan first, That way, you understand

00:33:27.550Derek Peters: Clearly, what's going to happen next, and then you can help kind of direct it and guide it where to go.

00:33:36.340Derek Peters: I haven't used plan mode yet. I may give that a try, but I usually do, I usually put it into, like, a project folder, and then plan, and then transfer it into GitHub.

00:33:48.480Derek Peters: Using, Quad Desktop. That's my preferred method. I'm not saying that's the method to use. Everybody…

00:33:56.310Derek Peters: This kind of flows regardless what you want to have.

00:34:01.440Derek Peters: And I'm learning from you all just as much as I hope you're learning from me.

00:34:07.280Derek Peters: That's good.

00:34:09.449Derek Peters: forward slash goal.

00:34:10.960Derek Peters: In Codex, alright.

00:34:13.460Derek Peters: So, once the plan's approved, you have a model… well, you… you hand the model a map.

00:34:19.920Derek Peters: And the map literally… well, and the map is literally the prompt on the slide, so walk through it. First, I pointed at the docks.

00:34:30.730Derek Peters: Read the PRD and the context file.

00:34:34.150Derek Peters: because… Because the docs are the spec.

00:34:40.010Derek Peters: You're basically feeding it everything it needs.

00:34:43.250Derek Peters: And the exact files to touch.

00:34:46.400Derek Peters: And an existing pattern to follow.

00:34:50.120Derek Peters: That one line is worth a whole paragraph of style notes.

00:34:55.340Derek Peters: So the… the rules come straight from the PRD.

00:34:59.220Derek Peters: return, return type JSON, Set the priority and the urgency in the text. Never invent a due date.

00:35:08.810Derek Peters: And it's also a great failsafe for bad output.

00:35:13.620Derek Peters: then…

00:35:16.050Derek Peters: let's say, then done when is the PRD's own success metric. So, a prioritized edible list in under 5 seconds.

00:35:26.170Derek Peters: Plus…

00:35:27.590Derek Peters: the tests that have to pass, because that's another thing, when you don't properly plan, a lot of people don't realize

00:35:34.600Derek Peters: Oh, you mean I have to do testing? That's just as important as planning, like, once you get to the end, then you're ready to test and roll it out to real users?

00:35:44.900Derek Peters: So… Now you have everything that you need to go forward.

00:35:52.010Derek Peters: So I'm just feeding the model the documents we already wrote, the PRD is the prompt.

00:35:57.230Derek Peters: On the next slide, we'll watch what happens, let's say, when you skip all of this. And let's… let's talk about, like, a good prompt versus a bad prompt.

00:36:05.510Derek Peters: So this slide… Is that the problem?

00:36:09.140Derek Peters: you'll see there's a good prompt versus a bad one, and again, we've all been there, like, especially when AI started rolling out, and then it started getting a lot more richer, and then now it's a tool that we use every day.

00:36:24.000Derek Peters: So, on the left is the bad prompt. Add an AI feature that turns notes into tasks. Make it smart.

00:36:30.460Derek Peters: And here's what you actually get. You know, what you get is it splits text into breaks. It will deliver you something that does somewhat work, you get excited, but there's no real proof behind it.

00:36:44.200Derek Peters: And then on the right is an example of a good prompt.

00:36:49.680Derek Peters: Implement, extract, and prioritize PRDV1,

00:36:54.200Derek Peters: You can point it to it. You can, you don't even have… you don't have to give it its… the whole file path, unless you have multiple versions. Especially in the first prompt, you want to be clear on, like, where it pulls from.

00:37:08.210Derek Peters: And then… return type JSON title priority due, except brain dump.

00:37:15.280Derek Peters: edible list in less than 5 seconds, plan first, no code until I approve. That's important, because if you're running auto mode, or something that just kind of goes on its own.

00:37:26.890Derek Peters: You lose a little sense of control, even though it can do something really awesome and really cool, and you're like, oh, wow, this is… this is creating something.

