26:50Theo - t3․gg · Essay
The Unexpected Death of Codex
OpenAI folded its beloved Codex app into a rebranded ChatGPT overnight -- Theo argues they just killed the best brand in AI coding.
July 11thThe argument in one line.
The same AI model produces meaningfully different output quality depending on which coding harness runs it, because Codex's bloated, over-prescriptive system prompt actively degrades design and reasoning while Claude Code's bounded Workflows feature makes multi-agent orchestration cheaper and more reliable.
Read if. Skip if.
READ IF YOU ARE…
- You use AI coding agents like Claude Code or Codex regularly and want to know which harness gets more out of the same underlying model.
- You're curious why AI-generated UIs from different tools look so different even when the models behind them are comparable.
- You write or tune system prompts, skills, or agent instructions and want a real example of prompt bloat actively degrading output.
- You're deciding whether to run an OpenAI-subscription model through Claude Code instead of the Codex CLI or desktop app.
SKIP IF…
- You don't use any AI coding agent day to day — the comparison won't land without hands-on context.
- You're looking for a step-by-step Codex tutorial — most of this video is a critique of Codex's system prompt, not a guide to using it.
The full version, fast.
Theo has been running OpenAI's GPT-5.6-Sol inside Claude Code instead of Codex, and the results have been better across the board — especially UI design and sub-agent orchestration. Digging into why, he found Claude Code's system prompt says almost nothing about front-end design, while Codex's system prompt (until he complained) carried a bloated, hyper-prescriptive front-end design section that forced generic, 'utilitarian SaaS' outputs regardless of context, plus odd behavioral quirks like mandatory 30-second status updates. Independent model reviews rated the Codex prompt 3-4/10 for a modern general-purpose agent. The bigger win is Claude Code's Workflows feature: instead of letting a model spawn sub-agents indefinitely (Codex's open-ended Ultra mode), Workflows are model-authored code that defines stages up front and then terminates, cutting token usage roughly fourfold for comparable output. Theo argues that orchestration model — not the UI — is the real reason to route even OpenAI models through Claude Code.
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 – 01:52
01 · Cold open
Theo sets up the video's premise: he's been running GPT-5.6-Sol inside Claude Code instead of Codex, and the results have 'absolutely broken' him.
01:52 – 03:02
02 · Sponsor: CodeRabbit
Sponsor read for CodeRabbit's PR-review platform, which breaks large pull requests into labeled, prioritized layers instead of a flat file list.
03:02 – 04:46
03 · What's wrong with Codex's sub-agents
Theo explains Codex's two sub-agent systems — a stable v1 and an unfinished, off-by-default v2 with layered message-passing — and says the resulting workflow feels chaotic.
04:46 – 06:05
04 · Claude Code Workflows explained
Claude Code's Workflows let a model write a programmatic, top-to-bottom orchestration plan that terminates, which Theo says uses roughly a quarter of the tokens Codex's open-ended Ultra mode does for the same task.
06:05 – 07:26
05 · Design showdown
Theo shows two AI-generated pages and asks viewers to guess which model made which — the Claude Code / Sol output is visibly cleaner than the Codex / Sol output from the same model.
07:26 – 09:51
06 · Hunting for design guidance in Claude Code's prompt
Theo reads Claude Code's full system prompt looking for the front-end guidance that must be making designs better, and finds the phrase 'front end' appears only three times, with no real design instructions.
09:51 – 13:26
07 · The excised Codex front-end guidance
Theo pulls a cached copy of Codex's system prompt and reads its recently-removed front-end design section live, including hard rules on layout patterns, icon libraries, and card border radii.
13:26 – 18:42
08 · Reading the full Codex system prompt
Theo reads more of Codex's official system prompt, tallies how often it says 'card,' and has GPT-5.6-Sol itself grade the prompt as 3/10 for general use.
18:42 – 23:56
09 · Autonomy, persistence, and the verdict
Theo covers Codex's 'implement by default' autonomy rule and its 30-second status-update instruction, then contrasts it with Claude Code's shorter, calmer system prompt before delivering his verdict on Codex's prompt.
23:56 – 26:33
10 · Gotchas running Codex models in Claude Code
Theo lists the rough edges he's hit running OpenAI models through Claude Code: occasional formatting mistakes, an unfamiliar link format, and a filename collision between two agents that Sol correctly avoided overwriting.
26:33 – 29:05
11 · The workflows payoff, and why not OpenCode or Pi
Theo explains this whole investigation started because he wanted to see Workflows run well with GPT-5.6-Sol, and argues OpenCode and Pi don't have a comparable bounded orchestration primitive yet.
29:05 – 30:48
12 · Wrap-up
Theo mentions a CLI proxy that can route other subscriptions, including a Twitter Premium Grok plan, through Claude Code, then signs off.
Lines worth screenshotting.
- Running GPT-5.6-Sol inside Claude Code instead of Codex produces visibly better UI designs from the same underlying model.
- Claude Code's system prompt mentions 'front end' only three times and contains no explicit design guidance at all.
