I cut my team from 10 people to 4.
Not because of money. Because of AI.
The surprising part? Less drama. More output. Fewer meetings about meetings.
Small teams aren't a downgrade. They're a filter.
5:58 AM · Jun 23, 2026 · 48.2K Views
142 389 2.1K
Theme 1 · less drama angle
A
Arya Pradana @aryapradana_
···
Going from 10 to 4 people taught me something nobody warns you about.
Half my "management problems" were never management problems.
They were just too many people in one room.
It is headcount.
The truth is: drama scales faster than output.
7:12 AM · Jun 23, 2026 · 31.7K Views
98 256 1.4K
Theme 2 · AI-assisted → AI-native
A
Arya Pradana @aryapradana_
···
Most teams use AI like a faster intern.
That's "AI-assisted." It's a dead end.
We rebuilt every workflow to assume AI does the first draft of everything.
That's AI-native.
The truth is: bolting AI onto old processes just makes slow things slightly faster.
9:03 AM · Jun 23, 2026 · 52.9K Views
173 612 3.3K
Theme 3 · mental model shift
A
Arya Pradana @aryapradana_
···
The hardest part of going AI-native isn't the tools.
It's unlearning "more people = more capacity."
The new model: leverage per person, not headcount.
It is a mindset.
The truth is: hiring used to be the answer because thinking was the bottleneck. It isn't anymore.
11:20 AM · Jun 23, 2026 · 39.4K Views
121 421 2.6K
Theme 3 · the role shift
A
Arya Pradana @aryapradana_
···
My 4 people don't have smaller jobs than the old 10.
They have bigger ones.
Each one now owns an outcome, not a task. AI handles the tasks.
It is ownership.
The truth is: AI didn't replace my team. It deleted the roles that were never really thinking.
2:45 PM · Jun 23, 2026 · 44.1K Views
134 503 2.9K
Threads
Theme 1 · 10 → 4
A
aryapradana_· 6h···
I went from a team of 10 to a team of 4. AI made it possible, but here's what nobody tells you:
The drama dropped before the output went up.
Fewer people. Fewer politics. More actual work.
It is a filter.
What did your team actually lose when it got bigger?
214 replies · 1,389 likes
Theme 1 · less drama angle
A
aryapradana_· 5h···
Unpopular take after downsizing 10 → 4:
Most "team culture problems" are just a math problem. Too many people, not enough clarity.
When the team shrank, the drama didn't shrink with it. It disappeared.
It is headcount.
Drama scales faster than output. Agree or disagree?
301 replies · 976 likes
Theme 2 · AI-assisted → AI-native
A
aryapradana_· 4h···
Honest question for founders using AI:
Are you AI-assisted or AI-native? Most are assisted — using AI like a faster intern. That's a dead end.
We rebuilt every workflow so AI writes the first draft of everything.
It is a rebuild.
Bolting AI onto old processes just makes slow things slightly faster. Where are you on this?
428 replies · 2,104 likes
Theme 3 · mental model shift
A
aryapradana_· 2h···
The hardest part of going AI-native wasn't learning new tools.
It was unlearning one belief: "more people = more capacity."
The new model is leverage per person, not headcount per problem.
It is a mindset.
Hiring used to be the answer because thinking was the bottleneck. It isn't anymore. What belief did AI force you to drop?
356 replies · 1,602 likes
Theme 3 · the role shift
A
aryapradana_· 1h···
People assume my 4 people have smaller jobs than the old 10. Opposite.
They have bigger ones. Each owns an outcome now, not a task. AI handles the tasks.
It is ownership.
AI didn't replace my team. It deleted the roles that were never really thinking. Harsh? Tell me where I'm wrong.
389 replies · 1,847 likes
Batch 2 — Generalist + AI Creation Workflow
Grounded in what we've actually built together · stack named, in-dev products kept vague
X / Twitter
Generalist · why now is the moment
A
Arya Pradana @aryapradana_
···
For years "jack of all trades, master of none" was an insult.
I'm a generalist. AI just made that my biggest advantage.
I can design, build, write, and ship without waiting on four specialists.
It is leverage.
The truth is: AI doesn't reward the deepest expert. It rewards the person who can connect the most dots.
8:10 AM · Jun 24, 2026 · 61.3K Views
188 724 4.1K
Workflow · idea → shipped
A
Arya Pradana @aryapradana_
···
The gap between "idea" and "live product" used to be a hiring plan.
Now it's an afternoon.
We sketch in Claude, build in Cursor, spin the UI in v0, ship our own tools. No handoff. No waiting.
It is collapse.
The truth is: the bottleneck was never talent. It was the distance between roles.
10:40 AM · Jun 24, 2026 · 54.7K Views
156 567 3.0K
Workflow · AI is a decision
A
Arya Pradana @aryapradana_
···
Using every AI tool for everything is a beginner move.
Picking the right one per job is the whole skill.
Claude to think. Cursor to build. v0 for UI. The model is a choice, not a default.
It is judgment.
The truth is: the people losing to AI aren't using it less. They're using it without deciding why.
1:25 PM · Jun 24, 2026 · 47.9K Views
142 498 2.7K
Workflow · one person, full pipeline
A
Arya Pradana @aryapradana_
···
A real product used to need a designer, a dev, a copywriter, a PM.
I ran that whole pipeline solo last month and shipped.
Not because I'm exceptional. Because the pipeline finally fits one head.
It is compression.
The truth is: AI didn't make me 10x smarter. It removed the handoffs that slowed me down.
4:05 PM · Jun 24, 2026 · 58.2K Views
167 631 3.4K
Threads
Generalist · why now is the moment
A
aryapradana_· 7h···
For most of my career, being a generalist felt like a weakness. "Jack of all trades, master of none."
Then AI showed up and flipped it.
I can design, build, write, and ship without four specialists in a chain.
It is leverage.