00:37:34.540Derek Peters: Always ask it to approve.

00:37:39.200Derek Peters: Don't… don't just have it go and create a bunch of stuff.

00:37:45.240Derek Peters: So here's the rhythm that keeps it on the rails.

00:37:49.330Derek Peters: Ask, review, test, refund, and repeat.

00:37:54.920Derek Peters: Small differences over big rewrites.

00:37:58.230Derek Peters: every time.

00:37:59.680Derek Peters: Because a small diff is something you can actually verify. Read every line. I know it's tempting to skim, or skip, or just be like, yeah, get what I needed and go forward.

00:38:12.560Derek Peters: and… It looks confident, but you own this.

00:38:16.610Derek Peters: Not the model. So, like, if… now, if you're doing something on your own, just to help yourself out, that's fine, but if you're… if you're shipping something that users are going to use.

00:38:27.330Derek Peters: Read it.

00:38:28.950Derek Peters: Even if it looks confident.

00:38:31.690Derek Peters: If, you know, the product that we're working on.

00:38:36.950Derek Peters: If Cadence quietly buries, call someone back.

00:38:41.070Derek Peters: That's an important task at the bottom as low priority.

00:38:44.400Derek Peters: the AI is not the one who has to answer for it. If you miss a task, and you're creating, let's say, this task management system.

00:38:53.800Derek Peters: And it's… Not prioritizing right, or taking these tasks and doing what it's supposed to with them.

00:39:03.030Derek Peters: then that's… The person that develops its issue.

00:39:09.260Derek Peters: And…

00:39:11.610Derek Peters: let's say, the move that people that I see come through, that people resist most, is when the model gets stuck and starts circling. Don't argue with it.

00:39:22.090Derek Peters: Just reset the context, open with a clean chat, start fresh.

00:39:28.390Derek Peters: And I know that, I mean, as good as these models are that are out here today, they can still get confused, they can still hallucinate. Arguing with a confused model just trains it to be confidently wrong.

00:39:43.950Derek Peters: at you. So, and… If you're using Clyde, sometimes it's very confident.

00:39:50.220Derek Peters: And you just gotta restart.

00:39:52.060Derek Peters: And just kind of reset the context.

00:39:59.420Derek Peters: Yeah, I do manually reset the context window, or I'll, I'll have, like…

00:40:08.500Derek Peters: a golden prompt that I can use to kind of start a new chat and put it in there. So, like, I'll save some prompts in the back end.

00:40:16.890Derek Peters: That's true, codecs compacting, that is such a great feature. Absolutely. And just have it compact the con… and do the context resetting itself, that is pretty solid.

00:40:29.500Derek Peters: What's happening?

00:40:30.890Derek Peters: Good feedback.

00:40:33.120Derek Peters: So, before we ship, Let me name 4 ways this all falls apart.

00:40:42.140Derek Peters: Because over a long session, You know, you went through all of this… You'll fill the form.

00:40:52.940Derek Peters: So… One, a 600 line diff.

00:40:58.640Derek Peters: You can't review.

00:41:00.850Derek Peters: Well, one, because it's 600 lines, so you don't, you just trust it.

00:41:06.180Derek Peters: Two, confident code that reads beautifully, and

00:41:10.640Derek Peters: It's quietly wrong. So you still need to go through it and just…

00:41:15.380Derek Peters: Check. Testing, too. Make sure you embed testing into your planning.

00:41:21.010Derek Peters: Three, arguing with the stuck model with the context rots into mush. You're just kind of going downhill, you're not going to win that battle.

00:41:29.790Derek Peters: Looks like you all have ran into this at some point with the context window and having to reset it, which is fantastic.

00:41:38.460Derek Peters: But keep in mind, the model didn't fail.

00:41:42.310Derek Peters: The workflow did.

00:41:45.290Derek Peters: And the fix is always the same three things.

00:41:48.660Derek Peters: smaller ask, Read the changes.

