Learning Library

← Back to Library

Prompt Engineering Lifecycle and Tools

Key Points

  • Prompt engineering is a “wild‑west” space that’s become essential to AI workflows, yet few have mapped out a systematic prompt life‑cycle.
  • The first stage—authoring and drafting—relies on interactive tools (Claude, ChatGPT, Prompt Perfect, Cursor) to iteratively refine wording and clarify mental models.
  • After drafting, prompts are version‑controlled like code, using platforms such as PromptLayer, PromptMethus, git‑based repos, or LangChain to ensure persistence, auditability, and team coordination.
  • Production‑grade prompts then enter a testing phase where multiple versions are evaluated for accuracy, cost, and hallucination risk, using tools like PromptFlow, Hegel’s evaluation suite, or custom eval frameworks.

Full Transcript

# Prompt Engineering Lifecycle and Tools **Source:** [https://www.youtube.com/watch?v=V0YhpeSOuzk](https://www.youtube.com/watch?v=V0YhpeSOuzk) **Duration:** 00:16:41 ## Summary - Prompt engineering is a “wild‑west” space that’s become essential to AI workflows, yet few have mapped out a systematic prompt life‑cycle. - The first stage—authoring and drafting—relies on interactive tools (Claude, ChatGPT, Prompt Perfect, Cursor) to iteratively refine wording and clarify mental models. - After drafting, prompts are version‑controlled like code, using platforms such as PromptLayer, PromptMethus, git‑based repos, or LangChain to ensure persistence, auditability, and team coordination. - Production‑grade prompts then enter a testing phase where multiple versions are evaluated for accuracy, cost, and hallucination risk, using tools like PromptFlow, Hegel’s evaluation suite, or custom eval frameworks. ## Sections - [00:00:00](https://www.youtube.com/watch?v=V0YhpeSOuzk&t=0s) **Systematic Prompting Lifecycle Overview** - The speaker outlines a structured life‑cycle for crafting LLM prompts—detailing stages like authoring, drafting, and versioning—and highlights the specific tools suited to each phase. - [00:04:11](https://www.youtube.com/watch?v=V0YhpeSOuzk&t=251s) **From Prompts to Agentic Workflows** - The speaker explains how prompts evolve into guided, multi‑step agents that rely on frameworks, tooling, and production‑level safeguards such as traceability, safety, and governance. - [00:07:19](https://www.youtube.com/watch?v=V0YhpeSOuzk&t=439s) **Prompt Ideation with Hey Presto** - The speaker highlights the common failure to define output formats during early prompt drafting, introduces the Hey Presto tool designed to guide fuzzy intent formation and format‑specific prompt tuning, and demonstrates its use for creating code prompts for a travel app. - [00:11:13](https://www.youtube.com/watch?v=V0YhpeSOuzk&t=673s) **Auto-Generating Slide Decks with Hey Presto** - The speaker shows how a single prompt in Hey Presto instantly creates a polished presentation, highlights seamless integration with Claude and ChatGPT, and candidly addresses audience skepticism about product promotion. - [00:14:21](https://www.youtube.com/watch?v=V0YhpeSOuzk&t=861s) **From Idea to Intent: Hey Presto** - The speaker outlines the need for a simple, community‑driven tool to transform casual ideas into precise AI prompts, describes building Hey Presto to fill that gap, and invites feedback and collaboration. ## Full Transcript
0:00Prompting is really weird because it's 0:01the most wild west software space I've 0:04ever seen, but it's also a highly 0:07leveraged critical part of AI workflows. 0:09There are dozens and dozens of prompt 0:11tools. And as far as I can tell, very 0:13few people have laid out or thought 0:15through the overall life cycle of a 0:18prompt and how we think about prompting 0:20systematically. That is what this video 0:22is for. I'm going to lay out how I think 0:24about prompting in a life cycle. And I'm 0:27going to lay out the tools that I think 0:30are relevant at each stage. So, let's 0:32jump into it. Let's start where most 0:34people begin. Authoring and drafting. 0:36So, you're writing, you're rewriting, 0:38you're testing prompt text. I'll be 0:40honest with you, a lot of people are in 0:42Claude or in Chad GPT for this, right? 0:43And they're like, "Hey, make it better. 0:45Hey, make it better." Or people are 0:46using tools like Prompt Perfect. Some 0:49people are in cursor if they're in code. 0:51This is all about hands-on 0:53experimentation and wording refinement. 