AI rewards the person who connects the most dots, not the deepest expert. Generalists, are you feeling this too?
412 replies · 2,233 likes
Workflow · idea → shipped
A
aryapradana_· 5h···
The gap between "I have an idea" and "it's live" used to be a hiring plan. Now it's an afternoon.
Our actual flow: think in Claude, build in Cursor, spin up UI in v0, ship our own tools. No handoff, no waiting on anyone.
It is collapse.
The bottleneck was never talent. It was the distance between roles. What's your idea-to-live time these days?
367 replies · 1,905 likes
Workflow · AI is a decision
A
aryapradana_· 3h···
Hot take: using every AI tool for everything is a beginner move.
The real skill is picking the right one per job. Claude to think. Cursor to build. v0 for UI. The model is a choice, not a default.
It is judgment.
The people losing to AI aren't using it less. They're using it without deciding why. What's your one go-to tool, and for what?
489 replies · 2,041 likes
Workflow · one person, full pipeline
A
aryapradana_· 1h···
A real product used to need a designer, a dev, a copywriter, a PM. Four people minimum.
Last month I ran that whole pipeline solo and shipped. Not because I'm exceptional. The pipeline finally fits one head.
It is compression.
AI didn't make me 10x smarter. It removed the handoffs that slowed me down. Where do handoffs still cost you the most time?
421 replies · 1,988 likes
X Thread — the full 10 → 4 story
Highest-leverage format on X: each post gets its own algorithmic boost. Post 1 is the hook — spend as much on it as the rest combined.
A
Arya Pradana@aryapradana_ · 1/9
I cut my team from 10 people to 4.
Revenue didn't drop. Output went up. And the office got quieter in the best way.
Here's exactly what happened, and the mental model that made it work 🧵
A
@aryapradana_ · 2/9
First, the honest part.
This wasn't a cost-cutting move dressed up as strategy.
AI changed what one person can do in a day. The 10-person structure was built for a problem that stopped existing.
A
@aryapradana_ · 3/9
The thing I didn't expect: the drama left before the productivity arrived.
10 people meant 10 sets of feelings, side conversations, and "can we sync about the sync."
4 people who each own an outcome? Almost none of that.
A
@aryapradana_ · 4/9
Most "culture problems" I spent years managing were really just coordination problems.
Too many handoffs. Too many people who needed to be looped in.
Shrink the team and a lot of that friction simply disappears.
A
@aryapradana_ · 5/9
But headcount wasn't the real shift. The workflow was.
We went from AI-assisted → AI-native.
Assisted = using AI like a faster intern, bolted onto old processes.
Native = AI writes the first draft of everything, humans edit and decide.
A
@aryapradana_ · 6/9
That gap is bigger than it sounds.
Bolt AI onto a slow process and you get a slow thing, slightly faster.
Rebuild the process around AI and you change what the process even is.
A
@aryapradana_ · 7/9
The mental model I had to unlearn:
"More people = more capacity."
The new one: leverage per person, not headcount per problem.
Hiring used to be the answer because thinking was the bottleneck. It isn't anymore.
A
@aryapradana_ · 8/9
So my 4 people don't have smaller jobs than the old 10.
They have bigger ones. Each owns a full outcome. AI handles the tasks underneath it.
That's the trade: fewer people, far more ownership each.
A
@aryapradana_ · 9/9
If you're staring at your org chart wondering why it feels heavy:
It might not be a people problem. It might be a model problem.
The truth is: AI didn't replace my team. It deleted the roles that were never really thinking.
If this was useful, follow for more on building AI-native teams.
318 1.2K 7.4K
Batch 3 — The Founder Journey, Continued
46 posts from Dewi's bank (account: Mas Arya), written in Arya voice + anti-AI pass · mirrored X / Threads · grouped by source
X / Twitter
Experimental 5 posts
Experimental · E01
A
Arya Pradana @aryapradana_
···
Giving an AI agent write access to our codebase was always going to break something.
It doesn't mean the experiment failed.
It means we found the edges before a client ever would.
It is a guardrail.
What keeps it safe?
Tight limits. Real reviews. Easy undo.
The truth is: you don't trust an agent until you've watched it fail and caught it.
From Dewi's bank · Thread · CTA: None
Experimental · E02
A
Arya Pradana @aryapradana_
···
Most ways to bolt memory onto an LLM app fall apart under real load.
It doesn't mean memory is hype.
It means a demo and a workload are different animals.
It is a stress test.
What survives?
Real traffic. Real cost. Real recall.
The truth is: the version that works in a demo is rarely the one you'd ship.
From Dewi's bank · Thread · CTA: Soft DM
Experimental · E03
A
Arya Pradana @aryapradana_
···
Running an AI feature at 10x traffic just to see the bill felt reckless.
It doesn't mean we were careless.
It means we'd rather find the cost cliff than fall off it later.
It is a dry run.
What changed after?
Leaner calls. Smarter caching. Honest quotes.
The truth is: a feature you can't afford at scale isn't done, it's a liability waiting.
From Dewi's bank · Thread · CTA: None
Experimental · E04
A
Arya Pradana @aryapradana_
···
Asking the design team to ship a working prototype in one afternoon sounds like pressure.
It doesn't mean we cut corners.
It means the tools finally make a tight timebox realistic.
It is a sprint.
What makes it work?
Clear scope. Fast tools. One owner.
The truth is: a deadline you can actually hit teaches more than a roadmap you can't.
From Dewi's bank · Thread · CTA: None
Experimental · E05
A
Arya Pradana @aryapradana_
···
Most AI demos look great because someone picked the perfect inputs.
It doesn't mean the product is fake.
It means you haven't seen it handle the messy stuff yet.
It is a magic trick.
So what did we do?
Ugly inputs. Edge cases. Bad data.
The truth is: the demo that never breaks is the one nobody pushed hard enough.
