Jagrit Chawla

Dumping high value alfa on AI and its applications.

Founder: Marktech Studios

[Jagrit Chawla @jagu_builds]

𝐘𝐨𝐮𝐫 𝐀𝐠𝐞𝐧𝐭 𝐖𝐫𝐢𝐭𝐞𝐬 𝐎𝐥𝐝 𝐖𝐞𝐛 𝐂𝐨𝐝𝐞

If you build landing pages with Claude or Cursor, you have seen it. It reaches for an old pattern or a bloated library instead of the browser API that already does the job.

Chrome and Edge shipped a fix. Modern Web Guidance is an agent skill plus a CLI that steers coding agents toward modern, secure, high-performance web APIs. Audit a legacy page, set Baseline targets, fix Core Web Vitals.

Repo: https://t.co/6OfeTXlKVJ
Docs: https://t.co/ATxeJ77MhC

My POV. A slow landing page burns ad money on every click. Make the agent build it right the first time.

[Jagrit Chawla @jagu_builds]

Daily AI updates | Day 12

𝐂𝐥𝐚𝐢𝐦 𝐘𝐨𝐮𝐫 𝐁𝐫𝐚𝐧𝐝 𝐈𝐧𝐬𝐢𝐝𝐞 𝐂𝐡𝐚𝐭𝐆𝐏𝐓

OpenAI shipped something D2C founders should care about more than any new model this week. On Nov 10 it launched Businesses in ChatGPT, in beta.

Translation: your brand can now have a verified profile inside ChatGPT, and people comparing products can talk to an agent that represents you.

Here is the part that matters. The old flow was simple. Someone asks ChatGPT for a recommendation, it names your competitor, done. You never saw it happen and you could not influence it.

That changes now. In Business Hub you can claim your business, verify it, and push your own product feed with live pricing and availability. ChatGPT stops guessing and starts quoting your data.

How to claim it:

1. Sign in to ChatGPT and open Business Hub from the sidebar.
2. Search your business name and select your listing.
3. Verify with phone, business email, website domain or Shopify.
4. Complete identity verification through Persona. That is what unlocks the blue check.
5. Update your description, logo, category, delivery and return policy.
6. Add a product feed so ChatGPT can show real prices.

Once verified you get Insights. Mentions, website clicks, and the discovery topics where your brand shows up. Aggregated over 28 days, no individual user data. Still, it is the first honest look at how AI describes you.

Eligible US ecommerce and online service businesses can also build a Business Agent. A shopper clicks your profile in a response and asks your agent questions directly. Separate chat, clearly labeled, powered by the info you provide.

It is free at launch. No Ads account or Enterprise agreement needed to claim.

The catch. Business Agents are US only for now, and local businesses can only claim an existing listing. Claiming a profile itself works in every market where ChatGPT runs.

My POV. Discovery is shifting from Google search to AI answers, and the brands that claim their profile early will own the answer when someone asks which one to buy. Do it this week, before your competitor does.

Sources:
https://t.co/CAgiHEb2QW
https://t.co/WjuGk1AIcg

[Jagrit Chawla @jagu_builds]

𝐁𝐥𝐞𝐧𝐝𝐞𝐫 𝐉𝐮𝐬𝐭 𝐆𝐨𝐭 𝐑𝐞𝐚𝐥-𝐖𝐨𝐫𝐥𝐝 𝟑𝐃

Blender 5.3 is out, and the headliner is native 3D Gaussian Splat support.

Import splats from PLY and SPZ files as point clouds, edit them with Geometry Nodes, then relight and render in Workbench, EEVEE or Cycles. No plugins needed.

Why it matters. Scan a real product or a space with your phone, turn it into a splat, drop it into Blender, and ship a 3D ad or a web hero. That used to need a specialist.

Blender is free.
https://t.co/JWBHe7Y9DH

My POV. The line between needing a 3D artist and doing it ourselves just moved.

[Jagrit Chawla @jagu_builds]

𝐓𝐡𝐞 𝐕𝐢𝐝𝐞𝐨 𝐒𝐤𝐢𝐥𝐥 𝐒𝐭𝐚𝐜𝐤

Video editors keep saying AI can't do real motion. Here's the stack that already does.

