§ 01 · The Hook
The bait, then the rug-pull.
Every developer has a tribe. Team Claude, team Gemini, team Codex — a social media full of loyalty pledges to models that will be superseded in six months. Mark Kashef's answer is a Claude Code skill called Council: a configuration-driven layer that keeps Claude as your daily driver while pulling opinions from any model on OpenRouter, then synthesizes and executes the best path.
Where the time goes.
00:00 – 00:29
01 · The tribal AI debate
Frames the false choice of model loyalty. Sets up the core question: why pick a side when you can use all of them?
00:30 – 01:04
02 · Introducing Council
Announces the skill concept. Animated whiteboard slides introduce the idea of Claude as hub, OpenRouter as router.
01:05 – 01:28
03 · Claude Code + OpenRouter explained
Brief visit to openrouter.ai. Positions it as the unified interface for all LLMs.
01:29 – 02:16
04 · What the council actually does
Animated flowchart: user prompt routes to Council skill, auto-detects category, routes to Codex/Gemini/Claude, returns 3 perspectives.
02:17 – 03:03
05 · Full skill file walkthrough
Sublime Text view of the SKILL.md: name, description, trigger phrases, invocation modes.
03:04 – 03:20
06 · Setup: OpenRouter API key
One setup step: OPENROUTER_API_KEY in .env.
03:21 – 04:31
07 · The config file
council_config.json: providers, defaults per category (bug_fix to Codex, frontend to Gemini 3.1 Pro, quick_check to Gemini Flash), categories with keyword triggers.
04:32 – 05:09
08 · Context packaging
What to send to external models: problem statement, relevant code snippets, what has been tried, what kind of answer is needed. Never send the full conversation history.
05:10 – 05:48
09 · Synthesis rules
After receiving responses: label by model, state agreement/disagreement with reasoning, execute the best approach. Claude remains the executor.
05:49 – 06:22
10 · Two terminals, same prompt
Parallel test: same Claude Code bootcamp landing page prompt in two terminals. One stays as control, one gets the council treatment.
06:23 – 07:07
11 · Baseline websites
Both sites look similar: generic dark-terminal aesthetic, similar typography. Sets the before-state.
07:08 – 07:57
12 · Running /council on the landing page
Invokes the slash command. Council plan: Gemini 3.1 Pro reviews frontend, Codex reviews copy. Runs both in parallel.
07:58 – 08:32
13 · Council results: frontend + copy audit
Gemini flags full-monospace body as UX anti-pattern, grid overflow on mobile, jarring FAQ animation. Codex scores copy 7/10, flags over-promises.
08:33 – 09:10
14 · Claude synthesis and implementation
Claude agrees on high-priority items, splits typography, fixes grid, adds focus states, tones down copy. Implements all changes.
09:11 – 09:35
15 · Improved bootcamp site
Result: cleaner layout, better pricing section, tighter copy. From one-off prompts to production workflows.
09:36 – 10:06
16 · Demo 2: VC analytics dashboard
Same pattern applied to a venture capital analytics dashboard with mock data. Baseline result is rough.
10:07 – 11:15
17 · Council audit on the dashboard
Gemini frontend audit + Codex copy audit. Claude synthesis includes a notable pushback: shorter dashboard labels preferred over verbose VC terminology.
11:16 – 12:14
18 · Bloomberg Terminal result and broader use cases
Final dashboard looks like a Bloomberg terminal. Skills can be stacked. Same-model comparisons (Haiku vs Sonnet vs Opus) mentioned.
12:15 – 12:22
19 · CTA
Free skill download link and Early AI Dopters community link in description.
Visual structure at a glance.
Named ideas worth stealing.
03:21
model
Council Skill Config
- providers (id, name, flagship model)
- defaults per category (bug_fix, frontend, architecture, refactor, general, quick_check)
- categories with keyword triggers
- context packaging rules
- synthesis rules
A JSON config that maps task categories to the best available model and defines how Claude should package context and synthesize responses.
Steal for Any multi-model workflow where you want automatic routing without manual model selection
02:17
list
Three Invocation Modes
- Auto-routed: say get a second opinion and the skill classifies the task
- Named model: explicitly ask Gemini or Codex by name
- Slash command: /council with a typed prompt
Three ways to trigger the Council skill, from fully automatic to explicit.