From Dewi's bank · Tweet+media · CTA: None
Owned product 5 posts
Owned product · OP01
A
Arya Pradana @aryapradana_
···
Good AI tools getting priced out of reach is normal now.
It doesn't mean you're stuck paying.
It means there's room for an open one you actually own.
It is ESflows.
Why build it?
Open source. Your key. No subscription.
The truth is: a tool you can open and change beats a tool you can only rent.
From Dewi's bank · Thread · CTA: Link (demo)
Owned product · OP02
A
Arya Pradana @aryapradana_
···
Losing context every time you switch AI models is just how it works now.
It doesn't mean you accept it.
It means there's a missing layer nobody built.
It is ViMemory.
What does it do?
Memory that follows. Across models. Yours alone.
The truth is: the model isn't your moat, the memory you keep between them is.
From Dewi's bank · Thread+video · CTA: Link
Owned product · OP03
A
Arya Pradana @aryapradana_
···
Paying monthly for convenience feels easy at first.
It doesn't mean it's cheap.
It means the meter runs whether you use it or not.
It is the math.
So why BYOK?
Your key. Your usage. Your bill.
The truth is: a subscription charges you for the days you didn't even log in.
From Dewi's bank · Thread · CTA: None
Owned product · OP04
A
Arya Pradana @aryapradana_
···
Promising that user data stays separate is easy.
It doesn't mean much on its own.
It means nothing until the separation is actually enforced.
It is isolation.
How does ViMemory do it?
Own namespace. Hard walls. No leaks.
The truth is: 'we promise' is marketing, 'it can't physically happen' is design.
From Dewi's bank · Thread · CTA: Link (install)
Owned product · OP06
A
Arya Pradana @aryapradana_
···
Our own product's first version was wrong in places.
It doesn't mean we shipped junk.
It means we're willing to run the teardown on ourselves.
It is honesty.
What did we change?
Clearer flows. Fewer features. Real fixes.
The truth is: a studio that won't critique its own work has no business critiquing yours.
From Dewi's bank · Thread · CTA: None
Previous client project 3 posts
Previous client project · PC02
A
Arya Pradana @aryapradana_
···
A client asking for AI in one specific feature is normal.
It doesn't mean that's where it belongs.
It means our job is to find where it actually helps.
It is judgment.
What guided the call?
Real use. Real value. Real fit.
The truth is: putting AI where the client asked is easy, putting it where it works takes a conversation.
From Dewi's bank · Thread · CTA: None
Previous client project · PC04
A
Arya Pradana @aryapradana_
···
Throwaway prototypes are treated as a normal cost of design.
It doesn't mean they have to be.
It means most of that waste is a process problem.
It is a straight line.
How do we skip it?
Real components. Early decisions. No dead ends.
The truth is: every throwaway prototype is a decision someone was scared to make sooner.
From Dewi's bank · Thread · CTA: Soft DM
Previous client project · PC05
A
Arya Pradana @aryapradana_
···
Inheriting a messy AI-assisted codebase is becoming common.
It doesn't mean the last team was bad.
It means fast got prioritized over lasting.
It is a rescue.
What did we fix?
Dead code. Silent costs. Weak seams.
The truth is: AI makes it easy to build something that works today and collapses next quarter.
From Dewi's bank · Thread · CTA: Teardown DM
Concept use case 4 posts
Concept use case · CU02
A
Arya Pradana @aryapradana_
···
An AI that answers confidently feels trustworthy.
It doesn't mean you should trust it.
It means confidence and accuracy aren't the same thing.
It is Cairn.
What did we design?
Cited answers. A verification path. Real sources.
The truth is: 'knowledge you can trust' isn't a better search bar, it's a way to check the answer.
From Dewi's bank · Thread · CTA: None
Concept use case · CU03
A
Arya Pradana @aryapradana_
···
Letting AI speed up money decisions sounds risky.
It doesn't mean you keep it out entirely.
It means you decide exactly where it's allowed to act.
It is Keel.
Where's the line?
AI triages. Humans approve. Money waits.
The truth is: the smartest AI in fintech is the one that knows when to stop and ask.
From Dewi's bank · Thread · CTA: None
Concept use case · CU04
A
Arya Pradana @aryapradana_
···
Trusting an AI agent's memory blindly is the default right now.
It doesn't mean it's safe.
It means you can't audit what you can't see.
It is Mneme.
What did we build in?
Traceable memory. Clear provenance. Real receipts.
The truth is: an agent you can't audit is just a confident stranger with your keys.
From Dewi's bank · Thread · CTA: None
Concept use case · CU06
A
Arya Pradana @aryapradana_
···
Most CRMs looking identical is just accepted.
It doesn't mean it's fine.
It means everyone shipped the default and stopped.
It is Meridian.
What did we change?
Warm paper. Mono numerals. Real character.
The truth is: 'it looks like every other CRM' is a choice, and usually a lazy one.
From Dewi's bank · Thread · CTA: Soft DM
Testimony client 2 posts
Testimony client · T02
A
Arya Pradana @aryapradana_
···
Praise for something being flashy fades fast.
It doesn't mean flashy is the goal.
It means the feedback that lasts is about what kept working.
It is the real pitch.
What do clients notice?
Still running. Still cheap. Still solid.
The truth is: the best review isn't 'it's amazing,' it's 'it still works and still costs the same.'
From Dewi's bank · Tweet · CTA: None
Testimony client · T05
A
Arya Pradana @aryapradana_
···
Selling speed is easy and everyone does it.
It doesn't mean speed is what clients remember.
It means trust is the thing that actually closed them.
It is trust.
How do we earn it?
Clear calls. Honest tradeoffs. No surprises.
The truth is: clients forget how fast you were long before they forget whether they could trust you.
From Dewi's bank · Thread · CTA: None
Call with client 5 posts
Call with client · CA01
A
Arya Pradana @aryapradana_
···
A client asking to 'add AI everywhere' is normal now.
