julzzz

Not average blondie - I build with AI, not just talk about it

[julzzz @ju1lzzz]

1.9M people installed a "legal team." Almost none turned on the one thing that keeps it safe

And it isn't a setting. It's the habit of reading what it hands back before you sign - because the plugin is a second reader, not a signer.

The real move is older than the plugin: point it at your own contract and let it surface what you'd skim past. One pass once caught an auto-renewal clause I'd have signed straight through.

Watch the one beat that matters:

00:21 - /review-contract runs

Nothing in between on the part that decides everything: it can be confidently, completely wrong, so you verify anything touching law yourself.

A plugin makes you faster. It doesn't make you covered. So which did you actually install - a tool, or a reason to stop reading?

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

The video's right: documentation is the unlock. The trap is treating it as a one-time build

A knowledge base is context, and context is a multiplier - it compounds while you maintain it and rots the moment you don't. The same base that turns "I tried it, it was okay" into compounding, month-over-month value is the one that quietly starts lying once it goes stale.

Building it takes an afternoon. Maintaining it is the actual job, and it's the step almost everyone skips. So the differentiator was never "do you have documentation." It's whether you run a loop on top of it - interrogate it, prune it, red-team what it tells you - or just trust the first confident answer at business scale.

Building the base is the easy 10%. Are you doing the other 90%, or just automating your blind spots?

Watch from 2:29 - where he builds a skill whose only job is to keep the base updated. That's the part that actually decides whether this works.

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

@x0Kairo Exactly - that's the gap between doing things faster and doing better things. Context without direction is just faster mediocrity

[julzzz @ju1lzzz]

A context file doesn't 10x anything. It multiplies whatever you were already doing - including your bad habits

Context is the multiplier on every other workflow. So if the workflow underneath is still "just give me the answer," all you've built is a faster route to personalized mediocrity.

The interview trick is real - ten minutes and the file writes itself. But the file isn't the win. Pointing it at the workflows that actually think is: make Claude interrogate you, red-team your plan, run the critique loop. Alone, it just makes generic answers sound more like you.

And nobody warns you it rots. Stale context quietly produces stale advice - the file you built once and forgot is lying to you by month three.

So before you copy the prompt: is your context file feeding a thinking process, or decorating a lazy one?

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

@GMINDSCOUK Right - it's a multiplier, not a filter. That's why the thinking process underneath matters more than the file itself

[julzzz @ju1lzzz]

@jurlycat The template wins because it doesn't force people to think - but that's exactly why it doesn't win at the margin

[julzzz @ju1lzzz]

The 9 steps build the machine. None of them teach you to drive it

The machine runs on tools. The driving runs on one habit: never ship the first pass.

You can finish all 9 steps, wire up every tool, and still use Claude like a search engine on Monday: type, read, close tab. The one step on that list that actually teaches you to drive is the context file, because it is the only thing that changes every future conversation. Context is the multiplier on everything else. One configured .md really does beat 50 random prompts - the rest of the checklist just gets you connected.

An assistant that "works while you sleep" is only ever as sharp as the file you handed it and the loop you run on what it gives back.

So before you spend the weekend building the machine: what are you actually going to do with the first draft it hands you?

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

There's no hidden code inside Claude. There's a persona you assign - and a loop most people never run

/god mode isn't a switch Anthropic buried in the model. The word just tells Claude to role-play omniscience, exactly like /steve_jobs or /forensic_analyst would. You didn't hack anything, you assigned a mask.

And a mask only touches the first output. My article makes the opposite bet: the first output is raw material, not the product. The return comes from the loop you run next - red-teaming the draft with "how does this die?" instead of "is this good?", then rewriting.

Same model, completely different return, and none of it came from a tag.

So which are you actually collecting: five magic words, or the one habit that compounds?

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

The future of AI coding isn't a smarter model. It's whoever structured their notes best

A better prompt buys you one good answer. A well-structured knowledge base buys you every answer after it.

That combo everyone's clipping isn't magic - it's the context-file idea taken to its visual extreme: give the model a memory. Ten minutes to set up, and every session after starts at depth instead of at zero. The model reads your whole project before it writes a line, which is why it stops behaving like a black box - there's nothing left for it to guess.

So the leverage quietly moves off the model and onto you. Not the person with the cleverest prompt, but the one who organized what the AI reads. Context is the multiplier on everything else you do with it.

When the model can already read everything you know, is a smarter model even the upgrade you're missing?

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

@jurlycat True, though the real risk shows up the moment you do add multiple agents without shared context

[julzzz @ju1lzzz]

Multi-agent AI isn't an org-chart problem. It's a memory problem

In my own pipeline the biggest multiplier was never more agents - it was a shared context file every step reads from before it acts.

Give one manager 5-7 subagents and you haven't added capability, you've added handoffs. And context dies at handoffs. Human names help the humans, not the machines - calling it "Nolan" makes delegation feel natural, but the agent still starts from zero unless something carries the why between layers.