- hyperframes, write HTML and get a video. New /reverse-any-video skill. https://t.co/tnnlP0oiYb
- remotion skills, Claude cuts real videos in React. https://t.co/JTdun9c8XC
- video-shotcraft, 124 cinematic shot recipes for product promos. https://t.co/cR676R99IJ
- blender mcp, Claude drives Blender and builds whole scenes. https://t.co/pPZi7GLVyV
- OpenMontage, a full agentic video studio. https://t.co/zR1sAIT3PU
- awesome-claude-video, 475 motion demos, each with its original prompt. https://t.co/4qZqpnSJ9F

That last repo is the one to study. Every project ships with the prompt that made it, so you can recreate it and then push your own version.

The editor's job isn't disappearing. It's moving up to taste and direction. The people who learn the stack will direct 5x the output.

[Jagrit Chawla @jagu_builds]

𝐎𝐰𝐧 𝐭𝐡𝐞 𝐑𝐨𝐨𝐦, 𝐍𝐨𝐭 𝐭𝐡𝐞 𝐅𝐞𝐞𝐝

Sam Parr on My First Million breaks down Moda Communities. A guy in Sioux Falls started it in 2018 because he had no clean place to buy and sell watches.

So he made a Facebook group. It grew to 600,000 members across 20+ communities. Watches under $10k, watches over, cars, gold. Over a dozen full-time staff. $5 to $10M a year.

The part brands miss is the gate. Dealers get a vetting interview. Then another member has to vouch for them. Then they pay $1k to $10k a year to list.

That's not a group. That's a paywalled trust network.

Now map it to D2C. Brands pour budget into rented reach, then watch CAC climb every quarter. Nobody owns the buyer list. Nobody owns the room. Wire the two together and the math flips: recurring membership revenue, high-intent buyers showing up weekly, a filter that keeps the place clean.

You don't need 600,000 people. You need 200 vetted ones paying every month for access you control.

The feed is rented. The room is yours.

[Jagrit Chawla @jagu_builds]

𝐄𝐢𝐠𝐡𝐭 𝐑𝐞𝐩𝐨𝐬 𝐅𝐨𝐫 𝐀𝐈 𝐀𝐝 𝐂𝐫𝐞𝐚𝐭𝐢𝐯𝐞

Someone catalogued the ad creative stack for Claude Code and Codex. All free, all open source.

- marketingskills, 50+ skills, start with customer-research and ad-creative. https://t.co/kDr7TYHeg3
- claude-ads, paid media across 12 ad platforms plus a creative strategist agent. https://t.co/gJvy53MExf
- marketing-skills, pulls competitor ads from Meta Ad Library and finds the gaps. https://t.co/vF1RBM3LSo
- OpenMontage, agentic video production, 12 pipelines, 700+ skills. https://t.co/zR1sAIT3PU
- last30days-skill, research across Reddit, X and YouTube before you write. https://t.co/V95M4s8BNY
- goose-skills, growth and GTM skills. https://t.co/Vw7QZDxzKQ
- advertising-skills, funnels and copy systems. https://t.co/j7XQSF5c4m
- ai-slop-detector, 120 patterns that make copy read like AI. https://t.co/EnBXOetXYd

The prompt that ties it together: give your agent a niche, ask it to find active ads launched in the last 30 days, then group them by hook and angle with source links.

Research first, then create. Most people skip step one and wonder why the ads flop.

[Jagrit Chawla @jagu_builds]

𝐀𝐝𝐨𝐛𝐞 𝐉𝐮𝐬𝐭 𝐎𝐩𝐞𝐧𝐞𝐝 𝐂𝐫𝐞𝐚𝐭𝐢𝐯𝐞 𝐂𝐥𝐨𝐮𝐝 𝐓𝐨 𝐀𝐠𝐞𝐧𝐭𝐬

Yesterday at Adobe MAX in Miami, Adobe previewed local MCP integrations for Photoshop, Illustrator, Premiere and After Effects.

Translation: your Claude or ChatGPT agent can open your actual project files and edit them. Not an export, not a copy. The native file, with layers, vectors, bins and sequences still editable.

It's a private beta landing in early 2027. But the direction is set.

Adobe also pushed the Creative Agent across Creative Cloud, added always-on agents that keep working in the background, and turned Firefly into a studio that pulls in models from Google, Kling, Runway, Luma and ElevenLabs in one place.

Here's what most people will miss.

The hard part of ad production was never the idea. It was the handoff. Brief to designer, designer to editor, editor to the person resizing 14 versions for stories, reels and feed.

That handoff is what eats the margin on a retainer.

