16:20Eric Tech · Tutorial
Understand-Anything vs Graphify: I Tested Both on My SaaS
A 16-minute live comparison of two Claude Code plugins that map your codebase — tested head-to-head on a real SaaS across six dimensions.
May 27thThe argument in one line.
Lauren Tan's open-source Pstack packages six stages of AI-assisted coding into reusable skills, and its most valuable ideas are running a design through multiple competing AI models before writing code, and attacking your own diff with an adversarial review before anyone else does.
Read if. Skip if.
READ IF YOU ARE…
- A solo developer using Claude Code, Cursor, or a similar agentic coding tool who wants a repeatable, named process instead of ad-hoc prompting.
- Someone maintaining an existing codebase who needs skills for learning how and why the code works before changing it.
- A developer curious how multi-model AI workflows, running the same design through several models and judging the results, work in practice.
SKIP IF…
- You're starting a project from scratch. Pstack's planning skills are built for existing codebases, not greenfield builds.
- You're not using Claude Code, Cursor, or another agentic coding tool with a skills or plugin system.
The full version, fast.
Eric Tech breaks down Pstack, Lauren Tan's open-source pack of Claude Code and Cursor skills, across six lifecycle stages: plan, design, build, test, deploy, and maintain. Pstack assumes you're working in an existing codebase, with skills to learn how and why code works before changing it. Its standout stage is design: an architect skill sketches types, an arena skill runs the same design through multiple AI models in parallel, and an interrogate skill judges and aggregates the results into one report. Build and test stages add TDD, a no-comments rule that stops AI from trusting stale comments over real code, and an adversarial attack-the-diff review. Eric is adopting attack-the-premise and type-system-discipline into his own skill set.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:55
01 · What is Pstack
Who Lauren Tan is, and Eric's diagram breaking Pstack into six stages of the AI software development lifecycle.
00:55 – 02:10
02 · Planning Skills
help/why/teach for learning existing code, figure-out and show-me-your-work for breaking tasks into tickets. Eric says he skips this stage for greenfield work in favor of Matt Pocock's or Superpowers skills.
02:10 – 03:16
03 · Design Skills
"Settle the code shape": architect sketches types, arena fans the design out across multiple AI models, interrogate judges and aggregates the results into one report.
03:16 – 05:33
04 · Build Skills
TDD (failing test first), typescript-best-practices, and deslop/no-comments to strip AI junk and stale comments out of a diff.
05:33 – 08:18
05 · Testing Skills
create-verification-skill builds end-to-end proof (launch, drive, capture evidence), swarm parallelizes test workers, attack-the-diff reuses interrogate plus a thermonuclear-code-quality-review pass.
08:18 – 09:54
06 · Deploy Skills
make-pr-easy-to-review, babysit the merge, fix-merge-conflicts, get-pr-comments. Maintain stage has no dedicated Pstack skill, Eric points to his own super-collect skill instead.
09:54 – 11:51
07 · Sponsor: Blotato
Mid-roll pitch for Blotato, a social-media automation tool connected as an MCP server, using Eric's own Instagram growth (0 to 7,191 followers) as the case study.
11:51 – 12:36
08 · Extra Skills
potato-mode routes a problem to the right skill, setup-pstack configures models per job, catch-up and plain-talk give fast recaps when switching context between projects.
12:36 – 14:56
09 · Principles + Learnings
Why principles are stored as indexed skills instead of one system prompt. Eric highlights attack-the-premise (stop re-patching, question the shared assumption) and type-system-discipline (make impossible states unrepresentable) as the two he's adopting.
Lines worth screenshotting.
- Pstack assumes brownfield work: its planning skills are built to learn and extend an existing codebase, not to start a project from scratch.
- The design stage runs the same problem through multiple AI models in parallel, then uses a separate model to judge and aggregate the results into one report.
- Running one model's design against another's is the core difference between Pstack and single-model workflows like Matt Pocock's skills.
- A no-comments rule exists because AI agents sometimes trust a stale comment over what the code actually does, so removing comments removes that failure mode.
- Separating write-the-test-first from prove-it-on-the-real-app treats unit tests and end-to-end verification as two different skills, not one.