0:55We're not testing the value of the 0:56prompt. We're just trying to figure out 0:58like there's a mental model we have of 1:00the perfect prompt. And whether it's 1:02right or not, we're trying to make the 1:05prompt that we write fit that model. And 1:07that's that authoring and drafting 1:08stage. And what we find in this 1:10situation is that the LLMs help us to 1:12organize our thoughts and help us to 1:14take these messy thoughts and kind of 1:16clean them up. The next stage in 1:18prompting is really around getting 1:20serious about versioning. And so teams 1:23will start to or individuals will start 1:25to say this is a prompt I use often. How 1:28do I keep it track of it when it needs 1:31to change? And so teams will store their 1:33prompts. They'll name them v1 v1.1. 1:36They'll describe differentiation with 1:38them. They'll diff them. They'll name 1:40them. And prompts are then almost 1:43treated like code. they become artifacts 1:45in the business because they get reused 1:47so much. Prompt layer definitely offers 1:50tooling capability here. Prompt methus 1:52does this. There are git based 1:54approaches that do this. There are a lot 1:56of other approaches as well. I know that 1:59lang does this. The idea is you are 2:02trying to solve for the persistence of 2:04the prompt. So there's one record for 2:06it. You need to make sure it's auditable 2:08and you make need to make sure that you 2:10can enable team level coordination on 2:11the prompt. The next stage after 2:13versioning is making sure that the 2:16prompt is regularly tested. And this is 2:18especially true if it's a production 2:20grade prompt that will be used for an 2:22LLM production system. And so you need a 2:24situ a solution where you can compare 2:26multiple prompts. You can evaluate 2:28outputs for accuracy. You can evaluate 2:30cost and hallucinations etc. Tools for 2:33this include Hegel's prompt tools prompt 2:35flow eval components prompt methus. 2:37Again there's some custom eval 2:39frameworks. There's a lot of other tools 2:41around eval. A lot of people are writing 2:44their own eval solutions and finding 2:46that those are more flexible and more 2:48effective because they allow them to 2:49write to the detailed setups they have. 2:52One of the places where we start to 2:53differentiate between individual 2:56builders and teams is right here at 2:59evaluation and testing. Because before 3:01if you were a serious individual, you 3:03might still have your own versioning 3:05system. You might still have a little 3:07notion database where you keep track of 3:09your individual prompts and it works for 3:11you. When it gets to eval though, teams 3:14who are using production grade prompts 3:16will build entire suites of tests, 50 3:20tests, 100 tests that they run in an 3:22automated fashion in a pipeline against 3:24a new version of a prompt. Whereas 3:25individuals, we're very unlikely to do 3:28that. we will probably test in an ad hoc 3:30way or if we are super super organized 3:33we'll have a library of 10 or 15 queries 3:35that we are running with the prompt to 3:37see if it works better but it's rare to 3:40get super serious and so as we move 3:42forward in this flow we see that these 3:44prompt tools start to become more team 3:46oriented this is one of the things that 3:49makes it really complicated when you're 3:52doing prompting because prompting is 3:53both an individual productivity choice 3:56and also something that supports teams 3:58and dealing with that makes it hard to 4:01write good software. But we're not done 4:03with evaluation. After evaluation comes 4:07constructing workflows or automation. 4:09And this is when prompts become mere 4:11steps and workflows. Like the prompt 4:13becomes a guidepost for an agent that 4:16might have tools, that might have memory 4:18that it calls, that might have 4:20conditional logic. And so then you begin 4:22to need to build like with Google's 4:24agent kit, with Langchain, with 4:27Langmith, with Hegel and Prompt tools, 4:30with prompt flow, with React agent for 4:32frameworks. This is all about multi-step 4:34automation and agentic behavior and 4:36prompting. Again, like this is part of 4:38why it's a tricky subject. It bleeds 4:39into this area because really, if you're 4:42trying to build an agent, you have to 4:44consider the prompt as the beating heart 4:47of the agent. The prompt is what helps 4:49you to predictably guide the agent. 