Now picture this. You remove a logo from one campaign asset, then tell your agent to cascade it across the whole set. It pulls the brief from Google Docs, works inside Premiere, and drops the finished cut in Slack. Adobe demoed exactly this.

For a studio running D2C accounts, that's not a nice-to-have. That's the gap between 6 clients and 15 with the same team.

The catch: Adobe says it's still deciding what the MCP endpoints should expose. Beta is early 2027. So don't rebuild your workflow today.

Do this instead. Write down every repetitive step in your current production chain. The 12 resize jobs, the version swaps, the color tweaks. That list is your automation roadmap for the day the beta opens.

The agencies that win the next 2 years won't be the ones with the best designers. They'll be the ones who mapped their boring work first.

Source: https://t.co/9Rf442YBxU

[Jagrit Chawla @jagu_builds]

𝐓𝐚𝐬𝐭𝐞 𝐈𝐬 𝐭𝐡𝐞 𝐌𝐨𝐚𝐭 𝐍𝐨𝐰

AI made execution cheap. Taste is what's left.

Feed yours daily.

https://t.co/u7UFP1pjjB, design inspiration from X
https://t.co/WE3PMrTD7R, 800+ 3D and motion website prompts you paste into Claude
https://t.co/jYaVAY4qC4, mobile UI animations
https://t.co/MLhl2vi01s, the fonts behind real brands
https://t.co/Dx7155nmqw, award-winning sites

Anyone can generate a page this week. Almost nobody can tell you why it's good.

Build the eye first. Tools second.

[Jagrit Chawla @jagu_builds]

𝐕𝐢𝐞𝐰𝐬 𝐀𝐫𝐞 𝐍𝐨𝐭 𝐂𝐨𝐧𝐯𝐞𝐫𝐬𝐢𝐨𝐧𝐬

Hector Serrano on Lenny's Podcast breaks down the organic UGC model he ran at Wispr Flow.

Here's how it works. You group a bunch of creators. They post about your product on their own Instagram and TikTok pages. You pay them on CPM, about $2 per 1,000 views, tapering as the numbers climb.

It did 500 million views in 2.5 months. On paper that's a monster. It converted almost nothing.

The lesson for D2C is blunt. Reach and revenue are two different engines. A $2 CPM looks great in a deck until you check whether anyone bought.

That's the shift Serrano made after. He moved to performance UGC: 50 creators, ads killed once they hit $100K in spend, 5,000 lifetime ads with 800 live at any time. Creators get a base, then a cut of what the ad actually drives.

Track creators by conversion, not views. Kill fatigue fast. Vanity reach is the most expensive thing you can buy.

Episode here: https://t.co/1yypiG0Wtb

[Jagrit Chawla @jagu_builds]

𝐎𝐧𝐞 𝐏𝐞𝐫𝐬𝐨𝐧 𝐒𝐡𝐢𝐩𝐩𝐢𝐧𝐠 𝐚 𝐓𝐞𝐚𝐦'𝐬 𝐎𝐮𝐭𝐩𝐮𝐭

Nobody's losing work to AI. They're losing it to one person running the right repos.

https://t.co/J0PLzMFwRL builds full-stack apps with AI. shadcn/ui gives clean components in minutes. Magic MCP generates React components inside your coding flow. Playwright MCP lets the agent click through a site and catch its own UI bugs. Webstudio is an open-source visual builder.

That's a designer, a frontend dev and a QA pass in one loop.

https://t.co/Lr0raWRR4R
https://t.co/dgiHfrGpHa
https://t.co/lqBDOr3nkQ
https://t.co/KpjXN37Z3I
https://t.co/qts95qgXcl

Learn the loop, not the tool.

[Jagrit Chawla @jagu_builds]

𝐓𝐡𝐞 𝐅𝐚𝐬𝐭𝐞𝐬𝐭 𝐖𝐚𝐲 𝐭𝐨 𝐖𝐫𝐢𝐭𝐞 𝐄𝐜𝐨𝐦 𝐀𝐝𝐬 𝐖𝐢𝐭𝐡 𝐀𝐈

Most people open ChatGPT and ask for a winning ad. You get the average of every ad ever made.

Do it backwards. Pull what's working first.

1. https://t.co/l0wm5KeuTx, find your niche, filter ads from the last 30 days.
2. Connect it to Claude through MCP.
3. Paste: "Find active ads in [NICHE] from the last 30 days. Analyze hooks, audiences, angles and pain points. Group similar ads, include source links."