It doesn't mean you do it.
It means one good question can save them a whole rebuild.
It is restraint.
What do we ask?
Where's the pain? Who benefits? Why now?
The truth is: the best thing you can sell a client is the AI feature you talked them out of.
From Dewi's bank · Thread · CTA: None
Call with client · CA02
A
Arya Pradana @aryapradana_
···
'Can you make it cheaper?' comes up on almost every call.
It doesn't mean the client is cheap.
It means they don't yet see what drives the cost.
It is a fair question.
What's the honest answer?
Running cost. Real maintenance. True scale.
The truth is: cutting the build price often just moves the bill to the running cost nobody mentioned.
From Dewi's bank · Tweet · CTA: None
Call with client · CA03
A
Arya Pradana @aryapradana_
···
Quoting fast feels like good service.
It doesn't mean it helps the client.
It means a few questions first save everyone money.
It is diligence.
What do we always ask?
Who's it for? What breaks? What's success?
The truth is: the quote you give before understanding the problem is just a number you'll both regret.
From Dewi's bank · Thread · CTA: Soft DM
Call with client · CA04
A
Arya Pradana @aryapradana_
···
A client who got burned by a cheap AI build comes in guarded.
It doesn't mean they're difficult.
It means they learned a real lesson the hard way.
It is caution.
What did we show them?
Working proof. Plain answers. No hype.
The truth is: the client who's been burned once is the one who'll value real work the most.
From Dewi's bank · Thread · CTA: None
Call with client · CA05
A
Arya Pradana @aryapradana_
···
'Our last dev used AI to build it fast' is meant as a flex.
It doesn't mean it was built well.
It means fast was the goal and lasting wasn't.
It is a warning sign.
What actually matters?
Still running. Still cheap. Still fixable.
The truth is: fast is the easy part, surviving six months is the part nobody brags about.
From Dewi's bank · Tweet · CTA: None
Discussion with client 5 posts
Discussion with client · D01
A
Arya Pradana @aryapradana_
···
Wanting AI to answer instantly and directly feels like good UX.
It doesn't mean it builds trust.
It means a fast answer with no source is just a confident guess.
It is sourcing.
What did we pick?
Show the source. Cite the basis. Earn belief.
The truth is: a direct AI answer feels faster, a sourced one is the only kind people come back to.
From Dewi's bank · Thread · CTA: None
Discussion with client · D02
A
Arya Pradana @aryapradana_
···
Talking a client out of a feature they asked for feels wrong.
It doesn't mean you're being difficult.
It means you're protecting the project from bloat.
It is the job.
What did that hour buy?
Sharper scope. Lower cost. Better product.
The truth is: the features you remove quietly do more for a build than the ones you proudly add.
From Dewi's bank · Tweet · CTA: None
Discussion with client · D03
A
Arya Pradana @aryapradana_
···
A messy kickoff discussion is normal.
It doesn't mean the project is doomed.
It means the mess just needs one clear page to land on.
It is a spec.
How do we get there?
One page. Shared words. Real signoff.
The truth is: a project rarely fails from a hard problem, it fails because nobody wrote down what they agreed to.
From Dewi's bank · Thread · CTA: Soft DM
Discussion with client · D04
A
Arya Pradana @aryapradana_
···
'Our competitors have an AI chatbot' is a common reason to want one.
It doesn't mean your users want one.
It means a competitor's roadmap became your brief.
It is a reframe.
What do we ask instead?
Who needs it? What for? Says who?
The truth is: copying a competitor's feature is the fastest way to inherit their mistakes.
From Dewi's bank · Thread · CTA: None
Discussion with client · D05
A
Arya Pradana @aryapradana_
···
Assuming the code is the hard part is natural.
It doesn't mean it's true.
It means the real fight is agreeing on what 'done' means.
It is alignment.
How do we settle it?
Define done. Write it down. Agree early.
The truth is: most projects don't stall on hard problems, they stall on two people picturing different finish lines.
From Dewi's bank · Thread · CTA: None
Revenue project 3 posts
Revenue project · R02
A
Arya Pradana @aryapradana_
···
A feature getting more expensive to run over time is expected.
It doesn't mean you accept it.
It means most teams never check the bill after launch.
It is the real metric.
What do we track?
Running cost. Steady spend. No creep.
The truth is: shipping a feature is easy, shipping one that still costs what you quoted is the hard part.
From Dewi's bank · Tweet · CTA: None
Revenue project · R03
A
Arya Pradana @aryapradana_
···
Treating every paid build as a one-off is normal.
It doesn't mean it has to be.
It means the second build in a space should reuse the first.
It is compounding.
How does it speed up?
Reused patterns. Known traps. Faster ship.
The truth is: the value of a paid project isn't the fee, it's the reusable pattern you keep afterward.
From Dewi's bank · Thread · CTA: None
Revenue project · R05
A
Arya Pradana @aryapradana_
···
Naming a high price for the first time is nerve-wracking.
It doesn't mean the number is wrong.
It means saying it out loud is the real test.
It is nerve.
What did pricing teach me?
Know your value. Say it plainly. Hold the line.
The truth is: pricing a build isn't a math problem, it's a confidence problem wearing a math costume.
From Dewi's bank · Thread · CTA: None
Project timeframe 4 posts
Project timeframe · PT01
A
Arya Pradana @aryapradana_
···
A short build timeline sounds too good to be real.
It doesn't mean corners were cut.
It means the steps are tighter when AI is in the loop.
It is a real timeline.
What does the speed need?
Clear scope. Fast feedback. One owner.
The truth is: an AI-native build isn't faster because you skip steps, it's faster because fewer people wait on each other.
From Dewi's bank · Thread · CTA: Soft DM
Project timeframe · PT02
A
Arya Pradana @aryapradana_
···
A second build going twice as fast feels like luck.
It doesn't mean it was.
