Prateek Gupta

AVP, Product Design @DreamMoney_IN - Simplifying money decisions for India 🇮🇳 Writer. Visual artist. Marathon Runner

[Prateek Gupta @PrateekG]

@Sanjeev87351 Noted and reported the issue Sanjeev ji

[Prateek Gupta @PrateekG]

🚨Hiring Alert 🚨

Building great products is one thing.
Building the intelligence behind them where the real magic happens.

Dream Money is hiring a Director of Data Science & Analytics, and I’m genuinely excited about what we are building.

This isn’t just another fintech role. You’ll be:

🔥 Designing and pushing a new fintech app that transforms personal wealth management for India

🔥 Leading AI, experimentation, and analytics that power breakthrough strategy

🔥 Working alongside brilliant minds who obsess over turning data into impact

If you’ve scaled ML systems, thrive in fast-paced execution, and want to lead teams through high-stakes innovation, let’s talk.

This role is for📍 Mumbai location.

The future of personal finance is being built right now. Come be part of it.

[Prateek Gupta @PrateekG]

As a design leader, you are no longer the maker.
Your craft is now your judgment and the systems, processes you build to make that judgment consistent, visible and teachable.

[Prateek Gupta @PrateekG]

@animeshprasadd @mytruetee Hey Animesh, there has been some logistical issue going on from amazon side. Order it from our website - https://t.co/CBYWfsahyY

[Prateek Gupta @PrateekG]

@animeshprasadd @mytruetee Okay, let me dig into this. Can you give me some reference number or email that you might have received. Something to track.

[Prateek Gupta @PrateekG]

@animeshprasadd @mytruetee Animesh, can you send me your order ID. Will look into this. DM me please

[Prateek Gupta @PrateekG]

Me right now 🤣

[Prateek Gupta @PrateekG]

Noticing this across companies I work with: the budget conversation shifted from “how many people can we hire” to “how high can we run our API bill.”

And companies are comfortable with it because an inflated API cost still beats the all-in cost of headcount.

Smaller, leaner teams. higher token budgets. That’s the new playbook.
This is why the job market looks the way it does.

[Prateek Gupta @PrateekG]

6) For designers specifically: ask Claude to first flow out the changes before touching or making changes to your prototype or code. Verify it if everything looks right, then implement. Kills the iteration loop which will save you a lot of tokens.

[Prateek Gupta @PrateekG]

Been optimizing Claude usage as now we are on paid subscription 🙈 and noticed some patterns worth sharing:

1) Stop burning tokens on setup use Projects or Memory so you’re not re-explaining context every chat. Every conversation you start without context burns 3-5 setup message.

2) Edit your original prompt instead of sending new ones (cuts token use by 80% over multiple rounds). Claude rereads your entire conversation history every single time. When it miss a mark, click edit on your original message and regenerate.

3) Start fresh chats every 15-20 messages or you’ll pay 250x more per request. Your first message cost around 200 tokens. By message 30, that same cost 250x. Copy Claude summary, start fresh chat.

4) Use right model for right thing. Use Haiku for quick tasks, brainstorming, grammar, formatting. Use Sonnet for content writing, analysis, coding. Use Opus for deep research and hard logic. Haiku all day for simple task free up 50-70% of your budget for the work that actually needs the bigger models.

5) Turn off features you’re not using web search, research mode, connectors, extended thinking all add overhead. Extended thinking also should be off by default and only switch it on when your first attempt wasn’t good enough.

Try this and see if your token efficiency becomes better.

[Prateek Gupta @PrateekG]

Moved to claude enterprise recently and noticed token consumption ramped up noticeably.

Was it the shift from team to enterprise, or has something changed in how claude processes requests?

Curious if others saw the same spike or if it’s something on our end.

[Prateek Gupta @PrateekG]

Spent thousands of tokens trying to automate our design process with claude.

We got the system, the structure, the speed but that final 5%? The craft. The polish. The handoff-ready refinement.

Has anyone solved this? or is this just the ceiling right now?

Genuinely curious if i’m missing something or if others have hit the same wall.

[Prateek Gupta @PrateekG]

A lot of people have been asking how I structured Claude skills for my team and how to start from scratch. So putting it all here.

The problem: Everyone prompting differently, getting inconsistent results, re-explaining the same context every conversation. We needed a shared knowledge layer behind Claude so the whole team gets consistent output without.

Here's how to do it step by step.

Step 1: Audit what you keep repeating
Before building anything, spend sometime noticing. What context do you paste into Claude every time? What questions does your team keep re-explaining? What workflows follow the same pattern every time? Write these down. This becomes your build list.

Step 2: Decide: markdown file or skill?
This is the most important distinction.
A markdown file in your project instructions is always loaded. Use this for foundational context you want present all the time, org, project or product background, brand voice, principles etc.
A skill only activates when the request matches its trigger description. Use these for knowledge or workflows that are only relevant for specific types of requests.

Step 3: Categorise your skills
Once something qualifies as a skill, it falls into one of three types:
Reference: stable knowledge. Domain context, specifications, guidelines, standards.
Capability: repeatable workflows. Generate, audit, review, parse, define.
Connector: orientation for external tools. Which files exist where, how things are organised. Claude might have tool access but no idea where anything lives without these.