Then give Claude your product URL and ask for 3 underused angles, each a 30-second UGC script.

Research beats generation.

[Jagrit Chawla @jagu_builds]

𝐄𝐯𝐞𝐫𝐲 𝐀𝐈 𝐁𝐮𝐢𝐥𝐭 𝐀𝐩𝐩 𝐋𝐨𝐨𝐤𝐬 𝐓𝐡𝐞 𝐒𝐚𝐦𝐞 𝐍𝐨𝐰

You can spot one from the thumbnail. Same centered card. Same gradient button. Same shadcn defaults, untouched.

It's not a tooling problem. It's that nobody changes the defaults.

Four libraries that get you off that path:

→ Bencho: https://t.co/Wg7xFehEoy
→ BoardUI: https://t.co/baHt6Iic6F
→ Transitions: https://t.co/RDMn54PGQj
→ Beautiful UI: https://t.co/wQ4gHy84oH

Design is the cheapest way to stop looking like everyone else. Nobody screenshots the app that looked like the last one.

[Jagrit Chawla @jagu_builds]

𝐘𝐨𝐮𝐫 𝐀𝐈 𝐕𝐢𝐝𝐞𝐨𝐬 𝐋𝐨𝐨𝐤 𝐁𝐨𝐫𝐢𝐧𝐠. 𝐓𝐡𝐞 𝐏𝐫𝐨𝐛𝐥𝐞𝐦 𝐈𝐬 𝐘𝐨𝐮𝐫 𝐒𝐭𝐚𝐜𝐤

Everyone's posting "one prompt" AI videos. Nobody shows what's behind them.

A model on its own gives you the same grey, floaty clips. The output you keep seeing is code plus tools the agent actually runs. Skip the tools and you get slop. Wire them up and you get something clients pay for.

Here's the free stack almost nobody mentions.

Rendering and motion
Remotion renders video with React. The Remotion Skills repo teaches your agent to build video the right way instead of guessing. HyperFrames turns plain HTML into video, built for agents in the loop.
https://t.co/JTdun9c8XC
https://t.co/fkrQGCGL4e
https://t.co/tnnlP0oiYb

3D and scenes
Blender handles real models, rigs and renders. Blender MCP lets Claude drive the scene directly, no tab switching.
https://t.co/MzQusu3ttP
https://t.co/pPZi7GLVyV

Voice and captions
faster-whisper turns any voiceover into captions. Chatterbox generates free voiceovers from plain text.
https://t.co/3g0U2kUUK7
https://t.co/tIzsBNV27F

Motion and polish
GSAP moves text and UI without the jank. Three.js puts 3D in the browser on your GPU.
https://t.co/oCH8p8XZKv
https://t.co/P0NuKJ7P6s

Self-check and export
Chrome DevTools MCP lets the agent watch its own render and fix mistakes before you see them. FFmpeg converts and exports the final MP4.
https://t.co/5WIkwAPj7N
https://t.co/jG76jir7lj

Assets
yt-dlp grabs reference clips and audio. ComfyUI runs your own image and video generation. Impeccable reviews and polishes the design.
https://t.co/DcLBBvR6dn
https://t.co/tqQbPNv7ju
https://t.co/7mBsUNkHG2

If you only build one loop, build this: Remotion renders, faster-whisper captions, Chrome DevTools MCP checks, FFmpeg exports. Everything else is an upgrade.

The honest part. None of this is one prompt. It's a pipeline. The people getting good output wired these tools together and let the model orchestrate the steps.

My take: the "one prompt" caption is marketing. The real skill is knowing which tools exist and how they connect. Start with Remotion Skills and one MCP server this week. That alone puts you ahead of most people posting today.

[Jagrit Chawla @jagu_builds]

𝐒𝐭𝐨𝐩 𝐏𝐫𝐨𝐦𝐩𝐭𝐢𝐧𝐠 𝐓𝐡𝐞 𝐒𝐚𝐦𝐞 𝐓𝐡𝐢𝐧𝐠 𝐄𝐯𝐞𝐫𝐲 𝐃𝐚𝐲

If you retype the same instructions every week, you don’t need a better prompt. You need a Skill.

Anthropic’s Agent Skills are folders with instructions, scripts and resources. Claude loads one only when the task matches, so it doesn’t eat your context.