- Codex's system prompt, before it was quietly edited, devoted roughly a quarter of its length to a hyper-prescriptive front-end design section.
- That Codex design section forced SaaS, CRM, and operational tools to look 'quiet, utilitarian, and work-focused' regardless of what was actually being built.
- The Codex prompt specified exact UI implementation details, like 8-pixel card border radii and a single required icon library.
- The word 'cards' or 'card' appears in the Codex system prompt roughly 18 times — more than 'front end' and 'UI' appear in Claude Code's entire prompt.
- Codex's system prompt contains a literal instruction to provide user status updates 'every thirty seconds,' explaining a widely-noticed quirk.
- GPT-5.6-Sol itself, when asked to grade the Codex system prompt, rated it 3/10 as a portable prompt for modern models and only 7/10 for the narrow Codex runtime it was actually written for.
- Codex's 'autonomy and persistence' instruction tells the model to implement changes by default unless the user explicitly asks for a plan first, which explains why it tends to start building before a plan is agreed on.
- Claude Code's Workflows let a model write its own programmatic orchestration plan upfront, so the process actually terminates — unlike Codex's open-ended Ultra sub-agent mode, which can run indefinitely.
- Using Workflows instead of Codex's Ultra mode cut token usage to roughly a quarter for the same task, with equal or better output quality.
- Codex has two sub-agent systems: a stable v1 that spawns one level of sub-agents, and an unfinished, off-by-default v2 that lets sub-agents spawn their own named sub-agents with layered message passing.
- It took a YouTuber publicly reading the Codex system prompt on camera to get OpenAI to remove its most-criticized front-end design section.
- Other open coding harnesses like OpenCode lack a comparable Workflows feature, so they still require the user to hand-build sub-agent orchestration.
- A CLI proxy layer lets other subscriptions, including a Twitter Premium Grok plan, be routed through Claude Code the same way OpenAI's Codex subscription can.
Why the exact same AI model designs and orchestrates better in one harness than another.
PROMPT & HARNESS DESIGN
The model doing the work can be identical — what actually changes the output is how much a harness's system prompt constrains it, and whether its sub-agent orchestration is open-ended or bounded.
01Cold open
- The same underlying model can produce measurably different output quality depending on which coding harness runs it — the harness, not just the model, matters.
03What's wrong with Codex's sub-agents
- Codex's stable sub-agent system spawns one level of workers and waits for them to report back; its newer, off-by-default v2 lets sub-agents spawn further sub-agents with layered message-passing, adding complexity most users don't need.
- An orchestration system that can spawn indefinitely tends to keep working long after a bounded plan would have finished, burning far more tokens for similar output.
04Claude Code Workflows explained
- Defining an agent's multi-stage plan as literal code that runs top-to-bottom, rather than letting the model decide sub-agent spawns dynamically, makes the process actually terminate instead of running open-ended.
- Letting the model write its own orchestration code, rather than the harness hard-coding it, produced noticeably better results in this comparison — cutting token usage to roughly a quarter for comparable quality.
05Design showdown
- AI-generated UI quality is not purely a function of the model — the same model produced a visibly better design when run through a different harness with a different system prompt.
- When comparing AI-generated output quality across tools, check what's actually in the system prompt before assuming it's a model capability difference.
06Hunting for design guidance in Claude Code's prompt
- A system prompt with almost no explicit design instructions can still produce better design output than one with heavy prescriptive guidance — over-specifying constraints doesn't guarantee quality.
- Searching a system prompt for a specific keyword is a fast way to audit what guidance is actually driving an agent's behavior.
07The excised Codex front-end guidance
- Global instructions like 'always make this category of tool look a certain way' get applied literally and unconditionally by capable models, regardless of whether the constraint fits the actual project.
- Rules that specify implementation minutiae, like exact border radii or a single required icon library, remove a model's ability to adapt to context, and tend to make every output look the same.
08Reading the full Codex system prompt
- Asking a model to grade its own system prompt for portability is a cheap, useful self-audit.
- Spending a large fraction of a system prompt on one narrow scenario taxes every other type of task that never needed that guidance in the first place.
09Autonomy, persistence, and the verdict
- An instruction to implement changes by default unless a plan is explicitly requested explains why an agent sometimes starts building before you've agreed on an approach — worth checking for in any harness you use.
- Banning all visible in-app instructional text as a blanket rule breaks real product UX, because onboarding, validation, and accessibility hints are often essential, not decorative.
10Gotchas running Codex models in Claude Code
- Cross-model harness combinations can behave unpredictably in small ways, like formatting mistakes or unfamiliar output conventions, even when the underlying task completes correctly.
- When two parallel agents could produce the same output filename, a well-behaved agent should detect the collision and choose a new name rather than silently overwrite — worth checking for in any multi-agent setup.
11The workflows payoff, and why not OpenCode or Pi
- A feature isn't valuable just because it's technically simple to build yourself — having it built-in with good tooling and a system prompt that steers models to use it correctly is what makes the difference.
