There are two kinds of AI events. The first kind tells you that AI is going to change everything and leaves you with a deck of statistics and no idea what to do on Monday. The second kind walks you through someone's actual Tuesday morning. Ladies Building With AI, hosted by Marta at TomTom's Amsterdam office on a Thursday evening in late April, was very much the second kind.

Three speakers took the stage. Vidhu, a software engineer at TomTom, opened by establishing shared language: what Claude actually is, what it is not, and how engineers inside a major mapping company are using it without losing ownership of their work. Then Anna Sheronova, a founder and former Fortune 500 software engineer who moved to Amsterdam on a startup visa, walked the room through three tasks she has fully delegated to Claude Cowork. Last came Carolina Posma, founder of Postiv.ai, who killed three myths that are keeping non-technical founders and marketers from using Claude Code, and then built a branded landing page on a live screen to prove the point.

The room was a mix: startup founders, corporate employees exploring AI on the side, engineers, marketers, a few people who admitted mid-session they'd never opened Claude before. The questions were sharp. The takeaways were concrete. And by the time the pizza arrived, a significant number of attendees had opened their laptops, navigated to claude.ai/download, and started installing things.

This is a full recap of what they shared: the frameworks, the prompts they actually use, what it costs them, and the argument running underneath all of it that the most important AI skill in 2026 is one you probably already have.

Marta, host of Ladies Building With AI, gives a double thumbs-up next to the TomTom banner at the venue entrance, camera around her neck, visibly delighted.
Marta before the session began, with the energy of someone who knows the evening is going to land.

Part 01

What Claude actually is and why that framing matters

Vidhu opened with a question that might seem basic but was exactly the right place to start: "So, what actually is Claude?" In a room where technical fluency varied enormously, establishing shared language first was not condescension. It was the thing that made everything after it useful.

Vidhu, a software engineer at TomTom, presents a slide titled 'So, what actually is Claude?' explaining the difference between claude.ai, Claude Code, and IDE integrations.
Vidhu on stage at TomTom, explaining the Claude ecosystem from an engineer's perspective.

Claude, she explained, is an AI assistant built by Anthropic. It can read, write, reason, and code across a wide range of professional tasks. The key thing that confuses people (including technical people) is that there are three different products that share the same name and the same underlying model, but they work in fundamentally different ways and are designed for fundamentally different jobs. Conflating them is like confusing a spreadsheet with a database because both handle numbers.

The more operationally interesting part of Vidhu's talk was what she described from inside TomTom. Junior engineers are using Claude to generate code: drafts, boilerplate, first-pass solutions, but the engineering team owns the output. Architecture decisions, trade-off reasoning, security reviews, the choice of whether to ship: those stay human. Claude accelerates the rate at which code gets written. It does not replace the judgement that determines which code is worth writing.

That frame, AI as leverage, quietly ran through everything else that evening. Every speaker was delegating volume and holding onto judgement. The ones who do this well, Vidhu implied, are the ones who will compound. The ones who try to delegate judgement will keep finding themselves in the same place they started.


Part 02

It’s time to delegate: here’s exactly what

Slide reading 'It's time to delegate. 3 tasks Claude handles for me every day. A talk by Anna Sheronova.' in bold black type on white.
From Anna's deck

The opening slide. No preamble needed.

Anna Sheronova walked on stage and put up a single slide: It's time to delegate. 3 tasks Claude handles for me every day. It's a simple promise, and she kept it completely.

Anna spent eight years as a software engineer inside Fortune 500 companies before moving to the Netherlands on a startup visa two years ago. She now runs Nutripy, an AI-native product for gym owners focused on retention and lead conversion. On the side she runs an AI consulting practice, helping businesses adopt AI and automation. She has been a nomad: two years across Latin America with her partner and dog. She is, by disposition, someone who has thought hard about operating with high output and low overhead.

Slide titled 'My workday, before and after': Before shows a single circle labelled ME with one connection to Code. After shows the same circle connected to Meetings, Calls, Social Media, Lead Generation, SEO, and Marketing.
From Anna's deck

The before/after that set the frame for everything else. Same person. Radically different surface area.

The before-and-after slide hit the room harder than it looked. Before AI: Anna at the center, one line out to Code, because that was the one thing she had bandwidth for. After AI: Anna at the center, with lines running out to Meetings, Calls, Social Media, Lead Generation, SEO, Marketing. The size of the circle representing Anna didn't change. Her output surface area did. That's the operational reality behind all the abstractions about AI and productivity: it is not that you work fewer hours. It is that the same hours now cover an incomparably wider surface area when you stop trying to execute everything yourself.

"The job didn't get smaller. The team did. Claude is how I run a real company without a real headcount."

Anna Sheronova, Founder · Nutripy

Then she went specific: three tasks, the exact workflows, and what each one costs to run.

Daily delegation #1: Meeting prep, automated overnight

Slide titled 'Meeting prep: the old way' showing a checklist: Research the company (Google), Who is the person (LinkedIn), What we already talked about (Gmail), Past meeting notes (Notion), Check LinkedIn posts, Write an agenda.
From Anna's deck

The manual version. Every checkbox is a tab, a search, a copy-paste, and fifteen minutes you don't have.