It means the first build left you a stack to stand on.
It is compounding.
Where's the speed from?
Known patterns. Reused parts. Fewer surprises.
The truth is: your second build in a space is faster because of what you refused to throw away from the first.
From Dewi's bank · Thread · CTA: None
Project timeframe · PT03
A
Arya Pradana @aryapradana_
···
A fast build aging badly is common.
It doesn't mean speed is the enemy.
It means the speed came from shortcuts, not systems.
It is the difference.
How do we go fast safely?
Reused parts. Clean seams. No debt.
The truth is: fast from systems lasts, fast from shortcuts just postpones the rebuild.
From Dewi's bank · Tweet · CTA: None
Project timeframe · PT05
A
Arya Pradana @aryapradana_
···
Expecting an AI feature to be quick is fair.
It doesn't mean it's quick to finish.
It means a demo and a shipped feature live on different clocks.
It is two timelines.
What's the gap?
Fast to demo. Slow to harden. Worth it.
The truth is: the demo takes an afternoon, the part that survives real users takes the rest of the month.
From Dewi's bank · Tweet · CTA: None
Tool and stack 5 posts
Tool and stack · TS01
A
Arya Pradana @aryapradana_
···
Starting every project from a blank file feels disciplined.
It doesn't mean it's smart.
It means you rebuild the same foundation every time.
It is one stack.
What do we standardize?
Vite. React. Tailwind.
The truth is: the blank file isn't a fresh start, it's a tax you pay at the beginning of every project.
From Dewi's bank · Thread · CTA: Link (porto)
Tool and stack · TS02
A
Arya Pradana @aryapradana_
···
Picking AI tools for what's exciting is tempting.
It doesn't mean it's wise.
It means hype fades and the bill stays.
It is a stack choice.
What earns a spot?
Low cost. Easy upkeep. Proven fit.
The truth is: the most impressive AI stack on launch day is rarely the one you can still afford to run a year later.
From Dewi's bank · Thread · CTA: Link
Tool and stack · TS03
A
Arya Pradana @aryapradana_
···
Obsessing over which AI tool to use is normal.
It doesn't mean it's the real risk.
It means the bigger gap is having no plan when it fails.
It is your fallback.
What should you ask?
What breaks? Then what? Says who?
The truth is: your AI tool choice matters less than your fallback, and most teams don't have one.
From Dewi's bank · Tweet · CTA: None
Tool and stack · TS04
A
Arya Pradana @aryapradana_
···
Sending your data to a hosted model is the default.
It doesn't mean it's the only way.
It means self-hosting is a tradeoff worth weighing.
It is a choice.
Why self-host the embeddings?
Nothing leaves. No fees. Full control.
The truth is: 'free and private' usually costs you some setup, and that's often a bargain.
From Dewi's bank · Thread · CTA: None
Tool and stack · TS06
A
Arya Pradana @aryapradana_
···
Dropping a tool you tried feels like wasted time.
It doesn't mean the test was a waste.
It means knowing what to drop is its own skill.
It is honest churn.
What survived this quarter?
Real use. Clear win. Earned keep.
The truth is: the tools you quietly drop teach you more about your work than the ones you keep.
From Dewi's bank · Thread · CTA: None
Following trend 5 posts
Following trend · FT01
A
Arya Pradana @aryapradana_
···
Everyone shipping the same trend at once is normal.
It doesn't mean it's ready.
It means most people are reacting to the demo, not the workload.
It is a real test.
What did we do?
Ran it live. Pushed it hard. Watched closely.
The truth is: the trend that looks effortless in a thread usually behaves very differently inside a real build.
From Dewi's bank · Thread · CTA: None
Following trend · FT02
A
Arya Pradana @aryapradana_
···
Trying a new tool the day it drops is exciting.
It doesn't mean you've tested it.
It means a toy demo and a real feature are different things.
It is a 24-hour verdict.
How did we judge it?
Real task. Real load. Real result.
The truth is: the only honest review of a new tool is the one written after you put real work through it.
From Dewi's bank · Thread · CTA: None
Following trend · FT03
A
Arya Pradana @aryapradana_
···
Hype around a new trend is normal.
It doesn't mean it's worthless.
It means the hype isn't the part that lasts.
It is a quieter question.
What actually matters?
Running cost. Real upkeep. Long haul.
The truth is: the trend fades fast, the bill to run it six months later is what you actually live with.
From Dewi's bank · Tweet · CTA: None
Following trend · FT04
A
Arya Pradana @aryapradana_
···
Cloning a popular AI product is everywhere right now.
It doesn't mean the clones work.
It means most copy the surface and miss the core.
It is a teardown.
What do they miss?
The hard part. The real loop. The reason it works.
The truth is: copying what a product looks like is easy, copying why it works is the whole job.
From Dewi's bank · Thread · CTA: Teardown DM
Following trend · FT06
A
Arya Pradana @aryapradana_
···
Letting skills go stale between projects is easy.
It doesn't mean you're falling behind on purpose.
It means practice needs a deliberate slot.
It is a habit.
What keeps us sharp?
One pattern. Each month. Rebuilt fresh.
The truth is: a team stays sharp not from big projects, but from small reps nobody is paying them to do.
From Dewi's bank · Carousel · CTA: Link
Threads
Experimental 5 posts
Experimental · E01
A
aryapradana_· 2h···
We gave an AI agent write access to our own codebase for a week. Yes, things broke.
That's not a failure. It's how you find the edges before a client does.
We added three things after: tight limits, real reviews, easy undo.
It is a guardrail.
You don't trust an agent until you've watched it fail and caught it. Would you ever give one write access?
From Dewi's bank · Thread · CTA: None
Experimental · E02
A
aryapradana_· 2h···
I tested three ways to give an LLM app memory. Two were demos. One held up under a real workload.
That gap is the whole lesson. A demo and a workload are different animals.
It is a stress test.