4:52Finally, you have deployment tools. So, 4:54prompts are embedded in real 4:56applications and they need to be 4:59tracked. They need to be up all the 5:01time. They need to run correctly. Prompt 5:04layer comes back through here. Langmith 5:06comes through here. There are model APIs 5:08that are directly available from OpenAI 5:10and Enthropic. You have to have 5:12production robustness, safety, 5:13traceability, governance. You can see as 5:16we've gone through this overall flow 5:18that we have gone heavily into the teams 5:21and companies world where prompting is a 5:24piece of code that companies need to 5:26maintain and they need production 5:28tooling for that system. I am here to 5:30tell you that you are missing a stage in 5:33the prompting layer. Let's go back 5:35through. What if we thought of our first 5:38piece, authoring and drafting as stage 5:41two, not stage one. Because it is. 5:44Because when you think about where you 5:46want to go with prompting, it's actually 5:48intent formation and discovery that has 5:51to happen first and then you get into 5:53authoring. And what I have found sitting 5:56down with people over the last few 5:57months is that intent formation for 6:00individuals is really hard. And this, 6:02I'll be honest with you, this is 6:03definitely something that is more true 6:06the earlier you are in your AI journey 6:08and you are still trying to figure out 6:10how how prompting actually works. But it 6:13remains true the farther you go in your 6:16journey if you need help to trade time 6:19for expertise. Basically, if you're 6:21trying to write a prompt quickly and 6:22formulate your intent quickly and you 6:24don't have the time to do it and you may 6:27be an advanced prompter, you still run 6:29into the same issue because you have a 6:31fuzzy goal. You summarize this, draft a 6:33plan, analyze a sentiment, and now you 6:35need to get to a structured, 6:36unambiguous, high lever prompt that is 6:39going to clarify the objective and the 6:41constraints and the steps. And you know 6:42that you should, right? Like we know we 6:44should. It's like eat your vegetables. 6:46There are not great tools at this stage 6:48and most people use chat GPT. I'll be 6:50really honest with you, most people use 6:51chat GPT for this or they use Cloud for 6:53this and that's fine. But what I have 6:55found is that those tools by themselves 6:58aren't super well suited to the kind of 7:01intent formation that we're doing at 7:03this stage. And there's a really really 7:06simple reason why. When you're crafting 7:08a prompt in Claude or Chad GPT or 7:11Gemini, you are crafting it and 7:13implicitly you are assuming the prompt 7:15will work in that particular LLM. 7:16there's not a cross LLM compatibility 7:19check going on there. You are also 7:21typically not getting the LLM to help 7:25you think through what is the output 7:28format for this prompt and artifact that 7:31you're building. And I that that sounds 7:33really abstract. So let me make it 7:35concrete for you. If you have a fuzzy 7:37intent, you typically know the output 7:40needs to be a deck at the end of this 7:42process. But what you don't often do is 7:44you don't often say, "Please tune this 7:46prompt in such a way that it's specific 7:49for writing a deck because you're at the 7:51fuzzy stage. You're trying to think 7:53through the content first." And content 7:54comes before format. What I'm picturing 7:57for you is the realworld complexity that 8:01I feel that others feel when they're 8:03trying to craft prompts. And to be 8:04honest with you, there hasn't been a 8:07great tool for that. And that's why I 8:09built one. I built Hey Presto to solve 8:12for specifically this ideation intent 8:14formation piece. And if you come back to 8:16me and you say, "Nate, I'm using Chad 8:17GPT." I will say, "Bless you. Have a 8:20wonderful time. This tool is not for 8:21everyone." Let me show you how it works 8:23and then we'll get into the rest of it. 8:24All right. Here we are in the prompt tab 8:28on Hey Presto. And I have just pasted in 8:30some very rough notes for an app I keep 8:32trying to build called my family travel 8:34app where it shows me the different 8:36destinations around the world. and I 8:37say, "Please help me craft a prompt to 8:39build code for my app." All I do is I 8:41say, "I want to build code." And it 8:42gives me an expanded prompt. Look at 8:44that. It's going to give me lots and 8:46lots and lots of detail on the travel 8:49app. It's going to give me a suggested 8:51file structure, suggested data model. 