- Before adopting a new agent harness, check specifically whether it has a bounded, structured orchestration primitive, not just generic 'sub-agents' — that's the real differentiator here.
12Wrap-up
- Subscription-routing layers can let you run one provider's paid plan through a completely different coding harness's interface.
Terms worth knowing.
- Codex
- OpenAI's agentic coding CLI and desktop app that lets models like GPT-5.6-Sol read, write, and run code in a user's workspace.
- Claude Code
- Anthropic's terminal-based coding agent, which can also run non-Anthropic models like GPT-5.6-Sol through a proxy layer.
- GPT-5.6-Sol
- The OpenAI model variant discussed throughout the video as the shared model being compared across the Codex and Claude Code harnesses.
- Sub-agent
- A secondary AI agent spawned by a top-level agent to handle a specific piece of work and report back when it's done.
- Workflow (Claude Code)
- A model-authored code file that defines fixed stages, prompts, and sub-agents up front so a multi-step task runs to a defined end instead of looping open-endedly.
- Ultra mode
- Codex's mode for open-ended, potentially unbounded sub-agent spawning and orchestration, off by default because it is still being tuned.
- System prompt
- The hidden instructions prepended to every conversation with an AI agent that shape its default behavior, tone, and constraints.
- CLI proxy
- A routing layer that lets a coding harness like Claude Code run models or subscriptions from other providers, such as an OpenAI Codex plan or Grok, through its own interface.
Things they pointed at.
Lines you could clip.
09:40
“But can we take a moment to appreciate that it takes a fucking YouTuber to get slop removed from Codex? Why is this my job?”
05:20
“Ultra can just go forever, workflows won't.”
15:43
“I must issue an apology to my friends on Pi. I underestimated your game.”
28:46
“The workflows in Claude Code are the best implementation I have seen of orchestrating sub agents on real day to day work for people who are employed that I've ever seen by far.”
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.
analogystory
00:00If you had told me a few months ago I would have to make this video, I would have looked at you very, very strange. Because I never thought I would be praising Claude Code as much as I'm about to, especially in the context I'm about to praise it in. You see, I've been using five six a lot, but I haven't been using it in Codecs.
00:15I'm using it in ClotCode. I'm sure you have a lot of questions, as you probably should, because this is kind of dumb. It is so dumb that I tried it out during my early access window and was impressed.
00:27And now that I'm getting more and more frustrated with Codecs as a harness, a CLI, and more, I decided to go explore using Five Six Soul in Claude Code a bit more. And the results have absolutely broken me.
00:40It is not just because like Claude Code is a better UI, it kind of is better than the Codec CLI, but the real benefits are the things Claude Code does differently. From how it thinks about sub agents, to its system prompt, to its integrations, and more. I already did a video where I praised the good parts of Claude Code.
00:57I did not think I would be revisiting this topic so soon, but yeah. I have some issues with Codecs. This video is going to be a bit chaotic.
01:06I'm obviously going to show you how to set up Five Six Soul in Claude Code, but I also am gonna talk a lot about the why. Both why I'm so frustrated with Codex, and also what Claude Code is doing better to make me want to reach for it instead.
01:20And also a hint of why I'm not reaching for things like Pi as cool as they are. No shots fired at Pi at all, but I really need to focus in on Claude Code and Codecs because there is a new generation of capability with these models, and it is very, very important that we understand what these harnesses are doing right and wrong if we want to get the most out of them.
01:40But first, I wanna make sure you get the most out of today's sponsor. AI has gotten really good at writing code and it's even gotten decent at reviewing code. So why is it that code review has never been worse?
01:49I'm so tired of just going through these PRs. Like, look at this. Why is it an alphabetical list of files in my code base?
01:56This PR has a thousand lines changed. Are you seriously gonna tell me I have to sit here and go through every single one of them? Wouldn't it be nicer if this was broken up based on what's actually happening and not the order of the code in the code base?
02:08It'd be really cool if one of our code review agents could do this breaking up for us and give us a way better interface to review our code in, wouldn't it? Oh, it's open, isn't it? Surprise, CodeRabbit has a code review platform now, not just a GitHub plugin, but an actual full website that is way easier to read changes from.
02:24They break up the PR into different layers, which describe specific changes that the PR makes, making it way easier to review. Each of these sections is a summary and the details that are important called out, and the pieces inside of the layer are prioritized based on what you should be caring about.
02:39They even have fancy little guides on the side here that show labeled sections that might be worth checking out. This is particularly useful on gigantic PRs because they're basically impossible to use in the GitHub interface. Since this one breaks things up, it's way easier to find the parts that actually matter and give useful feedback to your team.
02:56Merge bigger PRs faster with more confidence at soydev.link/coderabbit. So with that, let's get started with what's wrong with Codex.
03:05There are layers to this one. I don't wanna crash out again about the rebranding of Codex as the ChatGPT desktop app or about the shit show of the CLI not getting updated as much as the desktop app. Generally speaking, I still think for most people, the best way to start using and taking advantage of more agentic engineering is the Codex desktop app on one of the Codex subscriptions.