The version that wows in a demo is rarely the one you'd actually ship. How are you handling memory right now?
From Dewi's bank · Thread · CTA: Soft DM
Experimental · E03
A
aryapradana_· 2h···
We deliberately ran one of our AI features at 10x traffic, just to see the bill. It was a little scary on purpose.
Better to find the cost cliff in a test than fall off it with a client on the line.
It is a dry run.
A feature you can't afford at scale isn't finished, it's a liability waiting. Do you load-test for cost, or just for speed?
From Dewi's bank · Thread · CTA: None
Experimental · E04
A
aryapradana_· 2h···
Friday experiment: can the design team ship a working prototype in a single afternoon, AI tooling only? We logged it live.
Not about cutting corners. The tools finally make a tight timebox realistic.
It is a sprint.
A deadline you can actually hit teaches more than a roadmap you can't. What's the fastest you've gone idea to prototype?
From Dewi's bank · Thread · CTA: None
Experimental · E05
A
aryapradana_· 2h···
Most AI demos work because the inputs were cherry-picked. So I broke ours on purpose, with ugly inputs and bad data.
That's the only honest test. Anyone can win with perfect inputs.
It is a magic trick until it isn't.
The demo that never breaks is the one nobody pushed hard enough. What's the worst input you've thrown at your own product?
From Dewi's bank · Tweet+media · CTA: None
Owned product 5 posts
Owned product · OP01
A
aryapradana_· 2h···
Magnific and ElevenLabs Flows got too expensive for how we work. So we built ESflows instead: open source, bring your own key, no subscription.
We're a design studio. We felt the cost and the lock-in firsthand, so we made the thing we wished existed.
It is ESflows.
A tool you can open and change beats one you can only rent. What's a tool you wish you owned outright?
From Dewi's bank · Thread · CTA: Link (demo)
Owned product · OP02
A
aryapradana_· 2h···
We kept hitting model limits and losing context every switch. Annoying enough that we built a memory layer that follows you between models.
That's ViMemory. There's a 49-second demo if you want to see it move.
It is ViMemory.
The model isn't your moat. The memory you keep between models is. How much context do you lose switching tools?
From Dewi's bank · Thread+video · CTA: Link
Owned product · OP03
A
aryapradana_· 2h···
People ask why ESflows is bring-your-own-key instead of a subscription. Here's the honest math on who actually pays for 'convenience.'
A subscription bills you whether you used it or not. BYOK means you pay for your own usage and nothing else.
It is the math.
A subscription charges you for the days you didn't even log in. Which model do you prefer, and why?
From Dewi's bank · Thread · CTA: None
Owned product · OP04
A
aryapradana_· 2h···
In ViMemory, every teammate's memory lives in its own isolated namespace. The isolation is enforced, not just promised.
That distinction is the whole point. A promise is marketing. A wall you can't get around is design.
It is isolation.
'We promise it's separate' is marketing. 'It physically can't mix' is design. Which one are your tools giving you?
From Dewi's bank · Thread · CTA: Link (install)
Owned product · OP06
A
aryapradana_· 2h···
Our own product's v1 was wrong in a few places. So we ran the same teardown on ourselves that we'd run on a client.
Not fun, but necessary. A studio that won't critique its own work has no business critiquing yours.
It is honesty.
The hardest teardown to run is the one on your own product. When did you last audit your own work?
From Dewi's bank · Thread · CTA: None
Previous client project 3 posts
Previous client project · PC02
A
aryapradana_· 2h···
On one project the client wanted AI inside a specific feature. We put it somewhere else entirely, and the build was better for it.
Our job isn't to place AI where it's requested. It's to find where it actually earns its spot.
It is judgment.
AI where the client asked is easy. AI where it works takes a real conversation. Ever moved a feature against the brief?
From Dewi's bank · Thread · CTA: None
Previous client project · PC04
A
aryapradana_· 2h···
We took a client from brief to shipped product without a single throwaway prototype. Not luck, process.
Most prototype waste is really just decisions people delayed. Make them earlier and the dead ends disappear.
It is a straight line.
Every throwaway prototype is a decision someone was scared to make sooner. How many do you burn per project?
From Dewi's bank · Thread · CTA: Soft DM
Previous client project · PC05
A
aryapradana_· 2h···
We inherited an AI-assisted codebase from a previous build. Here's what we found, and what we fixed.
The last team wasn't bad, they just optimized for fast over lasting. Common trap right now.
It is a rescue.
AI makes it easy to ship something that works today and quietly collapses next quarter. Inherited anything like this lately?
From Dewi's bank · Thread · CTA: Teardown DM
Concept use case 4 posts
Concept use case · CU02
A
aryapradana_· 2h···
For Cairn, we designed 'knowledge you can trust' as a verification lifecycle, not just a prettier search bar.
Confidence isn't accuracy. An answer you can't check is a guess in a nice font.
It is Cairn.
Trustworthy AI isn't a better search bar, it's a way to verify the answer. Would you trust an AI answer you couldn't trace?
From Dewi's bank · Thread · CTA: None
Concept use case · CU03
A
aryapradana_· 2h···
For Keel, a fintech concept, AI triages the approvals but never moves money on its own. A human always signs off.
The design question wasn't 'can AI do this' but 'where is it allowed to act.' Restraint is the feature.
It is Keel.
The smartest AI in fintech is the one that knows when to stop and ask. How much would you let AI decide alone?
From Dewi's bank · Thread · CTA: None
Concept use case · CU04
A
aryapradana_· 2h···
For Mneme, memory and runtime for AI agents, we designed provenance you can actually audit. You can see where every memory came from.
Most agent setups ask you to trust a black box. You can't audit what you can't see.
It is Mneme.
An agent you can't audit is just a confident stranger holding your keys. Would you let one run unmonitored?