8:53And keep in mind, all of this is 8:55editable. I have made this is ideation 8:58stage, right? I have made no 8:59commitments. Nothing is running. I can 9:02go back through and edit all of this. I 9:05don't have to believe or buy any of 9:06this. And I can also change it. And so 9:08if I want to change the stack, this is a 9:11Python stack for those of you who are 9:13engineers. If I want to change it and 9:15say, look, I'm not writing this in 9:16Flask. I'm writing this in React. That 9:18makes more sense. All I have to do is 9:20append a uh write this in React uh and I 9:25can just regenerate and it will just 9:27regenerate the uh regenerate the prompt 9:29and it runs pretty quickly. While we're 9:31uh letting it show up, you have 9:33different tones you can get with this. 9:34You can select different paragraphs or 9:36options. JSON, email, table, numbered 9:39list, step by step, structured sections. 9:41Oh, it generated. Great. Um, and look at 9:45that. It's getting into React. It's 9:47getting into Tailwind, and it just 9:49rewrites it. And so, it takes the detail 9:51you give it, and in this case, I gave it 9:54a fair bit, and it will turn it in to a 9:58high-grade prompt. And if you're 9:59wondering, does this actually work? I 10:01will show you what I've been working on 10:03in lovable. The answer is yes. So, here 10:05we are in Lovable. We're crafting a cute 10:07little app. We can add our destination. 10:11It's not done yet. We can add a family 10:12member. Little energy levels here for 10:15the family member, whether they're an 10:17adventure, relaxer. The app isn't done 10:19yet, but it's coming together. And I was 10:21able to use the prompt to craft the 10:24initial piece of the app. And this is 10:26not just for lovable, right? Like, this 10:28is something I could have put into 10:29anti-gravity from Google. It's something 10:31I could have used uh for any other 10:32vibecoded application or I could even 10:34even have used it in cursor or other 10:36spaces to get started. It is designed to 10:39be agnostic of the tool but focused on 10:42the outcome you want which I think has 10:44been a space that's been missing 10:45strategically for us because when we're 10:47forming intent when we're at that fuzzy 10:50initial stage that I'm arguing we don't 10:52have good words for we need the 10:54flexibility to go anywhere. Let me give 10:56you one more example. This one's around 10:58PowerPoint text. So, here we are with a 11:01PowerPoint deck summary of Andre 11:03Carpathy's famous early 2025 software 11:063.0 talk. It's gorgeous. It looks great. 11:10Software is changing again. I just love 11:13the look of this. All generated by one 11:15prompt from Hey Presto. Super easy. And 11:18I got oh, I don't know, 15 16 I got 20 11:21slides out of it. Right. Embrace the 11:22Iron Man suit. I love this. What a 11:24finish. It makes sense. It's ready to 11:27It's ready to share. 11:29So, how did I do this? Very simple. All 11:32I did was I said, I want to make a deck 11:34about this article. I put in my notes. 11:36Here's my notes on Andre's article. Dum 11:38dum dum. I said, I want it to be a deck. 11:41And I got create a deck, right? It has 11:43everything. It's laid out. It's got the 11:45slides laid out. It's all perfect. And 11:47you can change this, right? You can 11:49adjust it. And by the way, if you're 11:50wondering, why would I go to Hey presto? 11:53That's sort of annoying. Uh, I get it. 11:56And maybe you don't want to. But if you 11:58do want to, we've made buttons so it's 12:00really easy to go into Claude or to go 12:02into chat GPT directly. And so you can 12:04open it up and it will just pop open the 12:06prompt right in Claude. Here's the 12:08prompt opened up right in Claude. Super 12:10super easy. Now, this is the part where 12:12people usually roll their eyes because 12:13they're like, "Oh, Nate has sold out. 12:15He's trying to like shill and like push 12:16his stupid little product on us." And 12:18I'll be honest with you, I could be the 12:21person who does that and I'm not going 12:22to because I'm not interested in saying 12:26this is the best product for everything. 