- A verification skill that launches the app, drives it like a user, and captures screenshot proof is what lets an agent trust its own claim that a feature works.
- Attacking your own diff with an adversarial model, before a human reviews it, catches regressions a single pass would miss.
- Principles live as separate, indexed skills instead of one giant system prompt, because a bloated system prompt gets read on every single message whether it's relevant or not.
- Attack-the-premise exists for the failure where an AI fixes a bug, the bug returns, it fixes again, and the bug returns again: the fix is to question the shared assumption, not attempt a third patch.
- Type-system-discipline means designing types so an impossible state, like a payment that is both paid and pending, literally cannot be represented in code.
Six stages that structure how AI ships code.
AI CODING WORKFLOW
Pstack packages six coding stages into reusable skills, and the ideas worth stealing are running a design through multiple models before building and attacking your own diff before anyone else does.
01What is Pstack
- Pstack is Lauren Tan's full open-source skill pack, built from her experience as a React core team member and principal engineer at SpaceX and Cursor.
- Treat it as a six-stage map (plan, design, build, test, deploy, maintain) rather than a single tool, so individual stages can mix with skills from other packs.
02Planning Skills
- The help/why/teach skills exist to make an agent learn how code works and why it was built that way before touching it, which matters most on an unfamiliar codebase.
- Figure-out and show-me-your-work turn a big task into smaller tickets and log the reasoning to a doc, so the decision trail survives past the chat session.
- Pstack's planning skills are weakest for greenfield builds; pick a lighter-weight framework there instead of forcing this one to fit.
03Design Skills
- Architect sketches the types first and writes code second, forcing the shape of a feature to get decided before implementation starts.
- Arena runs the same design problem through several different AI models at once, instead of trusting one model's first answer.
- Interrogate then has a separate model judge the competing designs and aggregate the best parts into a single report, rather than picking one output wholesale.
04Build Skills
- TDD here means writing the failing test first, then the fix: the discipline comes before the code, not after.
- A dedicated typescript-best-practices skill enforces strict typing rules on every build, useful on any strongly-typed stack even outside React.
- The no-comments rule removes explanatory comments from a diff because an agent sometimes trusts the comment over the actual behavior of the code, letting the two drift apart silently.
05Testing Skills
- Create-verification-skill builds a dedicated skill that launches the app, drives it like a real user, and captures screenshot or video proof, rather than trusting the agent's word that a feature works.
- Swarm runs multiple agents on different parts of the test surface in parallel and aggregates the results, instead of one agent working the whole scope sequentially.
- Attack-the-diff reuses the interrogate skill to have multiple models try to break a finished change before a human reviews it, treating adversarial review as a repeatable stage, not a one-off ask.
06Deploy Skills
- Make-pr-easy-to-review tidies a pull request for human reviewers as its own distinct step, separate from writing the code itself.
- Babysit watches a merging PR and triggers fix-merge-conflicts when a teammate's parallel merge collides with it, instead of leaving conflict resolution to be discovered later.
- Get-pr-comments turns review comments into a worked list of fixes, closing the loop between human feedback and the next agent run.
08Extra Skills
- Potato-mode acts as a router: describe the problem in plain language and it picks which skill in the pack applies, which matters once a skill pack grows past a handful of entries.
- Catch-up and plain-talk skills exist specifically for context-switching between projects, giving a fast recap of what was done so far instead of re-reading the whole session.
09Principles + Learnings
- Principles are stored as individual, indexed skills instead of one system prompt, because a system prompt gets re-read on every message whether or not it's relevant to the current task.
- Attack-the-premise is the fix for a loop where the same bug keeps coming back after repeated patches: stop patching and question the shared assumption behind both failed fixes.
- Type-system-discipline designs types so an impossible state, like a payment being both paid and pending at once, can't be represented in the code at all.
Terms worth knowing.
- Pstack
- Lauren Tan's open-source pack of AI coding skills, workflows, and principles for Claude Code and Cursor, organized across six software-lifecycle stages.
- Brownfield project
- A software project built on an existing, already-functioning codebase, as opposed to a greenfield project started from scratch.