00:41:54.060Derek Peters: And implement or incorporate some sort of acceptance test as you're going through this.

00:42:02.080Derek Peters: when you're… when your AI sessions fall apart, And it will come back…

00:42:10.730Derek Peters: And just remember that, like, just kind of reset the context, remind it what it's doing, remind…

00:42:17.440Derek Peters: remind… yeah, reminded what it's doing, and just kind of move forward. Don't get aggravated, as aggravating as it can be, but

00:42:26.060Derek Peters: And keep your prompts clean and informative.

00:42:29.460Derek Peters: Yes, don't let it compact.

00:42:31.940Derek Peters: So let's talk about shipping.

00:42:34.200Derek Peters: So this is where AI-built products quietly… Can, if they're…

00:42:40.400Derek Peters: Vibe-coded, or no thoughts behind it, this is where they go to die.

00:42:45.290Derek Peters: It might be a beautiful demo on someone's laptop, never reaches a single user.

00:42:51.450Derek Peters: Because shipping started to feel scary, started to feel a little nervous.

00:43:01.730Derek Peters: So we're going to…

00:43:04.430Derek Peters: my goal is not to make shipping sound boring, but my goal is, you know, let's start small, let's flag it, let's measure it. The entire goal of shipping, and if we go back to, like, let's say Cadence, as an example, or task management software.

00:43:21.360Derek Peters: you know, the first slice in front of real people without anyone in the building holding their breath. Like, you want to make it to where, like, hey, I want to demo this to a user, or I'm going to go in.

00:43:33.840Derek Peters: Shipping is… this proof of concept, getting feedback, getting real user feedback. So let's talk about the ship

00:43:41.640Derek Peters: Ship behind a flag.

00:43:43.830Derek Peters: So… The trick to shipping fast?

00:43:48.600Derek Peters: Is not shipping fast, fast, fast, fast.

00:43:53.060Derek Peters: And what I mean by that is, just… It's okay to submit, but… Going too fast?

00:44:03.240Derek Peters: can sometimes cause problems. But the secret to shipping fast is shipping small, and shipping often.

00:44:11.290Derek Peters: So… with feature flags, And with feature flags.

00:44:18.680Derek Peters: That's what makes it… makes it safe.

00:44:23.270Derek Peters: You have those small little iterations that you can ship out.

00:44:27.360Derek Peters: But you're shipping in small chunks. You're shipping often, and you're testing those.

00:44:34.560Derek Peters: So let me say what shit behind the flag actually means. Well, what it means to me.

00:44:39.950Derek Peters: Because, to me, it's the whole foundation of the slide that you're seeing right now.

00:44:45.650Derek Peters: A feature flag is just an on-off switch wrapped around your new code.

00:44:50.730Derek Peters: If the flag is on, Run a new extractor. If it's off, don't.

00:44:55.690Derek Peters: You deploy the code to production with that switch turned off.

00:45:01.470Derek Peters: It lives… well, it lives on the server.

00:45:04.640Derek Peters: But it's invisible to users until you decide to flip it.

00:45:07.860Derek Peters: So… Shipping and turning on becomes two separate reversible steps.

00:45:14.850Derek Peters: And that's what takes the fear

00:45:17.750Derek Peters: Yeah, it takes the fear out of releasing.

00:45:20.610Derek Peters: So, you merged unfinished work, Because it's switched off in production.

00:45:28.130Derek Peters: That keeps Maine reasonable at all times.

00:45:31.310Derek Peters: Which means you are never trapped in, let's say, a 6-week branch that fights you on every rebase.

00:45:38.490Derek Peters: And if you've ran into merge conflicts, that can be aggravating.

00:45:43.990Derek Peters: Then, you roll out gradually. Internal users first, then a small percentage.

00:45:50.140Derek Peters: And if you're working, let's say, on a solo product, grab some folks that can give you some feedback on that, and just do some user testing as well.

00:45:59.680Derek Peters: But internal users first, Then a small percentage, watching the…

00:46:05.140Derek Peters: Keep an eye on it the whole way.