12:27That is why I spent 15 minutes going 12:30through all the stages of the prompt 12:33tool chain and talking about all the 12:34other great tools that you have to go 12:36after for prompting. I do not believe in 12:38a world where there is one prompt tool 12:40for everything. And that in turn drives 12:43the way I'm thinking about pricing. If 12:45we are thinking about pricing for prompt 12:47tooling, it needs to change depending on 12:49whether you're a team solution or an 12:51individual solution. And in my case, I 12:53want to make this something that is very 12:56very affordable if you're already a part 12:58of the Nate Substack community. I don't 13:00want this to feel extra fancy or 13:02complicated. And so the easiest way I 13:04can figure it out is to just say if you 13:06are a member of the Substack, you will 13:09get a coupon code for 70% off forever on 13:13the prompt tool. which makes it like 13:15seven bucks a month or something. And 13:17I'm doing that because I want it to be 13:19super easy for you to use. And if it's 13:21useful, great. And fantastic, right? And 13:24on my side, I'm committing to continuing 13:26to make it better. And so, I'm going to 13:28be opening up a Slack channel into my 13:31working Slack for people who are using 13:34Hey Presto so that they can give direct 13:36feedback to me and a couple of other 13:37builders who are working on this so that 13:39we could make it better. My goal is to 13:41solve that fuzzy intent stage. So if you 13:45remember, we have these multiple stages 13:47in the prompt tool chain. And I'm going 13:48to show you what this looks like using a 13:51slide created by Nano Banana. So here we 13:53are. These are the six stages of the 13:55prompt life cycle. Right? We have spent 13:57almost all of our tooling and thinking 13:59time around stage two and stage three. 14:02Writing, rewriting, testing prompt text. 14:05Stage three, storage and versioning. 14:07This is where individuals are really 14:08seeing prompt tools. And then there's a 14:10lot of team level prompt tools as we get 14:12into these more purple stages, right? 14:13Evaluations, workflow construction, 14:15deployment, and production integration. 14:18I I really think this is important. If 14:21you want to get from casual ideiation to 14:24like clean intent, there needs to be a 14:27good tool for that. And I just haven't 14:30found a tool that more easily helps me 14:32to name this is what I want to build. 14:35It's a deck. It's a doc. It's 14:36communications. You can do Slack 14:38communications. It's code. Whatever it 14:40is. And that's why I built Hey Presto. 14:43If it's not for you, if you want to hack 14:44around and chat GPT, bless you. It may 14:47work well. I've certainly done it 14:49before. And if you disagree with me and 14:50you think intent formulation is not 14:53correct for prompting, I would love to 14:55hear what you think the correct initial 14:57stages because I do not believe the 14:59correct initial stage is writing or 15:01authoring. Either way, I had fun making 15:03Hey Presto. I think a lot of the future 15:06of tooling and tooling development is 15:08talking with communities, hearing what's 15:10going on, and trying something. And I'm 15:12offering Hey Presto in that spirit. If 15:14you're part of my Substack community, 15:16this is going to feel really affordable. 15:18And it's going to be something that I'm 15:20going to be plugged into on Slack so 15:22that I can evolve it with you. My goal 15:24is to make it useful. I don't need to 15:26make it super big. I just want to serve 15:28the community around a need that I 15:30found. I also think it's good to 15:32actually build things and launch them if 15:34you talk about AI all the time. And so 15:36that's another reason I'm putting this 15:38out there. I think it's important to 15:39have AI tools that solve problems if I 15:42talk about AI tools a lot. And so that's 15:43part of the reasons why I've been 15:44involved in building this and digging in 15:46and constructing this. If you guys are 15:48interested, I'm happy to do a little bit 15:49of a deeper dive on how I built the 15:51tool. I wanted to start with a prompting 15:53structure first because I think that we 15:55have missed a framework for prompting 15:58and I felt poorer until I could name the 16:01stages. I felt like I had trouble 16:03understanding my own thinking until I 16:05could name the different stages of the 16:06prompt tool chain and talk about the 16:08tools because this is a world where like 16:10no shade on all these other tools. 16:12They're great. That's part of why I took 16:13the time to introduce them. Like I think 16:14that you need multiple tools for these 16:17different stages and my goal is just to 16:18help you with that intent and initial 16:21sort of piece that I think hasn't been 16:23as clean as it needs to be. So, hope you 16:26enjoyed it. If nothing else, now you 16:28have some vocabulary to talk about the 16:29stages of the prompt and I think that 16:31will be helpful and maybe you got some 16:33good tool tips for some of the other 16:34later stages. That's another reason I 16:35made this video. I think there's a lot 16:37of great prompting tools out there. 16:38Cheers.