03:28This hurts because obviously, I wanna be pushing people towards t three code, but the Codex app is a really good starting point, so much so that it inspired me to build my own open source alternative, but I have some issues right now. The biggest issue by far is how Codex thinks about sub agents.
03:47I just filmed a video about Ultra Mode that's already out. I would recommend checking that out if you want a deeper dive on this. The simplest TLDR I can give is that there's two versions of sub agents in Codex.
03:59The v one is stable and very simple. It lets the top level agent spawn a bunch of sub agents one level deeper with specific work and tasks to do, and then gets a response when they're done. V two, which is the new unfinished version that you have to manually enable, has the ability to copy the whole context window over by default, which is annoying, and also let the sub agents spin up their own named sub agents beneath them with layering and message passing between them.
04:26It takes a very simple hierarchy and makes it a bit absurdly complex and adds a ton of tools that are necessary for all of these things to communicate, and the result's a bit chaotic and clearly still being tuned. That's why it's off by default.
04:39Go check out my video I just did about Ultra if you want a deeper dive on how sub agents are currently working in Codecs. The thing I want you to know right now is that I don't like it, and I find that the sub agent work flows, specifically the workflow feature in Claude Code is much better. Because instead of having sub agents spawn randomly when the model decides to, the model will upfront define a programmatic workflow with different stages, different prompts, different sub agents at each stage that can spawn work for other stages, all in a single JavaScript file that is executed top to bottom because code is a great way to define something dynamic.
05:17Since workflows are code, they actually end, which I have found to be a huge win for token efficiency when using the five six models.
05:25They tend to just kinda go forever with Ultra, and I have seen way, way less utilization on my limits when I'm doing this through workflows instead, even for the same task. Ultra can just go forever, workflows won't. The output quality is the same if not better from workflows, and the token usage is like a fourth as much.
05:42So much more efficient way of doing things. OpenAI is shipping the code in Codex itself, and the model has to call the code correctly. ClaudeCode lets the model itself write the code for its workflow and for its sub agents, which I have found to be much better overall.
05:58I found it to be so much better that I wanted to try it myself, and that's what inspired me initially to start using Five Six inside of ClaudeCode, Not Fable in Claude code calling five six over the Codec CLI, but with five six as the actual model doing the work in Claude code itself.
06:14Since then, I've been using it a ton. And the more I use it, the more impressed I get. I have a question for you guys.
06:20Which model do you think designed this page? When I look at this, it screams Claude to me. This obviously looks like something that Claude made, right?
06:31Like Fable or Opus, clearly. As someone who's looked at a lot of pages that were generated with Codex, it could be something GLM or otherwise distilled off of Claude, but this screams Claude to me.
06:44For reference, here's a similar page I did with Codex in 05/06 Soul. I don't know about you guys, but the difference in quality here is pretty aggressive to me. I think both are slop to be clear.
06:57This one hurts me a lot less than the one that Codex did. It shouldn't be too surprising to you based on the topic of this video that this page was made with Sol inside of Claude Code. So why the hell is Five Six Sol making designs this much better in Claude Code?
07:15You probably are gonna have the same assumption I had, which is that they have some things in the system prompt for Claude Code that make it more likely to do good designs. So I investigated. I read the full Claude Code system prompt, trying to find this front end guidance so that I could bring it over to Codex and maybe have Codex design better.
07:35So I hunted, and the word front end appears three times. First, it appears with the instructions that for UI and front end changes, start the dev server and use the features in a browser before reporting the task as complete.
07:48This is telling the model to confirm its work. Cool. Makes sense.
07:53Not really front end guidance. Don't worry though, the word front end appears two more times on the page. Oh, they're both here.
08:00As examples of memory. Strange. Maybe UI appears in here?
08:07Oh, yeah, it does twice in the thing we were just at. For UI or front end changes. Does that mean the system prompt has nothing about design?
08:17Why is it designing things so much better then? Something has to be up. I also remember that OpenAI has been bragging about Five Six being way better at design, even though I never saw that.
08:27But when I was testing Five Six, almost all of the work I did in it was done in Codex. Turns out I was checking the wrong place.
08:36If you were waiting for the crash out, it begins now. This is the official Codex system prompt.
08:43You are Codex, an agent based on GPT five. You and the user share one workspace, and your job is to collaborate with them until their goal is genuinely handled. You can tell from the first sentence what you're in for here, can't you?
08:57Oh, boy. I don't wanna spend this whole video reading this to you guys and explaining how horrible it is. Okay.
09:04That's a lie. I do kind of wanna do that. It is very tempting.
09:06Because I did read this whole thing. And however bad you think it is, it is comically worse. I think they might have listened to me and removed the part I wanna bitch about the most, so I'm gonna go find a cached version so I could do that there.
09:19Thank you to Felipe for uploading this for me. It is very helpful to have. Here is the official front end guidance that was in Codex up until recently, I I think at least.