00:46:07.880Derek Peters: Then, you roll it out to everyone.

00:46:11.000Derek Peters: If, let's say, the extractor returns something weird.

00:46:14.300Derek Peters: You flip a, you know, let's… yeah, return something weird, you flip a switch, You don't need…

00:46:21.360Derek Peters: you don't need to do, like, an emergency 2 a.m. deployment.

00:46:25.770Derek Peters: the prompts… And the prompts on the sl… got tongue twisted there. And the prompts on the slide?

00:46:34.340Derek Peters: And it… you'll see it right there, like, wrap the extractor in a flag cadence, extract V1 to fall off with fallback to manual entry, roll out to internal 5%, 100% gated error rate less than 1%.

00:46:50.200Derek Peters: So… When you're shipping features, It's always good to ship small.

00:46:58.110Derek Peters: Shift often, and… Have a way to kind of easily cut them on and off for slow rollouts.

00:47:08.040Derek Peters: So here's a habit.

00:47:12.170Derek Peters: That most people miss completely.

00:47:15.220Derek Peters: They use AI to write code.

00:47:18.200Derek Peters: And then go back to reviewing all of it by hand.

00:47:21.280Derek Peters: That's complete… that's completely backwards, because I don't know if you've been in that situation, that's a lot of work.

00:47:27.970Derek Peters: Generate test alongside the feature.

00:47:31.400Derek Peters: not as an afterthought. If you have a feature, That you're planning to implement?

00:47:39.070Derek Peters: Work on that feature.

00:47:42.310Derek Peters: Test it.

00:47:45.350Derek Peters: If it's for users that, you know, if it's for real users that are going to use it, get some user feedback.

00:47:51.550Derek Peters: And then roll it out.

00:47:53.550Derek Peters: And test in increments.

00:47:57.980Derek Peters: So, the pop line on the slide runs the test. So, type checks, and it has an AI reviewer to read off the

00:48:06.180Derek Peters: Read off for what humans tend to skim past.

00:48:09.740Derek Peters: Leap Secrets, missing test.

00:48:12.970Derek Peters: broken conventions, and that's another thing that I want to call out, too. Like, the number…

00:48:19.470Derek Peters: I'm not saying, like, review it line by line, but if you're working on those small features, you want to…

00:48:26.910Derek Peters: Yeah, it's impossible to read, you know, read it line by line, but what I'm saying is.

00:48:33.440Derek Peters: Don't wait till you have… 6 features.

00:48:37.470Derek Peters: And thousands of lines of code in different files

00:48:42.420Derek Peters: to, to roll something out. Roll those features out in increments.

00:48:49.830Derek Peters: So… And make sure that it's testing, you know, generate tests with

00:48:56.440Derek Peters: With the feature, the happy path, empty dunk.

00:48:59.260Derek Peters: and off-schema response that fail, or that must fail safe, not crash, make the required CIG on every PR, and have it run through and just test your code. Have it read through your code, and do the testing.

00:49:17.170Derek Peters: So, shipping isn't the finish line. It's where you finally get real information.

00:49:23.910Derek Peters: So, instrument the feature before you announce it.

00:49:26.990Derek Peters: A dark lunch teaches nothing.

00:49:30.320Derek Peters: So… Back to our product, so if we take a look at Cadence first week, on the slide.

00:49:38.260Derek Peters: 61% of people pasted a brain dump.

00:49:43.570Derek Peters: Beating target, an average of, let's say, we got 61% pace of… Brain dump.

00:49:50.820Derek Peters: Target was 50%. Organized, so 6.4 tasks extracted per session.

00:49:57.880Derek Peters: 34% tweak a priority, and then next, you know, why is this priority? And then there's an explainer in there.

00:50:14.320Derek Peters: That's… yeah, that does harden the build. Each component should be a sub-dependent on itself, absolutely.

00:50:22.450Derek Peters: Do you think?