Steal for Any Claude Code skill that should work both as a slash command and as a natural-language trigger
Lines you could clip.
00:51
"Instead of bickering about what model is the best today, you can just focus on what works."
04:37
"You are having these office hours with different language models that are aware of what you have tried, what you wanna do, and where you might be stuck."
05:15
"Why not have the predictions of one improve and ameliorate the predictions of the other?"
05:47
"I want Claude to always pick the path of least resistance even if that path was not carved by itself."
10:49
"Codex suggestion to rename markups to unrealized value uplifts is technically correct, but for a dashboard, the shorter labels are fine."
Things they pointed at.
How they asked for the click.
12:15
link
"I will attach the skill that I showed you in this video in the second link in the description below."
Low-friction: free download. Community upsell (Early AI Dopters) mentioned separately as the first link. No price stated.
Word for word.
HOOK opening / re-engagementCTA the pitch metaphor
00:00HOOKGo on any social media site right now, and you'll find different tribes rooting for different language models. You have team Claude, team Codex, and team Gemini, as well as team every other language model on the planet.
00:11HOOKEveryone's got a laundry list of why their model choice is the absolute best compared to others. But instead of being pushed to pick a side, why not just use all of them? What if you could keep Cloud Code as your daily driver,
00:24HOOKbut tap into the intelligence of Gemini and Codex and whatever is the latest and greatest whenever you need it. I built a skill around this, and in this video, I'm gonna show you how you can do it too. It's a mega skill that lets you tap on the shoulders of giants and leverage any model you want depending on where you think these other models fit best.
00:42HOOKLet's dive in. So the general concept is very simple. Instead of bickering about what model is the best today, whether it's 5.3
00:49HOOKcodex, 5.9, 6.1 opus, whatever's gonna be the next set of models, you can just focus on what works. And the way I've constructed the skill that I will give you by the end of this video is basically combining Claude code with OpenRouter.
01:02HOOKIf you're not familiar with what OpenRouter is, it is essentially a service that lets you access all kinds of language models. So even though I'm primarily harping on Gemini and Codex as the alternatives, you can choose whatever you want.
01:14HOOKIf you think that Quen models are amazing at coding, you can leverage those, and you can now create a pseudo counsel for your Claude code language model. So the way I'll put this counsel in action is I'll ask Claude code to do a task that I would typically do just with no language models whatsoever. And then I'll show you the power of being able to say, let's get a second opinion on this website, this app, this whatever that you can bring in a third party source like Gemini to take a look at your front end and be like, step aside.
01:42I got this. Here's my advice on what you should do. And then Claude Code can take said advice and implement it.
01:48Same thing with codex. I actually use codex quite a bit for bug fixing. Once in a while, Opus will get stuck on a bug, and instead of constantly spinning, I'll bring in something like codex to take a look at the entire code base and give me its best opinion on what's going wrong.
02:01Now you might have your own opinions on which models are best for what. I totally respect them. Everyone has different use cases and different applications.
02:08So without further ado, let's take a look at this skill. And what I've called it is just counsel. You can name it whatever you want, and I created it so it's a slash command as well.
02:17I can do slash counsel, and it will know that I'm trying to invoke the third party powers of a different language model or multiple for a particular problem. The description is get second opinions from competing AI models, OpenAI, Google. Again, you can add whatever you want here via open router without leaving clawed code.
02:33Important. This skill takes priority over another skill that I have in my system, so you can ignore that part. Trigger phrases would be get a second opinion, ask Gemini, ask GPT, ask Codex, consult.
02:44It just gives it a series of different ways to be triggered and invoked that are semantically similar. And the last thing I say here is never call Gemini or OpenAI's APIs without my direct intention to do so. In terms of the rest of the scale, it's pretty straightforward.
02:57We have the overall setup where you'd have to just grab an API key from open router. You would throw it into the dot n file, and then beyond that, you'd have three invocation methods. You already saw a few in the description.
03:10This goes a little bit deeper into others. And this config file is actually the most important part of this skill, and this is the part where you will wanna chime in as the user. And I'll grab it over so we can take a look, and we'll zoom in so you could see exactly how it's structured.
03:24So if you are nontechnical, there's nothing you actually have to worry about from a coding perspective. You just have to verbally tell Claude code to change this file whatever way you want.