- Arena (skill)
- A design skill that runs the same problem through multiple AI models or agents at once, then keeps the best parts of each competing result.
- Interrogate (skill)
- A review skill that has a separate AI model judge and aggregate competing results, reused in both the design stage and the testing stage.
- Deslop
- A cleanup pass that strips AI-written junk out of a code diff, such as stale comments, deep nesting, or untyped variables, before it's reviewed.
- Thermonuclear Code Quality Review
- An intentionally harsh, adversarial review skill that pushes for major structural simplification of a finished change.
- Attack the premise
- A principle for breaking repeat-fix loops by questioning the shared assumption behind two failed fixes, instead of attempting a third patch.
- Type-system-discipline
- A principle for designing types so that invalid combinations of state are impossible to represent in the code at all.
Things they pointed at.
Lines you could clip.
01:44
“I won't probably use her planning skill because you can see that she's not really big on a planning stage.”
04:54
“The reason why she says no comments in the code, because oftentimes AI sometimes just look at the comments rather than the code.”
13:39
“It just feels like AI just not able to fix that bug entirely... stop and have AI here to question about the assumption.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
This is Lauren Tan. She's currently the principal engineer at SpaceX, the technical staff at Cursor, and also a staff engineer at Meta. And you can see that she's also the core member at the React team.
And recently, she has open sourced her entire knowledge, her workflows, skills, and principles into an open source repository called Pstack. And that's why in this video, I'm going to share with you exactly what is Pstack, what are the skills that we have, as well as I'm going to share with you how are they different compared to the skills that we have reviewed in the past.
And of course, if you stick to the end of this video, I'm also going to share with you my core takeaways after I review her own software development skills. So with that being said, if you're interested, let's get into the video. All right, so to get started, you can see here that I create an entire diagram to break down all the workflows that she has, as well as her skills and also the principles that she follows whenever it comes to writing code.
and i have broke down her skills into six stages which is the entire ai software dome life cycle which i have made a video on which you can check out here so in this case let's start with the first stage which is the planning stage but essentially what it does here you can see for planning stage is whenever you build the applications you start with planning but her skill here you can see is mostly working for brown field project which means that the project here already have existing code on it right you can see that we have two workflows learn the code first and also plan a big job right so for example let's say if you want to learn how the code works you have the help skill if you want to know why the code works this way then we have the why skill and the teach skill here combines the how and the why and that's basically the workflow here to learn the code first now if you want to break a task or a project into a smaller tickets you can also be using the figure out skill and also show me your work skill which will basically log these decisions onto a proper doc or md file
But personally, I won't probably use her planning skill because you can see that she's not really big on a planning stage. It's mostly just focusing on improving the existing applications, like making a big migrations or big change. Then you can see that her skill is mostly up to par for that.
But in terms of like solo builders or you try to build an application from scratch or whatever it is. My preference are mostly on the Matt Pocock skills or the Sue Power skills. So those are the two skills that I highly rely on rather than using her planning skills.
And in that stage we have is the designing stage. Now that's going to be the part where we're going to design how we're going to do the implementation before we actually going to write a single line of code. So right here, you can see we have a single workflow here called settle the code shape.
Now you can see that we have the architect. Now the architect here is going to trigger the arena here to basically try to have multiple different models and multiple different AI agents here to fend out and try to do the design. Okay.
And once the, we have multiple different agents here come back with multiple different designs, they create it. We can then do trigger a skill called the interrogate, which will basically have a model here to judge the results and then be able to aggregate the results back into a single report for the design. So you can see that it's really different compared to Matt Pocock's or any other skills that we have reviewed in the past, because in the past we were mostly rely on just one single model, but her.
approach is mostly rely on having multiple different models right like opus gpt or grog whatever models you could think of and having them to fan out and then try to compare the results and be able to take the best parts from each models here and then aggregate it into one single report for the design before we move on to the next stage, okay?
So that's her design stage right there. And then the next part we have is the building stage. Now the building stage here, you can see we have two workflows.
One is to follow the test -driven development here to write in code. Now you can see that for writing test first is actually also using the TDD skill. And I have also judged the comparisons between TDD skills from Pstack, Matt Pococks, and also Sue Power.