00:50:24.080Derek Peters: So let's bring this… To kind of full circle.

00:50:29.070Derek Peters: So, what we went over was writing a vision, a requirement document, and a roadmap before you even get into coding. Those are…

00:50:39.500Derek Peters: The three most important documents… to me.

00:50:44.860Derek Peters: That can be created and generated before you get into developing a product.

00:50:54.390Derek Peters: And… and those are, like, those are the spec the AI builds from. So… and then we talked about prompting.

00:51:02.160Derek Peters: We talked about, you know.

00:51:05.860Derek Peters: We drove the build straight out of the requirements document, planned first, we tightened the loops.

00:51:12.800Derek Peters: We had it go through each line.

00:51:16.160Derek Peters: We talked about, you know, testing, And…

00:51:20.560Derek Peters: as long as you keep those in mind as you're progressing to build, that's what separates you from, let's say, somebody that's just 5 coding, creating things, which is… is still great, but…

00:51:34.530Derek Peters: Is there proof in the real world? Probably not.

00:51:38.070Derek Peters: the better you can plan the what and the why and the who.

00:51:44.030Derek Peters: The better your product's gonna turn out.

00:51:47.460Derek Peters: And the quality of what you ship is decided long before and long after the model types a single character, a single line of code.

00:51:55.850Derek Peters: So, before we open it up, Two quick things.

00:52:05.040Derek Peters: What are some things, you know, if you're working on some things, and…

00:52:09.440Derek Peters: you've been toying around? It sounds like we have a lot of experienced individuals on the call, too. Like, it'd be interesting, like, what are you working on? That… that would be fun.

00:52:29.200Derek Peters: Okay.

00:52:32.640Derek Peters: I love the interaction.

00:52:34.490Derek Peters: You all are awesome.

00:52:36.800Derek Peters: Okay.

00:52:38.680Derek Peters: So, before I make, before we go on to the next slide, here's…

00:52:45.400Derek Peters: what I actually want you to do.

00:52:48.600Derek Peters: Because… because of talk, you don't…

00:52:53.100Derek Peters: Let's see, because you don't act on entertainment, you know, 3 moves, 3 things that you can do.

00:53:00.030Derek Peters: to kind of help you is write a one-page PVD on the next project or next idea that you have.

00:53:06.430Derek Peters: Turn it into a requirements document, scope it out.

00:53:10.470Derek Peters: Write out, you know, what's in scope, what's out of scope.

00:53:15.640Derek Peters: you know, that keeps it shippable, it keeps it simple, it also makes it easy to develop a roadmap that you can actually deploy, and then make an AI plan first. Like, work with the LLM to make a plan and a roadmap for you. So, before you write a single line of code, have it propose a plan, and then you just approve it.

00:53:35.650Derek Peters: No new tools, it sounds like it's everything that everybody's using.

00:53:40.270Derek Peters: But, but that's it.

00:53:44.680Derek Peters: And before we get into any Q&A that's on there,

00:53:51.010Derek Peters: We… we do have our applications open. If there's anybody that's interested in joining us here at Gauntlet, would love to invite you to apply.

00:53:59.780Derek Peters: And, I also want to say thank you so much for your time.

00:54:04.330Derek Peters: as well.

00:54:05.920Derek Peters: And, we'll also be sharing that information with you as well. Oh, thank you, Sunny. There it is.

00:54:14.310Derek Peters: So that's it.

00:54:15.810Derek Peters: We kind of went through one product example from this ideation to…

00:54:23.250Derek Peters: To shipping the, you know, shipping the features?

00:54:27.110Derek Peters: And kept it into an AI-first planning methodology, so AI-first product planning.

00:54:34.170Derek Peters: So, same loop, real documents, real prompts, and every prompt and every result that you saw here was actually from my iteration going through my planning process.

00:54:46.200Derek Peters: And I will tell you that it does… Make a huge difference.

00:54:54.350Derek Peters: The more planning you can do up front.