09:29Let me double check quick. Was an easy way to do this. Yes, they changed it.
09:33They listened to me. I crashed out at my friend who works on front end stuff at Codex to get this fixed, and they did. But can we take a moment to appreciate that it takes a fucking YouTuber to get slop removed from Codex?
09:47Why is this my job? Let me show you what I made them remove. This is the previously present official front end guidance.
09:56You follow these instructions when building applications with a front end experience. First off, build with empathy. If working with an existing design or given a design framework in context, you pay careful attention to existing conventions and ensure that what you build is consistent with the frameworks used and design of the existing application.
10:14It gets so much worse from here. I see people in chat already freaking the fuck out. You have no idea how bad this gets.
10:21It's probably the worst counter psychosis I've ever had was reading this. This feels like it was written for GPT four one. You think deeply about the audience of what you are building, and use that to decide what features to build, and when designing layouts, components, visual styles, on screen text, and interaction patterns.
10:39Using your application should feel rich and sophisticated. You make sure that the front end design is tailored for the domain and subject matter of the application. For example, SaaS, CRM, and other operational tools should feel quiet, utilitarian, and work focused, rather than illustrative or editorial.
10:57Remember that thing that OpenAI models do, where when you tell them to do a thing, they do the thing? This just told the model, if you're doing anything that looks like a Sass, that it needs to look utilitarian, quiet, and work focused.
11:11Always. And it does, always. But this isn't even the bad section of the front end guidance.
11:17You make sure to use icons and buttons for tools, swatches for color, segmented controls for models, toggle slash check boxes for binary settings, sliders slash steppers slash inputs for numeric values, menus for option sets, tabs for views, and text or icon plus text buttons only for clear commands, parentheses, unless otherwise specified.
11:40Cards are kept at eight pixel border radius or less unless the existing design system requires otherwise. You do not use rounded rectangular UI elements with text inside if you could use a familiar symbol or icon instead. You build tooltips with names and describe unfamiliar icons when the user hovers over it.
11:57Cool, fine. Point three, you use Lucid icons inside buttons whenever one exists instead of manually drawn SVG icons. If there's a library, use that icon library instead.
12:07Whatever. The fact that it's prescribing specific libraries for icons and specific border radiuses for cards is absurd.
12:17But we're just getting started here. You build feature complete controls, states, and views that a target user would naturally expect for the application.
12:26You do not use visible in app text to describe the application's features, functionality, keyboard shortcuts, styling, visual elements, or how to use the application. You should not make a landing page unless absolutely required.
12:40When asked for a site app game or tool, build the actual usable experience as the first screen, not marketing or explanatory content. When making a hero page, you use a relevant image, a generated bitmap image, or an immersive full bleed interactive scene as the background with text over it that is not in a card. Notice how all of the pages that Codex designs are the same?
13:00It's because it tells it exactly how to make it look here. It tells it how to structure the homepage to make sure the h one tag is the brand, product, place, or person name, or a literal offer slash category. Chad is begging me to stop reading because this is hurting them so much.
13:18This is an absolute shit show. And sadly, Never Drac, this isn't a skill. This isn't a thing that you can add to Codex.
13:26This is the official Codex system prompt directly. Up until yesterday, every single time you sent a prompt to any OpenAI model in Codex, this all came along with it.
13:38You have burned hundreds of thousands, if not millions of tokens on this slop.
13:45And it took me, probably the first person to ever actually read this, to get it removed. I'm not trying to talk myself up here.
13:53In fact, I'm trying to do the opposite. I don't get why it's my job as a fucking YouTuber to be the one to get this fixed.
14:01And I'm at the point now where I am working on building a whole new Codec system prompt from scratch by hand. No AI generated anything in it. Prompts should be by hand, especially global ones and skills and things like that.
14:13My system prompt will be written by hand by the end of this coming week. And if it works out well, I will share it with all of you and harass OpenAI to copy my changes. But there is so much more in here.
14:24I don't wanna just sit here and complain because I God. You do not put UI cards inside of other cards.
14:32Do not style up page sections as floating cards. Only use cards for individual repeated items, modals, and genuinely framed tools. The word cards is on this page six times.
14:43The word card is 12. I feel sick. I'm gonna ask you guys a question to the power users of Codex.
14:51Have you ever noticed that when you set it off to do some work, like babysitting a PR, or keeping track of sub agents, or just doing something long and difficult, that it loves to set timers.
15:05Specifically, it really loves to set thirty second timers. Have you guys seen this weird fixation on thirty second time limits in codecs?
15:16Because I did, and I was suspicious of this. Would you like to know where that came from? You provide user updates frequently, every thirty seconds.
15:27I cannot say enough bad things about this system prompt. It is enough of a reason to never trust Codex again. The fact that this has been shipping for as long as it has, and is burning our tokens, wasting our time, and making our outputs worse, is hilarious.
15:43And with this, I must issue an apology to my friends on Pi. I underestimated your game. To be fair, this is your fault too, because you never told us how bad the codec system prompt was.