For a D2C studio: one skill for product descriptions in brand voice, one for ad variants, one for landing page QA.

Build once. Reuse forever.

Docs: https://t.co/4SPVgQUsc2

My take: your best prompt is the one you never have to write again.

[Jagrit Chawla @jagu_builds]

𝐁𝐮𝐢𝐥𝐝 𝟓𝟎 𝐀𝐠𝐞𝐧𝐭𝐬, 𝐊𝐞𝐞𝐩 𝟓

Seema Amble, a16z partner, on the a16z Podcast with Ivan Makarov and Figma CFO Praveer Melwani.

She talked to a growth-stage CFO whose team built 40 to 50 agents in a month or two. Her guess: 90% are dead already. Nobody opens them.

Her point is that the graveyard isn't the failure. The muscle is. Once someone builds one agent, they stop doing tedious work by hand and start asking why it isn't automated yet.

Now translate it to D2C. Most brands run one AI pilot, it underperforms, they quit and call AI overhyped. Wrong read.

The teams that win build 20 throwaway things. A review miner that pulls objections from 300 Amazon reviews. An ad angle generator that dies after two weeks.

Let agents be disposable. If every AI workflow has to survive forever to be worth building, you'll never start.

Build five, keep one. Volume beats polish when the first draft costs nothing.

Episode here: https://t.co/QoeS8BYCWm

[Jagrit Chawla @jagu_builds]

𝐘𝐨𝐮𝐫 𝐋𝐚𝐧𝐝𝐢𝐧𝐠 𝐏𝐚𝐠𝐞 𝐅𝐚𝐢𝐥𝐬 𝐓𝐡𝐞 𝐒𝐚𝐦𝐞 𝐖𝐚𝐲 𝐄𝐯𝐞𝐫𝐲𝐨𝐧𝐞 𝐄𝐥𝐬𝐞’𝐬 𝐃𝐨𝐞𝐬

Someone ran 420 landing pages through the same 36 checks. One page per site, mostly indie and SaaS. The same failures kept showing up.

Worst offenders, straight from the study:

Lists features, never says what you get: 91%
Doesn’t say who it’s for: 89%
Proof sits far from the button: 89%
Headline names the product, not a result: 84%
No social proof at all: 80%
Button says "Get started" instead of what you get: 73%

Read that again. 9 out of 10 pages describe the product and forget the customer. The feature list is doing the job the headline should be doing.

Every one of these is a message problem, not a traffic problem. You can fix all five in an afternoon, no designer needed.

If you run a D2C brand, this is your cheapest conversion win this month. Not a new ad account. Not a redesign. Copy and layout.

The fix, in order:

1. Headline names the result, not the brand. "Lose 4 kg in 30 days" beats "Welcome to FitCo."
2. Say who it’s for in the first screen. "For busy dads over 35." The reader should feel caught.
3. Put proof next to the button. Reviews, order counts, before and after. Right where the thumb lands.
4. Button says what they get. "Start my free week" beats "Get started."
5. One line on why you over the alternative.

Selling supplements? The headline says the outcome and the time frame. Not the ingredient list.

The study comes from PageCritic. Free audit, 36 checks, a conversion score in about a minute: https://t.co/WaCkHHi60h. The full rubric is public at https://t.co/MGvYcwgE0h. 25 of the 36 checks run in code in a real browser, on desktop and phone.

Run your own page first. Then run a competitor’s. The gap is almost always copy, not product.

My take: most brands pour money into traffic to fix a page problem. Fix the page first. It’s free and it compounds.

[Jagrit Chawla @jagu_builds]

𝐘𝐨𝐮𝐫 𝐋𝐚𝐧𝐝𝐢𝐧𝐠 𝐏𝐚𝐠𝐞 𝐋𝐨𝐨𝐤𝐬 𝐆𝐞𝐧𝐞𝐫𝐢𝐜

Because you start from a blank prompt. Start from a reference instead.

https://t.co/Ra3QRD3fdV, real landing page inspiration, filterable.
https://t.co/WE3PMrTD7R, 800+ 3D and motion website prompts you paste into Claude.
https://t.co/YBIsj1aIp5, prompts for ad visuals and social.
https://t.co/jYaVAY4qC4, UI and motion animation inspiration.
https://t.co/u7UFP1pjjB, web, branding and 3D reference.

Workflow: pick 3 references, write down the exact section order and the type scale, then prompt Claude with that structure, not "make me a good website."