An org chart moves tasks down. It doesn't move understanding sideways. Structure is the visible 10%; the memory layer underneath is the 90% that decides whether your virtual company thinks or just forwards.

So before you draw the hierarchy: who actually holds the context between your agents?

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

@Bullrun8x Right, taste and judgment can't be templated - that's the real skill gap

[julzzz @ju1lzzz]

50 UI styles won't make Claude a designer. One decision will.

A palette library solves the one problem AI never actually had: access to nice-looking parts. The bottleneck was never the components - it was knowing which one the problem calls for, and why. That's UI versus UX: one is what you see, the other is what survives contact with a real user.

Notice why the good design skills actually work. Frontend Design doesn't win on more gradients - it wins because it forces a choice of aesthetic first, then builds to it, instead of defaulting to purple-gradient slop. The leverage is the decision before the code, not the size of the shelf.

A component MCP plus 97 palettes gives you a fast decorator. Worth setting up - but the first output is raw material, not the product.

If beautiful parts made you a designer, stock libraries would have replaced the profession years ago. Why didn't they?

[Bober_smart @Bober_smart]

https://t.co/p4teS18nnJ

[julzzz @ju1lzzz]

@calebturnerlab Well put - the 'why' is the actual craft, everything else is just tooling

[julzzz @ju1lzzz]

@framexin Exactly - more components just means more ways to make the wrong choice faster

[julzzz @ju1lzzz]

@jurlycat This is the real split - it's not tier vs tier, it's operators vs people still pricing their time instead of their output

[julzzz @ju1lzzz]

AI isn't replacing the middle designer. It's replacing the one who refuses to operate it

The video's right that the €2,000 tier gets eaten first. It's wrong about where that work goes.

That budget doesn't vanish. The owner behind it was never going to learn Framer AI or write clean prompts - they don't have the time, and they never wanted a design, they wanted the task done. So the tier stays open, just for a different role: the operator who runs the same skills I use to move 3-4x faster and ships the whole thing turnkey in an afternoon.

"Go premium" is only half the advice. The half nobody says: the bottom doesn't close, it just stops paying designers and starts paying operators.

So the real split isn't top segment vs bottom. It's who runs the AI versus who competes with it. Which side are you building toward?

[Bober_smart @Bober_smart]

https://t.co/p4teS18nnJ

[julzzz @ju1lzzz]

@nathanbrooksx This is such a needed reframe - AI handling the 'task done' part actually frees designers to focus on the judgment and meaning that make design good, not just finished. Exciting to see where that shift takes the field.

[julzzz @ju1lzzz]

The hack exposes a bigger question: what's left to hire for when anyone can generate the output?

The first output is raw material, not the product - and the whole role now lives in what you do to it after.

Anyone can make Claude produce a plausible answer in an interview. That skill is now free. What still isn't free is the loop on top of it: asking "how does this die?" instead of "is this good?", red-teaming the output, running two critique passes before anything ships. Same model in every hand. Completely different return.

So the hack works - it just moves the real test to week one, when someone asks why the confident output broke in production. The people gaming the filter and the people who can actually interrogate a model's answer are about to be very easy to tell apart.

The screen measures who can generate. The job is about to measure who can tell when the model is confidently wrong.

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC

[julzzz @ju1lzzz]

@crimson_moti Exactly, couldn't agree more

[julzzz @ju1lzzz]

You don't need 2,277 skills. You need these 6 pointed at a workflow. Here's which

I picked by one rule: what actually breaks if I uninstall it. Six passed.

Document analyst. Point it at your own contracts, bank exports, a year of notes, and tell it to flag what contradicts, what's unusual, what you missed. Highest-ROI skill in the whole list - it reads every line with equal attention, which you never do.

Memory / context. A living file of who you are, your goals, your standards, the decisions already made. Ten minutes to set up, and every future chat starts at depth instead of zero. This one is the multiplier on all the others - without it, every skill resets to a stranger each session.

Critique / editor. Runs the loop most people skip: generate the draft, attack its three weakest points, rewrite only those. The second and third pass is where quality actually lives - the first output was never the product.

Red-team / skeptic. Feed it any real decision - pricing, a launch, a job change - and force it to argue how the plan dies, not why it works. Confirmation is free everywhere. Structured disagreement is the rare asset.

Data / spreadsheet. Hand it a messy CSV or an export and let it clean, compute, and chart. Turns "the numbers are in there somewhere" into an answer you can act on.

Automation / schedule. For the recurring stuff - a morning digest, a weekly review - so the workflow runs without you re-typing the same prompt every time.

Notice what they share: not one is impressive on its own. Each is a tool bolted onto a workflow you already run. The store hands you 2,277 skills - the number that matters is how many are pointed at something.

Which of the six would change your week if you actually used it?

[julzzz @ju1lzzz]

https://t.co/6kyuSZiOWC