03:33So in this case, I walk through the providers I really care about, and then it figured out from the open router API how to call those models. So when we look at defaults, then you can say that when there's a bug fix that Opus, for whatever reason, can't fix will bring in the powers of 5.3 codecs.
03:49And when it comes to front end, we can say to use the powers of Gemini 3.1 pro preview or whatever latest model that is out by the time you're watching this. And you can do the same thing for architecture,
04:01refactoring, anything in general, or a quick check. So a quick check on front end could use Gemini three flash, which is still a very capable and pretty cheap model.
04:10And then as we go down, you have different categories where we go deeper as to what to look for. Let's say the words bug fixes, error diagnosis, etcetera, and when to invoke different things like the bug fix, like the front end, and like the architecture language model. And just to make sure we're sending the right information to the right model at the right time, we have this section that's called context packaging.
04:30So I said when sending context to another model, include, one, the specific question or problem statement, two, relevant code snippets. This is probably the most important because if you want a proper consult, you should provide snippets of the existing code to see the opinions of other language models.
04:46And then this one is really important. What other approaches have been tried? So in a way, you're having these that are aware of what you've tried, what you wanna do, and where you might be stuck, and then they can give their holistic opinion in that way.
05:00And this is the last part I'll show that brings everything together, which are these synthesis rules. So after receiving a response from the council, present the raw response clearly labeled with the model name, state agreement or disagreement. So I want Claude Coe to chime in and be like, I
05:15agree somewhat with Gemini's take or Codex's take, but it's made me think of another solution. That's the goal. These are language models.
05:22So why not have the predictions of one improve and ameliorate the predictions of the other? The last element here is the execute best approach. So end of the day, you're gonna have different language models have egos about which one is right.
05:33But, ideally, I want Claude to always pick the path of least resistance even if that path wasn't carved by itself. Now that's a run through of the skill. Let's actually put it to the test.
05:43So if we open a terminal and I say something in one terminal like, can you spin up a landing page for a new Claude code boot camp? And it's a boot camp for someone that is more on the intermediate end.
05:56Create a full landing page, a whole program, come up with pricing, and spin up the resulting web page on localhost. So we will send this in two different terminals. Naturally, with language models, you'll have two slightly different results.
06:10With one, we'll keep it as is. With the other, I'll show you how just bringing in different opinions from different language models can change the entire equation. After a few minutes, we got two different websites from the two different terminals.
06:23Now the first one looks like this. Nothing too crazy. Looks like a standard websites.
06:30Could be a little bit easier on the eyes from a font perspective. And the second one is the exact same command. I just asked it to create a brand new folder.
06:39So this one looks like this. So stop prompting, start shipping, little bit better hero statement, but the rest of this looks pretty similar.
06:48You can see the fonts are very similar. So what can we do to improve one versus the other? So this first tab, the website no cancel, will act as our control.
06:57We're not gonna change this at all. Whereas we're gonna make only one set of changes using different language models just to show you how quickly you can improve things. So if we go to website council
07:08and we go to the bottom, I'll use my slash command slash council, and then this should now pose a question back to me. I'm already going to preemptively draft it and say the Okay.
07:20I want you to take a look at the web page that was put together, and I want you to audit it from both a front end perspective as well as a copy perspective. I want you to send snippets of the front end code to Gemini 3.1 Pro, and I want to send the copywriting
07:36to Codex to get their opinions on what are the best practices if we should keep things the same or improve them and if we should improve them in what way. So a bit of a mouthful, and as I inhale oxygen, I will send this over and show you the results. And after four minutes, the skill comes back with a proposal for a front end as well as a copywriting set of changes.
07:58On the front end, it says what's working. So at least the language model is understanding that things don't need to be fully rewritten from scratch, but it does flag some issues.
08:08So this goes into the spacing, the grids, the fact that mobile friendly is not optimized, and then it proposes some changes.
08:16Then the same thing happens with codex on the copywriting side. It then flags. It's basically like having a LLM as a judge, except you get to tap into an LLM as a judge from a different language model that might have different pros and cons.
08:29And then this is the part that we should really care about, which is Claude code synthesis. So it agrees. It doesn't have an ego in this case.
08:36With both reviews on the high priority items, here's what I would prioritize. So now you give Claude Code the driver's seat taking consultations from these models.