And here you can see this is the entire breakdown for the difference between all skills from TDD. So if you want to figure out which one is the best TDD skill, you can check out this one, right?
For example, quick fix bugs. new features, strict teams disciplines, right? You can see that we have different types of skills because there's so many test -driven development skills that we have reviewed.
So if you want to know exactly which test -driven development skills is best fit for you, you can check it out inside of this page right here, which I'll make sure to put this in the end of this video. But essentially, once we have our TDD skill built, right, we also have our TypeScript best practice because Lauren Tan here, the creator of PStack, she's mostly focusing on React, right?
So React is mostly using TypeScript. So that's why you can see that we have a skill here called TypeScript Best Practice, which will enforce AIs here to building application using the TypeScript Best Practice. And furthermore, we also have our cleanup diff.
So essentially what this do here is we have our dslop here. First, it checks the difference between the main branch as well as the current branch that we have. And then furthermore, you can see there's some additional checks, right?
For example, check for extra comments. I'll try cache like basically like air exceptions or any variables that doesn't have types, right? Deep nestings.
Those should be removed inside of the code, right? It follows the best practice here in terms of like writing code that are not AI slot. And the goal here is to try to keep the behavior the same by running tests to make sure of that.
And it gives you a one to three sentence summary by the end of it. Now, one thing that I want to point out here is you can see remove extra comments. Now, why does she actually want to remove extra comment here?
Because here you can see it has a skill called no comments. Now, the reason why she says no comments in the code, because oftentimes AI sometimes just read the comments instead of the code. So what happened is this.
Let's say we have a comment, say that this function returns A, but this function in the code, it returns B. So what do AI listens to? Well, the AI sometimes just look at the comments rather than the code.
So that's why you can see she removes the comment part because this way it's going to remove the noise so the AI actually look at the code rather than the comments. So you can see that that's the entire building process for Loren 10's PStack. Okay, so next stage we have here is testing.
Now you can see that we have three stages for testing, three workflows, and let's walk through these workflows one by one. Now, the first workflow we have here is prove the application works. Now, here you can see it uses a skill called create verification skill.
Now, essentially, how is it different compared to TDD skill? Because TDD skill is mostly focusing on writing functional or unit testing, and then create verification skill is mostly focusing on creating end -to -end tests, right? Now, one thing to know here before you create end -to -end testing is that you have to set up the environments, you have to set up the database, you have to set up the servers, you have to set up...
pretty much how user is going to launch it, right? You can see launch the applications, you know, be able to diagnose it if there's any problem with it, be able to drive the actual user behavior, collecting the evidence, do the cleanups. When it says clean up part, maybe sometimes it needs to tear down that temporary database or temporary servers that you have spin up and having some helper function here to help you along the way, right?
So you can see it's going to writing a skill here that builds this kind of end -to -end flow, right? And rather than just writing an end -to -end testing, yes, writing end -to -end testing is key, but building that entire flow, how AI agent here is going to spin up, what commands is going to run, that requires a skill here to create this kind of verification, right?
And you can see there's some additional rules, like one feature per folder, rather than just putting features in one single folder, right? We want to divide that into different folder for different features and then try to run its own steps once end -to -end. and then points you to the maintaining skill right so and here you can see there's additional skills after that it's called swarm so rather than just have one ai agent here runs the entire thing end to end right does the launching does the testing for different pages we're going to swarm by having multiple different ai agents here right you could also choose different model here and having multiple different ai agents to focus on small parts of it and at the end aggregate it into one single report okay so that's basically They prove it on the real application part.
Now, the last part that we have here is attack the difference. And you can see that it's actually reusing the interrogate skills from the earlier in the design scope, right? You can see that we have the interrogate to have multiple different AI agents here, try to break the changes.
So essentially what it do here is it knows the difference, right? So compare the main branch versus your current feature branch, see what are the code changes that it has made. And then what we do here is we have different AI models here and try to spin up and try to see if we can actually break.
this change that we have did, right? And then we're also using a skill called Thermonuclear Code Quality Review, which I have no idea how they come up with this name, but you can see it does a extra harsh review layer to see if there's any way that we can be able to make the code here much more simpler, right? So it's very similar to the ponytail skill that we have reviewed on our channel before.