From Dewi's bank · Thread · CTA: None
Concept use case · CU06
A
aryapradana_· 2h···
Most CRMs look identical. For Meridian we gave one warm paper and mono numerals, on purpose.
The sameness isn't inevitable. Everyone shipped the default and stopped pushing.
It is Meridian.
'It looks like every other tool' is a choice, and usually a lazy one. When did a default last go unquestioned on your team?
From Dewi's bank · Thread · CTA: Soft DM
Testimony client 2 posts
Testimony client · T02
A
aryapradana_· 2h···
Best feedback we got this month wasn't about how flashy something looked. It was that the build still works and still costs what we quoted.
That's the whole AI-native pitch in one line. Durability over dazzle.
It is the real pitch.
The best review isn't 'it's amazing,' it's 'it still works and still costs the same.' What feedback actually sticks with you?
From Dewi's bank · Tweet · CTA: None
Testimony client · T05
A
aryapradana_· 2h···
A client told me what mattered wasn't speed, it was that they trusted the build. That stuck with me.
Everyone sells fast. Almost nobody sells trust, even though it's the thing that actually closes.
It is trust.
Clients forget how fast you were long before they forget whether they could trust you. What makes you trust a builder?
From Dewi's bank · Thread · CTA: None
Call with client 5 posts
Call with client · CA01
A
aryapradana_· 2h···
A client asked us to 'add AI everywhere.' One question on that call saved them a rebuild: where's the actual pain?
Adding AII everywhere is easy to say yes to and expensive to undo. Restraint is the service.
It is restraint.
Sometimes the best thing you can sell a client is the AI feature you talked them out of. Ever said no to more scope?
From Dewi's bank · Thread · CTA: None
Call with client · CA02
A
aryapradana_· 2h···
On a scoping call: 'can you make it cheaper?' Here's the honest answer about AI cost we always give.
It's a fair question. They just can't see the running cost yet, only the build price.
It is a fair question.
Cutting the build price usually just moves the bill to a running cost nobody mentioned. How do you handle the cheaper question?
From Dewi's bank · Tweet · CTA: None
Call with client · CA03
A
aryapradana_· 2h···
Five questions we ask on every first call before we quote anything. They save both sides money.
Quoting fast feels helpful but usually isn't. A few questions up front beats a wrong number.
It is diligence.
A quote given before you understand the problem is just a number you'll both regret. What's your first question on a call?
From Dewi's bank · Thread · CTA: Soft DM
Call with client · CA04
A
aryapradana_· 2h···
A prospect came in burned by a cheap AI build. Guarded, and fair enough. Here's what they feared and what we showed them.
They weren't difficult, they'd learned a hard lesson. We met it with proof, not promises.
It is caution.
The client who's been burned once is the one who'll value real work the most. Ever won someone over after a bad experience?
From Dewi's bank · Thread · CTA: None
Call with client · CA05
A
aryapradana_· 2h···
Heard on a call: 'our last dev used AI to build it fast.' Fast isn't the flex people think it is.
Fast was the goal, lasting wasn't. That's how you end up with a rescue job six months later.
It is a warning sign.
Fast is the easy part. Surviving six months is the part nobody brags about. Would you rather have fast or durable?
From Dewi's bank · Tweet · CTA: None
Discussion with client 5 posts
Discussion with client · D01
A
aryapradana_· 2h···
A client debate today: should AI answer directly, or always show its sources? We picked sources.
A fast answer with no source is a confident guess. Sources are slower to read and far easier to trust.
It is sourcing.
A direct answer feels faster, but a sourced one is the only kind people come back to. Which would you ship?
From Dewi's bank · Thread · CTA: None
Discussion with client · D02
A
aryapradana_· 2h···
Spent an hour talking a client out of a feature they wanted. Best hour of the project.
Felt wrong in the moment. But protecting a build from bloat is the job, not a betrayal of it.
It is the job.
The features you quietly remove often do more for a build than the ones you proudly add. When did saying no pay off?
From Dewi's bank · Tweet · CTA: None
Discussion with client · D03
A
aryapradana_· 2h···
How we turn a messy client discussion into a one-page spec everyone signs off on.
The mess is fine. It just needs somewhere clear to land before anyone builds.
It is a spec.
Projects rarely fail from a hard problem. They fail because nobody wrote down what they agreed to. How do you capture a kickoff?
From Dewi's bank · Thread · CTA: Soft DM
Discussion with client · D04
A
aryapradana_· 2h···
Client: 'competitors have an AI chatbot.' Us: 'do your users actually want one?' That question reframed the whole project.
A competitor's roadmap had quietly become their brief. Worth catching before you build.
It is a reframe.
Copying a competitor's feature is the fastest way to inherit their mistakes. Ever built something just because a rival had it?
From Dewi's bank · Thread · CTA: None
Discussion with client · D05
A
aryapradana_· 2h···
The hardest part of a build is rarely the code. It's getting everyone to agree on what 'done' actually means.
Two people can picture totally different finish lines and not realize it until week six.
It is alignment.
Most projects stall not on hard problems but on a finish line nobody defined. How do you nail down 'done'?
From Dewi's bank · Thread · CTA: None
Revenue project 3 posts
Revenue project · R02
A
aryapradana_· 2h···
A feature we built a while back still runs on the cost we quoted. That's the metric I actually care about.
Most teams celebrate launch and never look at the bill again. The bill is where the truth lives.
It is the real metric.
Shipping a feature is easy. Shipping one that still costs what you quoted is the hard part. Do you track cost after launch?
From Dewi's bank · Tweet · CTA: None
Revenue project · R03
A
aryapradana_· 2h···
A paid build in one vertical became a reusable pattern we now ship faster. The fee was nice. The pattern was the real win.
Most teams treat every project as a one-off and rebuild from scratch each time.
It is compounding.
The value of a paid project isn't the fee, it's the reusable pattern you keep afterward. What are you reusing across builds?