Step 4: Write trigger descriptions carefully
Every skill has a description that tells Claude when to activate. This makes or breaks the system.
Write them like you're onboarding a new hire. Specific scenarios, keywords, and example requests - not vague categories. Include what the skill is NOT for. Bad triggers mean skills fire when they shouldn't or miss when they should.

Step 5: Share with your team
On paid plans, skills are shared at the team level. Update once, everyone benefits on their next conversation.
If you're on free plans, you can still collaborate. Keep your skill files in a shared Google Drive folder or a GitHub repo. Team members copy them into their own Customize section. Not as seamless, but it works, one person maintains, everyone pulls the latest version.
Either way, treat it like shared infrastructure. We maintain a dedicated Slack channel where anyone can request new skills, ask questions, or flag when something's producing bad output.

Step 6: Start small, iterate
Don't build everything at once.
First - create markdown files for your most foundational context.
Next - build one or two reference skills for your most asked questions.
Then - add a capability skill for one workflow your team repeats weekly.
Later - add connector skills only when you integrate external tools.
Finally - bundle into plugins once a domain's skill set stabilises.

[Prateek Gupta @PrateekG]

Your AI is getting dumber mid-conversation? It's probably not the AI. It's your context window.

Here's what I learned building 28 Claude Skills.

Every time you ask Claude to do something, it's reading: your prompt + skill instructions + all dependency files + any code or prototype in the conversation. After a few iterations, that context fills up. And when it does, Claude starts forgetting your early instructions, gives weaker outputs, or just hits a wall.

The fix isn't better prompts. It's leaner files and smarter structure.

What worked for me:

→ Skill files under 500-800 words each. Tables over paragraphs. Rules at the top, examples at the bottom.

→ Selective dependency loading — don't read the entire reference file, read only the sections needed for that specific task.

→ One task per prompt. Stop asking AI to do five things at once.

→ When a code file or prototype has been through multiple iterations in one chat, start fresh. Copy the final version into a new conversation and iterate from there. The old conversation's context is bloated with every previous version.

→ Remove anything from skill files that's written for humans, not for AI. If it doesn't change Claude's output, it's just burning tokens.

The rule: every extra word in your system gets multiplied by every prompt you run. Trim once, save thousands of tokens daily.

[Prateek Gupta @PrateekG]

Yes exactly, each capability skill mentions which reference skill it depends on. Before it acts, it read those reference first for context.

E.g. when you ask for design audit, it will first read design principles, component specs etc. Now when one principle changes or a new component gets added, all I need to do it update one skill file. Rest all skill file will pick from it.

[Prateek Gupta @PrateekG]

We've been using Claude as a design collaborator. But the more we used it, the more we realized throwing context into every chat is exhausting and inconsistent.

So I built a system (thank you Katherine, this is based on the structure you mentioned)

28 interconnected Claude Skills, organized into 3 layers designed so the entire product, design, and tech team works from the same knowledge base.

Here's how it's structured:

𝗟𝗮𝘆𝗲𝗿 𝟭: 𝗥𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲 (14 skills)
This is the foundation, the stuff Claude needs to "know" before it does anything. Design principles, component specs, content strategy, motion rules, accessibility standards, interaction flows, platform conventions, and even fintech-specific domain knowledge like legal, trust & security UX.

Think of it as the team's collective brain, always available.

𝗟𝗮𝘆𝗲𝗿 𝟮: 𝗖𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆 (13 skills)
This is where Claude actually does things reads PRDs, defines user flows, creates design tickets, runs usability heuristics, generates prototypes, audits designs against our system, and even handles compliance checks.

Some of these are standalone: they work on their own.
Others are reference-dependent: they pull from Layer 1 to stay consistent.
And a few are Figma-dependent: they connect directly to our Figma files for reviews and syncing.

The magic is in the dependency. When we update a reference skill (say, our component specs), every capability that depends on it automatically stays in sync.

𝗟𝗮𝘆𝗲𝗿 𝟯: 𝗖𝗼𝗻𝗻𝗲𝗰𝘁𝗼𝗿 (1 skill)
The bridge between Claude and our tools. Right now it's a Figma MCP wrapper that gives Claude specific context when reading or writing to Figma.

The whole point of this system is simple anyone on the team can use any skill and get a consistent, context-aware output without re-explaining who we are, how we design, or what our rules are.

Every time someone improves a skill, everyone benefits. It compounds.

Still early. Still refining.

[Prateek Gupta @PrateekG]

@JakeMetzDev @ridd_design Thanks @JakeMetzDev

[Prateek Gupta @PrateekG]

🚨 Hiring Alert 🚨

I am hiring a Motion Designer & Illustrator (2-5 yrs) at Dream Money, Mumbai.

We're making investing simple for millions of Indians. Need someone who can make it look and feel that way too.

If you think in keyframes, love illustration, and want to build a fintech app from the ground up, then this is for you.

[Prateek Gupta @PrateekG]

@tt__pixel Thanks. Happy to help @tt__pixel

[Prateek Gupta @PrateekG]

Getting first drafts using claude are now trivial. But that last 20%, the nuance, the edge cases, the finesse, the things that feel right costs more effort and time.

I feel the more context and prompt i add, it breaks more.