And you can see that essentially, try to see if we can be able to refactor the code even better, right? So those are the sets of skills, those are the sets of workflows that's in the testing stage, which are really, really interesting. Okay.
Now the last stage that we have here is deployment stage and deployment stage. You could think of it like we're trying to deploy the feature, right? So now we have everything that all tested are all ready to merge to the pull request.
So now what we can do here is first of all, make a pull request easy review, but obviously everyone has their own. writing style even our support skill we already have our own style for this so it really depends on how you want to have your pull request review you can also reference her skill for the pull request easy to review skill and see how you can do to tidy up for the pull request but essentially you can see that we have a set of skill for the entire pipeline like people to babysit the entire pull request right let's say we're going to merge it okay let's try to babysit it to make sure that it's actually able to fully merge we have the fixed merge complex let's say you're about to merge the pull request someone else on your team here is also merging the that too, which caused a complex, what's your changes, then how do we actually fix it?
And also get pull request comments. So list of the review comments and also fixing the CI, which is the CI CD pipelines, the CI checks, right? So these are the things you can see as she has built out this entire pipelines.
So if you're interested for that, you can also check it out as well. Okay. Now, essentially you can see for stage six, which is maintaining.
So how can you actually move to collecting these issues and cycle back into the planning stage you can see that she doesn't have any skill for that but we do have review tons of skills for maintaining skills like our super collect from super board skills you can also check that as well okay so pretty much those are the six stages right that we have from the software development cycle from her skills now we're going to take a look at is some of the secondary skills that you can literally trigger anytime anywhere you want inside of your development and we're also going to take a look at her principles at the end of this video Now, before we move on to the next part, if you have applications and your struggle with marketing, well, you got to check out this product called Potato.
And take my social media account, for example, Eric Tech Pro. I have grown this social media accounts from zero followers all the way to 7 ,000 followers just in a couple of months because I was able to post consistently and be able to also automate my entire process using this tool. So in this case, let me show you exactly how I was able to set this up.
Now, first of all, make sure to sign in with potatoes. Once you sign in, make sure to click on connect accounts and you can connect it with any social media accounts that you have. And once you've done that, make sure to click on settings and click on APIs.
And here you can be able to connect with your potato here with your large image model, because now you can be able to using large image model here to directly control your social medias. So in this case, I'm using ChatGPT. I'm going to click on that.
And it's going to tell you exactly the step -by -step guide, how you can be able to adding potato here as an MCB server instead of your ChatGPT. And once you have that set up, here you can see it asks, do you have access to Potato? And it says, yes, I do have access.
And it's telling you exactly which Instagram account you have access to. And then you can see I just paste the video links. For example, I have two videos that I posted on YouTube and I have AI here to kind of just cropped it with some clips.
And you can see that the clips here is shown here. And eventually I also ask it, hey, why don't you just go ahead and try to publish those reels, those shorts videos, and then also have one here to be scheduled for tomorrow's. And you can see that AI agent here is able to do that using the Potato MCPs.
And then furthermore, I also ask it to set up the auto comment DM rule. Or let's say someone actually comment any keywords, for example, real video or SDE video. Then behind the scene, we actually have an AI agent here that can actually respond to the users whenever someone comments for this kind of keywords instead of their comments.
So this way we can actually be able to grow our Instagram accounts to show engagements as well as be able to capture emails from people so that we can be able to add them into our email campaigns. So that's exactly how I was able to grow my Instagram accounts all the way to 7 ,000 followers using potato here, using this social media automation.
And if you're interested, make sure you check it out in the description below. So pretty much that's it. Let's get back to the real video.
Okay. So to get started, let's start with the secondary skills. And start off, we have the router skill, which is the potato mode.
So let's say if you have a situation you don't know which skill to trigger, you can trigger the potato mode here and explain what problem you have. And it's going to try to using the right skill here to tackle that problem. And then furthermore, you can see we have the setup skills, the catch up and plane talk skills.
the writing skills, as well as improving p -stack over time skills. And specifically, I really like the catch up and plain talk skills, especially for the recalls. So let's say if you switch different contexts, you were working on different projects and you want to get a recall of what happened so far, then you can see that there's some sets of skills I can use here to catch up on what happened so far, right?