15:54Because the problem with codecs isn't even implementation. A lot of it is just this. It is so bad.
16:00Don't worry though, it mentions goblins twice. The codec system prompt mentions goblins more than the Claude code one mentions front end in UI, and it mentions cards six times more than the Claude code prompt mentions front end at all.
16:16My chat is crashing out so hard over this right now. What in the fuck? This has been fucking with it interacting with Claude two.
16:23Please, God, stop this. Yep. You've all noticed this.
16:27And I'm the first person to figure out why somehow. I'm sure others have, and they just don't have a platform enough that they can be loud about it online. But like, what the fuck?
16:37I did mention before that you shouldn't be using agents to write your system prompts, your skills, and all of those things. Like, just do it by hand. They will come out way better, and you can adjust and learn as you go.
16:47That said, even most models should be able to see the problems with this system prompt. So I asked Five Six Sol in Claude Code, as well as Fable five in Claude Code, and of course, Five Six Sol in Codex, just to double check and be sure, to describe all of the things that are wrong with this system prompt and help me rate it.
17:06Five Six in Claude Code said the following. Seven out of 10 is a prompt for one tightly controlled codex runtime, four out of 10 is a general coding agent prompt, and three out of 10 is a portable prompt for modern models like Five six Soul. Obviously, it is coupled to codex, it's built for that, that's fine.
17:23It calls out that modern models generally understand tool schemas well without the main behavioral prompt repeatedly describing implementation details. These runtime instructions should be generated by the harness not embedded into the reusable agent prompt. Yep, sure, whatever.
17:35But it re describes tool calls multiple times throughout. The front end section is far too prescriptive. It's effectively a separate front end design constitution.
17:43It consumes about a quarter of the prompt even when the task concerns a CLI, database, back end, or a library. Cards must have a eight pixel radius, or letter spacing must always be zero. All of these things that I went over before.
17:58A capable model will obey these, but this doesn't mean the results improve. It may ignore an existing design system user request or domain convention to satisfy arbitrary global constraints. And then it has a bunch of actively questionable UX guidance, like it's weird callouts for icon only controls, especially for touch devices.
18:15Because touch devices don't have tooltips, remember. The ban on visible instructional text is more concerning. Empty states, onboarding, validation guides, accessibility hints, and contextual help are often essential product UI.
18:25Yep. Have you ever wondered why Codex never does good empty states? It's the system prompt.
18:29They also have a bunch of weird stuff in there like continue until solved that is probably part of why five six doesn't stop when it should sometimes. You ever wondered why Codex loves to just start building when you're in the early stages of planning? It's right here, autonomy and persistence.
18:45Unless the user explicitly asks for a plan, asks a question about the code, is brainstorming possible approaches, or otherwise makes it clear that they do not want code changes yet, you assume they want you to make the change or run the tools needed to solve the problem. In those cases, do not stop at a proposal, implement the fix.
19:02You hit a blocker, try to work it through yourself before handing the problem back. I am very thankful OpenAI lets you use your Codex sub other places right now. I love that I'm seeing in chat as I do this, people saying, oh my god, that's infuriating.
19:16I've been fucked by that exact prompt. I forgot it includes callouts about limiting dominant purple and purple blue gradients, beige cream sand and tan, dark blue and slate, and brown slash orange slash espresso palettes. Scan CSS colors before, yeah, you get the idea.
19:31They tried to make 5.5 better at design by putting a bunch of rules in and made it worse. They tried to make 5.5 more persistent by putting a bunch of rules in and made it worse or different. And all of these things were still applied on 5.6 where they should not have been.
19:44Apparently, the changes they made for the front end design stuff is only for 5.6, it's still applied for 5.5, even better. We're unlikely to change the current system prompt for 5.5 because it was used during training, but I'll pass your feedback along to the team.
20:01Even five six Soul in Codex had a lot of negative things to say. Specifically, they call it four out of 10 prompt implementation, and tore apart a bunch of things that are wrong with it.
20:11I think there's even more than that. The point I'm trying to make here is that the system prompt feels more thrown together than crafted. And this is what I've shared on Claude Code and the insane length of their system prompt and whatnot.
20:23It's a lot better here. I think they're actually writing a decent bit of it. You're an interactive agent that helps users with software engineering tasks.
20:30Use the instructions below and the tools available to you to assist the user. They added some inserts here about what to not do with, like, hacking and whatnot. It starts with descriptions of what the user will see.
20:39It says all text you output outside of tool use is displayed to the user. Output text to communicate with the user. Cool.
20:45That is fine. Tools are executed in a user specific permission mode. Yep.
20:48Makes sense. Doing tasks. The user will primarily request you to perform software engineering tasks.
20:53This may include solving bugs, adding new functionality, yada yada. They're highly capable and often allow users to complete ambitious tasks that would otherwise be too complex or take too long. You defer to user judgment about whether a task is too large to attempt.