08:45On the must fix, it says split typography, fix grid, add focus, smooth FAQ animation, same thing on the copy.
08:53And then it says, okay. These are the things that are worth testing that they mentioned. And then it says, do you want me to implement any of these changes now?
09:00I do. It takes five, ten minutes, and we get a much better version right here where it's deeper in terms of the outcome.
09:08It's better organized. Some things here, especially on the font, look a lot cleaner.
09:13The numbers here specifically look way better than they used to before, and overall, it's a lot tighter. It has some more FAQs. It has better pricing, more detailed and thoughtful pricing, and we did all this technically with just a single prompt.
09:27Now let's say that was a complete fluke. Let's apply it to another scenario and see how it behaves. So if I say here, create a new folder and create an analytics dashboard showing some mock data for a venture capital firm, and then I tell it to spin up on a different port.
09:41It will go and it then created something like this, which is okay. It's not the prettiest, but it's manageable.
09:49And then the other version looked pretty rough as well. But in this case, we did the exact same thing, and then we ran an audit sending the exact same prompt and just experiment,
09:59I sent this prompt right here without even invoking the slash command just to see if I sent it, would it be able to pick up the skill and then invoke it? In this case, it executes the council plan, asks me if we should proceed, then it goes through and brings a very heavy front end audit. So it tells me what's solid, what are the top fixes.
10:19Same thing with codex on the copy side, it scores it seven out of 10, and it comes up with both the issues as well as its proposed fixes. The main moral of this is that you don't have to just blindly follow what these other language models are proposing. They can just add some more brainstorming.
10:35They can make you think more out of the box, and you might not go with any of these. We might come up with your own from having seen what they come up with. And, again, Claude says here my take.
10:44I agree with both models on high priority items. And then where I'd push back slightly, Codex's suggestion to rename markups, markdowns to unrealized value uplifts is technically correct, but for a dashboard, the shorter labels are fine.
10:57So there is this level of judgment while still taking the different pieces of advice into account. In this case, I just had Cloud Code blindly follow their suggestions and integrate it into its plan, and it came up with this very Bloomberg terminal looking UI that I think could be better, but overall looked infinitely better than the initial one, again, with just one prompt taking external perspectives into account.
11:20And you can apply this skill to all kinds of scenarios, especially non scenarios where you can bring in different language models with different strengths. You can even take the same language model like Claude and go through different versions of it to see what versions might say about the exact same task. So you could check haiku, you can check sonnet, and you check opus all on the exact same task or the exact same copy or whatever it is you're trying to accomplish.
11:45So hopefully that shows you the power of creating a mega skill. And the beauty of these mega skills is you can combine them with additional skills. Let's say the front end skill and create some form of symbiosis between both of them.
11:56So the more skills you can stack and the deeper you can make them, the higher leverage you'll have. And like I said, I'll attach the skill that I showed you in this video in the second link in the description below. If you want access to my exclusive mega skills, then you'll want to check out the first link in the description below from our early AI adopters community.
12:13CTAAnd for the rest of you if you found this helpful, found it novel and you want to see more then please leave me a like and a comment on the video it really helps the video and the channel. I'll see you in next one.
— full transcript
§ 05 · For Joe
How to consult multiple AI models from a single session.
WHAT TO LEARN
The Council skill shows that the bottleneck in multi-model workflows is not access to models but the discipline of context packaging, routing, and synthesis.
- Auto-routing by task category removes the cognitive overhead of choosing which model to ask for each problem — bug fix goes to Codex, frontend review goes to Gemini, architecture stays with Claude.
- Context packaging is the real work: sending the problem statement, relevant snippets, and prior attempts (not the full conversation) gets better responses than dumping everything.
- A model that disagrees with another model is more useful than one that agrees — the disagreement surfaces the blind spots in each model's training.
- The executor role belongs to Claude: external model opinions are inputs to synthesis, not commands to follow blindly.
- Pushing back on technically-correct suggestions is what distinguishes synthesis from aggregation — Claude declining verbose VC terminology for a compact dashboard is the skill working as intended.
- The same consultation pattern applies to non-coding tasks wherever a second opinion from a different model perspective adds signal.
- Stacking skills compounds leverage: a frontend design skill plus a Council skill produces results neither could reach alone.