So I have grouped them into different categories. So if you want to get into more details, you can also check out this page right here. Now, lastly, we also want to talk about here is the principal skills.
And you can see that there's tons. of principle skills inside of Pstack. Now, why are they skills, not system prompts?
You might be asking. It's because if you try to keep all the rules and principles into a system prompt, that's going to be a huge file, which contains a lot of tokens. Imagine that every message you send to AI is going to read that system prompt, but those information might not be relevant to what you're trying to do.
So that's why you can see it keeps all those principles into several skills and having an index map on exactly when to trigger them, right? Now because there's so many principles here, I'm just going to narrow down exactly what I find really helpful based on what we have in our superboard. Now for the first principle, keep it small.
This is really similar to the ponytail skill, which is already something that we have. So I'm not going to talk about that. But for the design for shape, one thing that I want to talk about here is this interesting principle called attack the premise.
Sometimes we have AI here to fixing a bug. It fixed it, still happens, fix it again, it still happens again. And what happened here is that it's going to continuously where we have AI fixing and then blocked, fixing and then blocked.
And it just feels like AI just not able to fix that bug entirely. But what this skill, like I said, it does here is to stop and have AI here to question about the assumption. Are we even on the right track before we have AI here to actually focusing on that direction?
Okay. So this is a really helpful skill and definitely I'm planning on adding this into our super war skill. Now, the next part that we have for the principle is the co -structure and safety and specifically for the type system disciplines.
Sometimes a variable here could have different types or different status. So it's really important to reject any impossible states, right? For example, a payment here could not be paid and also pending at the same time.
So problems like these, we want to make sure to have a principle that basically solve them, which is something that I'm actually going to add this to my super board as well. So pretty much those are the principles that I see that are really valuable. And I probably will take this into my own development workflow.
And of course, there's probably some other principles that I have already covered in CyberSuperboard. So if you're interested for this, you can check out this page right here. And I'll make sure to put this entire page, this entire diagram here for Pstack inside of the description below so you can actually access as well.
So pretty much that's it for this video. If you found it helpful, make sure to like this video. And of course, comment down below, what are some other spec driven development framework you actually review on the channel?
So pretty much that's it for this video. I'll see you in the next video.
The Hook
The bait, then the rug-pull.
Eric Tech opens by establishing why this skill pack is worth fifteen minutes: Lauren Tan is a React core team member working across SpaceX, Cursor, and Meta, and she just open-sourced the entire workflow, skills, and principles she uses to ship code with AI.
Named ideas worth stealing.
00:39model
Pstack's Six-Stage AI Software Lifecycle
- Plan
- Design
- Build
- Test
- Deploy
- Maintain
Eric's own organizing frame for Lauren Tan's skill pack. Each stage maps to one or more named Pstack skills.
Steal forOrganizing or comparing any multi-skill AI coding pack.
02:10model
Settle the Code Shape (design workflow)
- architect
- arena
- interrogate
Architect sketches types and a rough design, arena runs that design through multiple AI models in parallel, interrogate has a separate model judge and aggregate the results into one report.
Steal forAny high-stakes design decision where more than one model's opinion is worth having before building.
06:04model
Create-Verification-Skill (Launch, Doctor, Drive, Evidence)
- Launch
- Doctor
- Drive
- Evidence
A generated skill that reads the repo to learn how to launch, diagnose, drive, and capture proof that a feature works end-to-end, one feature per folder.
Steal forAny agent workflow that needs proof a feature actually works, not just a claim that it does.
How they asked for the click.
VERBAL ASK
09:54product
“If you have applications and your struggle with marketing, well, you got to check out this product called Potato... I have grown this social media accounts from zero followers all the way to 7,000 followers just in a couple of months.”
Mid-roll personal case-study pitch for sponsor Blotato, using his own Instagram growth numbers as social proof, then a direct walkthrough of connecting it as an MCP server to ChatGPT.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