21:05For exploratory questions, respond in two to three sentences with recommendations and the main trade offs. Present it as something the user can redirect and not a decided plan. Don't implement until the user agrees.
21:14This is the opposite of the Codex one and it's actually quite nice to have the model ask some questions. Prefer editing existing files to creating new ones. That's nice.
21:22It says be careful to not introduce security problems. Friars need to be inserted anymore, but not the worst thing. Default to writing no comments.
21:29Only add one when the why is non obvious. A hidden constraint, a subtle invariant, a workaround for a bug, behavior that would surprise a reader. If removing the comment wouldn't confuse a future reader, don't write it.
21:39These are reasonable guidelines, and all of these are simple and pretty general. Executing actions with care. Carefully consider the reversibility and blast radius of actions.
21:48Generally, you could freely take local reversible actions like editing files or running tests. But for actions that are hard to reverse, affect shared systems beyond your local environment, or could otherwise be risky or disruptive, check with the user before proceeding. Just like the way this is written, the tone of it, the actual details of course as well, it's not that bad.
22:09It's actually decent. It gives a list of examples that are things it should verify before doing, like destructive operations, hard to reverse operations, actions visible to others.
22:19Meanwhile, I just saw poor Matt Schumer get his entire system, like, user folder deleted by Soul running in Ultra in Codex.
22:28So maybe Codex needs these warnings now, too. There's a simple tone and style section.
22:35Says only use emojis if the user requests it. Your responses should be short and concise. When referencing specific functions or pieces, use the pattern file path colon line number.
22:44Don't use a colon before tool calls. Your tool calls may not be shown directly in the output. Has things to not save.
22:50Memory is complex, I'm happy that they gave it enough instructions here to do it right. It gives info on the environment it is in. It gives the tools.
22:57It has instructions on how to create a pull request, which are a little more specific than I like and make it in formats I don't love. This is one of the few things I would actually change here. Their edit tool, I still hate how Claude code does edits, but that's a rant for another time.
23:09Schedule wake up is cool. It doesn't specify to wait thirty seconds. It gives it info on how to pick a good amount of delay, which is useful.
23:18It then gives a description of skills, tool search, and that's it.
23:22It's not that long. I read through the whole thing, and it wasn't too bad. Meanwhile, the more I read the codec system prompt, the more I questioned my whole life up until this point.
23:33Yeah. That's enough of the codecs crash out for now. I think a cleaning of houses in order on that side, though.
23:41This system prompt is so bad that it should just be deleted and rewritten from scratch, and they need to make some really, really intense benchmarks to measure success and failure, as well as auditing the histories to see how cringe its interactions are. This is bad beyond what is acceptable.
23:58So you have stylus what's wrong with Codex, and a good bit of what's better in Cloud Code. The system prompt is better, workflows are fucking awesome. It's UX in the terminal is better overall.
24:07The CLI for codecs is rougher than it should be right now. One last thing before we do the setup though, I wanna talk about the gotchas that I've been experiencing using codecs models inside of Cloud Code.
24:21Admittedly, there have been fewer than I expected. First off, I've noticed it sometimes loses track of what it's doing a little bit more.
24:29This could be changes in how context is managed, but I honestly think this is just something small in the system prompt. Here is an example where I had it making changes alongside two other agents. All three are making HTML files for me to go review the feedback it was giving, and two of them happened to come up with the same name for the HTML file.
24:49One was using Fable, and it wrote fine and made the URL fine. This one was using five six sole, and it tried to write the file.
24:59It already existed, so it said it would leave that file untouched and publish a new report with a distinct file name. So it did its own different file instead.
25:07Of note though, is that I told it to use my HTML plan skill. And my HTML plan skill explicitly tells the model to use the web host through the CLI I gave it to host this model.
25:20This skill is meant to give the model the ability to upload the HTML with a link that I can click and look at. It's one my favorite things I've built, my silly little post plan site. And it even says here, I'll keep it standalone and responsive, then publish through post plan as requested.
25:35It failed the write. It noticed the file already existed. It made the new file.
25:39Oh, wait. I was wrong. I thought it didn't host it.
25:42This is a link. It's just not familiar with the formatting here. I'm used to the model spinning out an actual link.
25:48This is somehow a clickable link, even though it doesn't look like that here at all. I was wrong. I thought it just didn't do that step.
25:54It did. I asked it, and here it gave me the actual link in text that makes more sense. I was wrong about that.
26:00Mostly it's good about these things. Like it does stay on task largely fine. It just doesn't know how to use the formatting great.
26:07And I have noticed this in other places too, like here when I did the number based formatting, it screwed it up a little bit. Like here, it's number one.
26:15It's heavily coupled to one specific runtime. Has a chunk, and then one again. Modern models generally understand yada yada.
26:22Then it has two twice, then it has three twice, and then four twice. I think this is something weird about how markdown formatting works in Claude code, but Five Six isn't doing it right, clearly.
26:33Remember, the reason I wanted to do this initially was for UltraCode and specifically for Workflows, because I wanted to see how good Five Six could do Workflows. And it does them great. There was a problem though.