00:54:57.550Derek Peters: Gives you a richer product that… Users will actually use.

00:55:04.980Derek Peters: So, the floor is wide open. We got, we got a few minutes. If anybody does have any questions.

00:55:11.700Derek Peters: Be more than happy to answer that for you.

00:55:17.340Derek Peters: So, Mark, you asked, you know, where does the UI design fit into this? It fits in… it fits into the very beginning of, because, like, whenever you're planning, like, your… let's say your product vision.

00:55:32.530Derek Peters: and your product requirements, like the PRD, you could either do it one of two ways. You could have, like, a requirement document for, like, your UI, UX piece of it, and have it, like, two separate documents. I do like to fold it up into the same PRD, that way I'm not having to point, to two documents.

00:55:52.890Derek Peters: or two different versions of the documents, because it can sometimes confuse the LLM, but you can fold that into your requirements document. That's a good question.

00:56:02.250Derek Peters: So, UI design actually follows the same set. You have a vision, you have your requirements.

00:56:08.180Derek Peters: You know, the what, The why, the who.

00:56:11.350Derek Peters: And… It's just… it becomes part of your requirements.

00:56:23.020Derek Peters: So, Josh, you said, is there a benefit to writing out code by hand anymore when AI can just write it for me?

00:56:29.970Derek Peters: I do think there is a huge benefit in having that background and knowledge, because

00:56:37.790Derek Peters: even though, like, let's say you have AI write it, you got the advantage of knowing, like, if you have an issue or, like, a bug or something, knowing how to read and write code is still an important skill today.

00:56:49.850Derek Peters: Because that way, you don't have to have it completely wipe something. If you…

00:56:56.960Derek Peters: you know, hey, I got this bug, find it, and then you realize that it's, like.

00:57:01.760Derek Peters: it rewrote 3 different files, but if, you know, still knowing how to write code, or at least jump-starting, you can even jumpstart writing your… getting your setup and start coding yourself, and then have the AI kind of be, like, a coding assistant to help you out. You don't have to have it do it

00:57:17.310Derek Peters: for you, it does kind of defeat the personal, like, you know, AI-first development.

00:57:22.180Derek Peters: But… Knowing how to write code yourself can also help save time in debugging.

00:57:34.510Derek Peters: Renee, what shipping flag tools are you using? So…

00:57:39.970Derek Peters: I put that into… kind of build it into my roadmap. I like to use Claude Code to… so I have it tied into a roadmap, and then I have each feature kind of incremented and cut out, but I like to keep those into my docs folder, into a README file, and then,

00:57:57.040Derek Peters: point the LLM to it, and then that way we know exactly what, shipping

00:58:02.750Derek Peters: shipping flag, or what feature gets shipped, and which ones to flag. So I use Claude Code for that.

00:58:14.190Derek Peters: And I may have missed Abby's question, so Carl, thanks for bringing it up right there. So, as a PM, I find one of the biggest challenges is slicing AI features into user stories.

00:58:24.920Derek Peters: That are both technically feasible and valuable enough to gain support.

00:58:31.090Derek Peters: In a larger enterprise where AI adoption tends to be slower.

00:58:38.770Derek Peters: close the… let's see, I'm just… I'm reading through this question here. So, how do we close the gap between PM's Quick Planning and Development and Classic Engineering? I'd be interested in what tools you have today

00:58:51.010Derek Peters: On that.

00:58:55.410Derek Peters: And there's so many… or, like, what AI features…

00:58:59.590Derek Peters: It doesn't have to be, like, what you're planning.

00:59:03.660Derek Peters: I have not tested Fable against Opus yet.

00:59:08.610Derek Peters: I just got into it right before this call.

00:59:12.280Derek Peters: I, I can't… bring myself to use Fable just yet. I'm sure it's fantastic.

00:59:18.170Derek Peters: but… burning through those tokens when Opus is

00:59:23.770Derek Peters: In most cases, Sonnet's good for what I need to use it for. Opus, I use for, like, more complex tasks.