My take: generic output is a research shortcut, not a model limit.

[Jagrit Chawla @jagu_builds]

𝐘𝐨𝐮𝐫 𝐄𝐝𝐢𝐭 𝐓𝐚𝐛𝐥𝐞 𝐉𝐮𝐬𝐭 𝐆𝐨𝐭 𝐂𝐡𝐞𝐚𝐩𝐞𝐫

Higgsfield dropped Miracle on Nov 7. It vectorizes any video into layers, so you edit one object, recolour one product, move one element, in real time, no re-render. The masking and rotoscoping that ate a full day is gone. Private beta, MCP integration coming.

Pair it with two open tools.

Remotion (https://t.co/lg3TexF9T4) builds motion graphics with React, data driven and version controlled, renders in bulk. 43.5k stars at https://t.co/fkrQGCGL4e.

Wan 2.2 generates video locally on your own GPU, so no per-clip fee.

My take: UGC volume stops being a budget problem and becomes a taste problem.

[Jagrit Chawla @jagu_builds]

𝐓𝐡𝐞 𝐌𝐞𝐜𝐡𝐚𝐧𝐢𝐜𝐚𝐥 𝐖𝐨𝐫𝐤 𝐈𝐬 𝐔𝐧𝐝𝐞𝐫 𝐓𝐡𝐫𝐞𝐚𝐭

DHH on Lex Fridman Podcast #501, saying the quiet part out loud about AI and craft. His line: you'll have a hard time coping with the new reality if the only thing you loved about the job was the mechanical part, building what someone else told you to build.

That's not just code. It's the whole marketing stack too.

The D2C brands still paying a retainer for someone to resize assets, schedule posts, tag products and pull weekly numbers, that layer is gone this year. Agents run it for close to zero.

What's left is the part a model can't own. Knowing which customer to chase. Which claim is actually true. Which 3 seconds decide the scroll.

The teams I watch winning are cutting the execution layer and buying judgment instead. Fewer hands on the keyboard, sharper taste above it.

Hire for the call, not the typing. The mechanical middle is where D2C margins go to die.

[Jagrit Chawla @jagu_builds]

𝐓𝐡𝐞 𝐆𝐮𝐲 𝐖𝐡𝐨 𝐁𝐮𝐢𝐥𝐭 𝐂𝐥𝐚𝐮𝐝𝐞 𝐂𝐨𝐝𝐞 𝐑𝐮𝐧𝐬 𝐈𝐭 𝐁𝐨𝐫𝐢𝐧𝐠

Boris Cherny built Claude Code. His own setup is "surprisingly vanilla." No 500 line prompt. No plugin stack. Just a handful of habits you can copy tonight.

Five Claudes in parallel. He runs 5 instances at once, each in its own git checkout or worktree of the same repo. Tabs numbered 1 to 5. Notifications tell him when one needs input. The single biggest unlock, per the Claude Code team.

git worktree add .claude/worktrees/my-worktree origin/main

One CLAUDE.md, checked into git, edited by the whole team. Every time Claude does something wrong, it goes in the file so it never happens twice. Their rule: always use bun, not npm. Yours might be: never invent a stat, always cite the source.

Plan mode first. shift + tab, twice. Refine the plan with Claude, then switch to auto-accept and let it one-shot. "A good plan is important to avoid issues down the line."

Slash commands for anything he does daily. /commit-push-pr commits, pushes and opens a PR. Commands live in .claude/commands/ and are shared with the team.

Subagents as automations: code-simplifier, verify-app, build-validator, code-architect. Each is one markdown file.

A PostToolUse hook auto-formats every file Claude writes, so CI never fails on style. And /permissions pre-allows safe commands instead of running dangerously-skip-permissions.

The most important tip, his words: give Claude a way to verify its work. "If Claude has that feedback loop, it will 2-3x the quality of the final result." For UI work he lets Claude open a browser, test the change, and iterate until it passes.

All of it is documented at https://t.co/AzysiskVOe, 142+ tips.

Now translate this to a D2C marketing studio. Same shape of work: research, first drafts, grading. Swap worktrees for campaign folders. Swap verify-app for a checker that opens the landing page and confirms the form fires. Swap the CLAUDE.md rule for your brand voice, banned words and offer.

My take: the teams that scale next year will not add headcount. They will wire up the feedback loop. Taste on the last mile, file-based machinery on everything before it.