26:45It is an append to the system prompt that says, hey, please please please use workflows for this task. This works great with Claude models. You can tell it to use them for your workflows and orchestration.
26:56For example, here I told it that I wanna better understand this code base, it's the Codex code base, and I want it to analyze with the workflow. Use five six Sol, five six Terra, and Fable five all on high reasoning levels to analyze the code base. I want it to have the info when I ask it a vague thing like this, so that it will orchestrate properly.
27:14So I'll probably put it in my system prompt in some way, or my global agents MD. There's a lot of places you can put it, but as long as you give Claude code the ability to know what you mean by Solterra and other things, or you just specify them correctly here, you're good to go. There's a couple other small catches, like for example here, the five six models don't report how many tokens they're using until the work is done.
27:36Fable will give you live updates about how many tokens it's using. It's a small thing, just one of the few I noticed. Overall, I've actually been really impressed with this.
27:47And I already see the comments coming in, so I'm going to address them now. Why not OpenCode or Py or OhMyPy or whatever harness is popular today?
27:57The reason is because they don't solve the problem I care about. Sure, their system prompts are less bad than Codex is, but they don't solve the sub agent orchestration problem, which is the thing I am most excited about here. Oh My Pie, absolute fucking slop.
28:12It froze my terminal for two minutes when I installed it, because they thought it was a good idea to post their entire 150 page change log in formatted markdown in my terminal, destroying my entire scroll back and destroying it entirely.
28:26It was real bad. Py is great, but you're gonna have to build all of these workflows and orchestration tools yourself. OpenCode has a bunch of hard coded sub agents that aren't very good.
28:35The v two of OpenCode might be better. I have not explored it at all yet. But none of those have workflows.
28:41And to be real, and to be very, very clear, because I started here, I did this for workflows. The workflows in Claude Code are the best implementation I have seen of orchestrating sub agents on real day to day work for people who are employed that I've ever seen by far. If you want to have an agent break up other agents with other models to do complex work, to take turns and have phases throughout, to kick off things that matter and ignore things that don't, and have an actual end.
29:11Workflows are the best solution I have found for this by far, and nothing else comes close. To be very clear, I don't think workflows are this super complex feature that no one else can implement. In fact, Codex told me to just add it to Codex myself when I asked it about that versus using it in ClaudeCode.
29:26You can absolutely do this. That all said, Claude Code has this figured out. They have a good system for it built in.
29:34They have good UI built for it too in the terminal at least, to the desktop app less so. They have a system prompt that properly steers models to do it right, and they have a harness that is good enough. One last cool thing that I'm sure you probably could have guessed, CLI proxy API supports things other than Claude code and codecs and subs from those platforms.
29:53So if you wanna bring in subscriptions from other places, like a cough cough Grok Build, because you have an existing Twitter premium account, you can here. It's actually a pretty good value. You get a slightly absurd amount of usage out of the $20 a month Twitter plan, if you use it with Grok Build.
30:10And if you don't wanna use Grok Build and you wanna use it in other things, you can now get Opus four five in Claude Code by just adding it here. I had a lot of fun with this one. And again, to be clear, I don't think everyone should go exclusively use their Codec Sub through Claude Code.
30:25I just thought it was a fun experiment with results that were much better than I expected. And now I'm just curious if you guys are interested enough to do the same yourself. Am I insane for going this deep, or is this actually kind of fun and interesting?
30:37You gonna set it up yourself, or maybe you're just gonna point your agent at this video and ask it to do it for you? Let me know how all of that goes. I'm genuinely curious.
30:43Well, keep an eye on the comment section, but until next time, peace nerds.
The Hook
The bait, then the rug-pull.
Theo opens by admitting he never expected to be praising Claude Code this much — not because it's a nicer terminal UI, but because running OpenAI's own GPT-5.6-Sol model through it produces better designs and cheaper orchestration than running that same model through OpenAI's own Codex.
Named ideas worth stealing.
03:30concept
Codex sub-agents v1 vs v2
- v1: stable, one level of sub-agents, waits for results
- v2: unfinished, off by default, sub-agents spawn their own named sub-agents with layered message passing
Codex's two generations of sub-agent orchestration, contrasted by Theo to explain why Codex's multi-agent behavior feels chaotic.
Steal forauditing any harness's sub-agent model before trusting it with unattended multi-agent work
05:00model
Claude Code Workflows
A model-authored JavaScript file that defines stages, prompts, and sub-agents up front so a multi-agent task runs top-to-bottom and terminates, instead of the model deciding to spawn sub-agents dynamically forever.
Steal forany custom agent pipeline that currently lets a model spawn sub-agents with no defined end condition
How they asked for the click.
VERBAL ASK
01:52product
“Merge bigger PRs faster with more confidence at soydev.link/coderabbit.”
Woven as a produced demo inside the cold open rather than a hard interruption — shows CodeRabbit's actual PR-review interface and specific features before delivering the link.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