00:59:30.850Derek Peters: But for day-to-day, Sonnet's pretty good. I haven't even opened the door of Fable.

00:59:41.710Derek Peters: Well, there's the answer. Josh ran out in an hour, so…

00:59:51.700Derek Peters: Now, one thing I do like about Opus, though, is, like, whenever you don't… like, you have a vision and you want it to interview, I have found that it does do a really good job in interviewing you with some half-baked information and helping you flesh out some ideas into a true vision. It does a really good job with that.

01:00:11.920Derek Peters: Oh, AI tools to develop features, okay.

01:00:17.290Derek Peters: Yes, that is rough in an enterprise environment, because there's a lot of security protocols,

01:00:25.520Derek Peters: in place, and I've ran into that before, as well.

01:00:30.050Derek Peters: and I think from there, it's just more about…

01:00:37.900Derek Peters: selling the benefit of what those tools or features can do, but if you're going to do that, like, let's say Carl and, I think it was Abby, there's a lot of great chats here, but…

01:00:50.680Derek Peters: I would come up with, like, a use case, and…

01:00:54.810Derek Peters: And some proof, like, what you can actually do with it, and how… and…

01:01:00.020Derek Peters: really sell the value of the return on investment that those AI features are going to help you as, let's say, a PM, and work with the team of developers. So an ROI, especially in an enterprise environment, is going to be super important, but the most important piece is the security piece, like.

01:01:18.950Derek Peters: what guardrails are you going to put around it to make sure that that enterprise information stays safe? Because that's what those corporations care about.

01:01:28.130Derek Peters: So, like, if you just come up with a good story and a good use case and show, like, what the return on investment is for the organization.

01:01:35.100Derek Peters: I've had a lot of good luck, getting ears to listen, I should say.

01:01:43.630Derek Peters: Does that help, Carl?

01:01:51.090Derek Peters: He turns into a house of cards.

01:01:57.300Derek Peters: All right.

01:02:00.330Derek Peters: Well, it does look like we're 2 minutes after. A lot of great conversation, a lot of great interactions, loved every minute of it.

01:02:08.970Derek Peters: Just wanted to say thank you all for allowing me to host you this evening for Gauntlets Night School, and really appreciate you coming. I hope you guys have a wonderful night, morning, and evening, depending on where you're at in the world. It's been great. Thank you all.

01:02:39.950Tom Babb: You can just log out, Derek, or you can just end the call.

01:02:43.700Derek Peters: Okay. Oh, no, I can't end it. It just lets me to leave, so you and Sonny were gonna…

01:02:48.070Tom Babb: Yeah, you got it, you got it. We're good, thank you.

Download transcript (.vtt)

FAQ

What is the AI-First Product Loop? +
It is a five-move workflow of a vision doc, a PRD, a roadmap, prompting an agent to build from those docs, and releasing behind feature flags. It turns vibe-coded demos into products real users actually use, with you directing the build as the tech lead.
Why do most AI builds go wrong? +
Because someone asks the model to invent the plan, so the fix is to plan first and have the agent build against clear docs.
How is the AI-First Product Loop different from vibe coding? +
Vibe coding produces a quick demo, while the AI-First Product Loop produces a shippable product by giving the agent structured intent before any code is written.
Why write a vision doc and PRD before prompting? +
They give the agent the context and constraints to build the right thing, so you avoid working out your intent through trial-and-error prompts.
What is Cadence? +
Cadence is a fictional AI task tracker Derek uses to trace one feature from idea to ship across all five moves.
Why release behind feature flags? +
So you can ship safely and in small steps, turning features on or off without a risky big-bang release.
What does being the tech lead mean here? +
You direct the build with specs and judgment rather than poking at it like an end user, and you slow down up front so the whole thing moves faster overall.

What's next?

Keep building with the rest of Night School, or apply to Gauntlet — ten weeks of technical intensity with the best AI engineers we can find.