Lock-in is a choice you make for convenience
Brazil made payments free. I almost added the cost back without noticing.
Final LATAM build for @claudeai for Small Business: Pix with the processor as a replaceable part. Swap vendors with a settings change. The agent never notices
Customizing @claudeai Cowork Small Business plugin for LATAM: payments, fiscal docs, WhatsApp.
Let AI read everything; gate what it can fire.
Founder, KOOKY AI Exchange.
In Brazil, a small business is a WhatsApp number. Storefront, order desk, support line, all one chat thread.
So building the WhatsApp piece of Claude for Small Business, the real question was: should the AI hold the customer channel at all?
https://t.co/O8eEsN0TPA
Senior managers and directors are leaving people leadership to come back as High-Impact Individual Contributors with their own agent stack.
Charging team-equivalent salaries. No management overhead.
The math changed. The career path is following.
I gave my AI agents a better memory. They got worse.
More memory is not more intelligence. Past a point it is noise with good search.What actually fixed it, and the one layer every "company brain" post skips:
https://t.co/G8J0WYqUZV
Week 2 wrap: the second LATAM adaptation for Claude for Small Business shipped. Not the way I'd scoped it on Monday.
The deletion was the deliverable. Full writeup:
https://t.co/NAjJeaPFI6
2 of 4 down. WhatsApp next.
W1 mapped the gap: QuickBooks → @ContaAzul . Today the ContaAzul skill starts. The load-bearing piece is NFe issuance, the part /close-month actually needs to close a Brazilian month. Same shape as W1: contract first, sandbox-verified, ships in days.
Every sale a Brazilian business makes needs an NFe. That's Nota Fiscal Eletrônica, the mandatory electronic invoice. /close-month in *@claudeai for Small Business* can't close a Brazilian month without it. That's why @ContaAzul is my next build.
MercadoPago looks like one product. Underneath it's two. Cards and Pix run on different internal systems that don't speak the same language. The skill hides the split behind one charge call. Skills give the artifact a portable contract any agent can read.
With the right scaffolds already compounding, week 1 shipped what a 3-person team (eng, writer, PM) delivers in a one-week sprint.
Production BR payments skill. Published article. Three cookbook recipes.
Calendar time: one working day. ~15:1 leverage.
LATAM patterns for @claudeai , open-sourced as a public fork of @AnthropicAI 's cookbook.
Recipe 01 (MercadoPago) is live. Recipes 02–04 ship with the next three builds.
Selective PRs upstream coming W4–W5.
https://t.co/KaCHgDOgfc
Started writing the MercadoPago connector for @claudeai for Small Business.
Skills it unlocks on day one: /cash-flow-snapshot, /close-month, /weekly-brief — anything that reads payments.
Ships next week. Brazilian SMBs can finally point Claude at the rail they actually use.
@AnthropicAI put the Connector layer behind MCP. That's the choice that makes the LATAM gap fixable.
A regional operator can ship MercadoPago on the same surface, same UX, same approval gates as the official Stripe connector. No greenlight needed.
That's rare in this space.
Pix moved more volume than credit and debit cards combined in Brazil last year.
It's not "an alternative payment method." It's the payment rail.
A small business plugin without Pix doesn't process payments. That's not a gap. That's a missing limb.
Mapped the 12 Connectors in Claude for Small Business against the BR small business stack.
Stripe → MercadoPago.
PayPal → Pix.
QuickBooks → Conta Azul.
Slack → WhatsApp Business.
Zero overlap with what BR SMBs actually use. Wrote the thesis here: https://t.co/TLAecuMiUj
Read all 31 Skills in Claude for Small Business this weekend.
They're well-scoped. /cash-flow-snapshot, /close-month, /crm-cleanup — each one a real workflow, not a demo.
The Skills layer is the strongest part of the launch. The Connectors are where the gap sits.
KOOKY OS is not starting with auto-posting.
It is starting with the loop:
task -> draft -> Humanizer -> WATSON review -> Renato approval -> manual publish -> validation
Publishing is easy to automate later.
Trust is harder.
So we are proving the operating loop first.