From Dewi's bank · Thread · CTA: None
Revenue project · R05
A
aryapradana_· 2h···
What I learned pricing our first serious AI build: the number wasn't the scary part. Saying it out loud was.
The math was easy. The nerve to name it plainly and hold the line was the actual work.
It is nerve.
Pricing a build isn't a math problem, it's a confidence problem in a math costume. What's the hardest part of pricing for you?
From Dewi's bank · Thread · CTA: None
Project timeframe 4 posts
Project timeframe · PT01
A
aryapradana_· 2h···
Day one to shipped in days, not months. A real timeline of how an AI-native build actually goes.
No corners cut. The steps are just tighter when AI handles the grunt work and fewer people wait on each other.
It is a real timeline.
An AI-native build is faster not because you skip steps, but because nobody's stuck waiting. What's your idea-to-shipped time?
From Dewi's bank · Thread · CTA: Soft DM
Project timeframe · PT02
A
aryapradana_· 2h···
Our second build in a vertical is twice as fast as the first. Not luck, compounding.
The first build left a stack to stand on: known patterns, reused parts, fewer surprises.
It is compounding.
Your second build is faster because of what you refused to throw away from the first. What do you carry between projects?
From Dewi's bank · Thread · CTA: None
Project timeframe · PT03
A
aryapradana_· 2h···
Fast builds age badly when the speed came from shortcuts. Here's how we go fast without the debt.
Speed isn't the enemy. Speed borrowed against the future is. There's a real difference.
It is the difference.
Fast from systems lasts. Fast from shortcuts just postpones the rebuild. Where does your speed actually come from?
From Dewi's bank · Tweet · CTA: None
Project timeframe · PT05
A
aryapradana_· 2h···
How long should an AI feature take? Less than you think to demo. More than you think to make it survive.
The demo and the shipped feature run on different clocks. Most estimates only count the first one.
It is two timelines.
The demo takes an afternoon. The part that survives real users takes the rest of the month. Which one are you estimating?
From Dewi's bank · Tweet · CTA: None
Tool and stack 5 posts
Tool and stack · TS01
A
aryapradana_· 2h···
Six products, one stack: Vite, React, Tailwind v4. We never start from a blank file, on purpose.
The blank file feels disciplined but it's really just rebuilding the same foundation every time.
It is one stack.
The blank file isn't a fresh start, it's a tax you pay at the start of every project. What's your default stack?
From Dewi's bank · Thread · CTA: Link (porto)
Tool and stack · TS02
A
aryapradana_· 2h···
Our AI stack is picked for cost and maintainability over hype. Every piece has to earn its place.
Hype fades. The bill and the upkeep stay. So we choose for the boring stuff that lasts.
It is a stack choice.
The flashiest stack on launch day is rarely the one you can still afford a year later. How do you pick your tools?
From Dewi's bank · Thread · CTA: Link
Tool and stack · TS03
A
aryapradana_· 2h···
Hot take: your AI tool choice matters less than your fallback when it fails. Most teams have no fallback at all.
Everyone debates which model. Almost nobody plans for the day it's down or wrong.
It is your fallback.
The tool you pick matters less than what happens when it breaks. What's your plan when the model fails?
From Dewi's bank · Tweet · CTA: None
Tool and stack · TS04
A
aryapradana_· 2h···
We self-host the embedding model so nothing leaves the server and it's free. Here are the tradeoffs, honestly.
It's not the default for a reason, it takes setup. But 'nothing leaves, no fees' is often worth it.
It is a choice.
'Free and private' usually costs you some setup, and that's frequently a bargain. Would you self-host or pay to skip it?
From Dewi's bank · Thread · CTA: None
Tool and stack · TS06
A
aryapradana_· 2h···
The tools we tried and dropped this quarter, and the one that earned a permanent spot.
Dropping a tool isn't wasted time. Knowing what to cut is its own skill, and a underrated one.
It is honest churn.
The tools you quietly drop teach you more about your work than the ones you keep. What did you cut recently?
From Dewi's bank · Thread · CTA: None
Following trend 5 posts
Following trend · FT01
A
aryapradana_· 2h···
Everyone's shipping the same trend this week. We ran it on a real build instead of a demo. Here's what actually held.
Most people are reacting to a polished thread, not a live workload. The two rarely match.
It is a real test.
The trend that looks effortless in a thread behaves very differently inside a real build. Tried it on anything real yet?
From Dewi's bank · Thread · CTA: None
Following trend · FT02
A
aryapradana_· 2h···
Tried a trending tool the day it dropped, on a real feature, not a toy. Honest 24-hour verdict inside.
Day-one excitement isn't testing. A toy demo and a real feature behave nothing alike.
It is a 24-hour verdict.
The only honest review of a new tool is the one written after real work runs through it. How do you vet new tools?
From Dewi's bank · Thread · CTA: None
Following trend · FT03
A
aryapradana_· 2h···
The hype around this trend will pass. The question that outlasts it: what does it cost to run in six months?
The excitement is real but temporary. The running cost is what you actually live with.
It is a quieter question.
The trend fades fast. The bill to keep it running is what stays. Do you ask the six-month question before adopting?
From Dewi's bank · Tweet · CTA: None
Following trend · FT04
A
aryapradana_· 2h···
Everyone's cloning a popular AI product. Most miss the one thing that makes it actually work. Here's a teardown.
The clones copy the surface and skip the core. The visible part was never the hard part.
It is a teardown.
Copying what a product looks like is easy. Copying why it works is the whole job. What's a clone that missed the point?
From Dewi's bank · Thread · CTA: Teardown DM
Following trend · FT06
A
aryapradana_· 2h···
We rebuild one trending UI pattern each month just to keep the team sharp. No client, no deadline.
Skills go stale quietly between projects. Practice only happens if you give it a slot.
It is a habit.
A team stays sharp not from big projects, but from small reps nobody's paying them to do. How does your team practice?