The manual version of meeting prep is a known tax on every founder's morning: research the company on Google, find the person on LinkedIn, dig through Gmail for the thread history, track down the Notion page, read the past meeting notes, check their recent posts, then synthesise all of it into an agenda. If you have three meetings on a Thursday and each one takes thirty minutes to prepare for, you have just allocated ninety minutes before you've done a single thing that moves the business.

Slide titled 'Asking Claude' showing the Claude Cowork interface with a long, detailed prompt for meeting prep that pulls from Google Calendar, Gmail, Notion, and web search, producing structured Notion pages per meeting.
From Anna's deck

The delegated version: one prompt that pulls from four sources and produces structured briefs in Notion for every meeting on the calendar.

Anna's delegated version runs as a scheduled routine. Claude Cowork pulls today's calendar events from Google Calendar, checks the last ninety days of Gmail for thread history with each attendee, searches Notion for existing pages about them, and runs one web search per external attendee for recent context. For every meeting, it produces a structured Notion page, titled by date and meeting name, with an at-a-glance summary of two to four lines, attendee context with their last email and a one-line note, recent context in three to six bullets, a likely agenda, exactly three talking points, and optional watch-outs. Anything it cannot verify is marked inline as "Unverified." Nothing is fabricated. The brief runs overnight. Anna wakes up to a full Notion workspace of meeting briefs, and the first thing she reads in the morning is context she didn't have to gather herself.

The pattern behind the prompt

Daily delegation #2: Content trendwatching on Threads

Slide titled 'Before and after' showing: left side with screenshots of Threads posts saved manually to a phone library. Right side: the same information organised as a structured Notion database with topics, hook patterns, engagement data, and sources.
From Anna's deck

671 screenshots saved manually to a phone. The same intelligence, now structured and searchable in Notion, automatically.

Anna's second delegation is a content intelligence workflow. Her personal brand runs on Threads. The manual version of staying current on what is working: what hooks are landing, what formats are getting shared, which was an endless loop: scroll the feed, screenshot winning posts, save them to a camera roll that was, at last count, 671 items deep and entirely unsearchable. The screenshots accumulate. The insights don't.

The Cowork version runs as a workflow that scrolls the Threads feed for her topics, rates relevance against her voice and audience, tags each strong post by hook pattern: niche callout rally, geo-founder callout, authenticity theater, calendar progression motivator, before-path-after, and saves each keeper as one structured row in a Notion database. The output is a classified, searchable library of what's working, with enough pattern data to brief a writer or prompt a content session. Hooks shift every week. The loop no longer ends.

Daily delegation #3: The SEO blog that runs itself

Slide titled 'One skill picks up where the last one dropped off' showing a chained workflow: researcher (produces brief.md) → writer (produces en-draft.md) → formatter (produces en-final.md) → reviewer (verdict: PASS or REVISION) → publish (live URL). A dashed arrow shows revision loops back up the chain.
From Anna's deck

The full content pipeline. Each box is a Claude skill. Each arrow is an automatic handoff. State lives in a CSV that can pause and resume anytime.

The third workflow was the one that quieted the room the most completely. Anna runs a full SEO content engine for Nutripy. It doesn't work like a chatbot or a "write me a blog post" prompt. It works like a production pipeline.

There are five skills in the chain, each one a markdown file: researcher, writer, formatter, reviewer, publisher. A master orchestrator skill reads a CSV of article topics, finds the next one that needs attention, and kicks off the appropriate step. Each step runs as a fresh sub-agent, with isolated context and no bleed between tasks. The researcher pulls market data, verifies claims against primary sources, scans the competitive SERP, and produces a brief. The writer produces a draft in Nutripy's brand voice. The formatter applies the content structure. The reviewer runs a quality pass and returns either a PASS or a REVISION flag with specific notes. The publisher posts live. State for the whole pipeline lives in the CSV. Pause anytime, resume anytime, from any step.

7.1K
total search impressions in the first two weeks of the Nutripy blog running on this pipeline, with an average search position of 5 and 109 total clicks. The blog was not being run before this system existed. A human content team didn't write any of it.
A Google Search Console graph showing total impressions over 3 months from January to April 2026: flat near zero for two months, then a sharp spike to over 1,200 daily impressions starting mid-April.
From Anna's deck

The Search Console graph from two weeks after the pipeline launched. That steep line going up at the end is two weeks of automated publishing.

The whole pipeline runs overnight. Anna doesn't lose daytime token budget to it. She doesn't manage it in the way you manage a content team. She reviews the CSV to see status. If a piece comes back REVISION, she reads the reviewer's notes, makes a call, and re-queues it. That's the entire job. The pipeline is the team.

Anna Sheronova presents the pipeline diagram slide to the audience at TomTom, the workflow chart visible on the large screen behind her.
The pipeline slide. You could hear the audience doing the mental math about what this would mean for their own content production.
Slide titled '5 tips before you go': 1) Ask Claude to research the web first. 2) When a hard browser task works, save the recipe. 3) Run big token-heavy work outside your work hours. 4) Use Routines: they run on a remote machine. 5) Watch what permissions you grant connectors.
From Anna's deck

Anna’s closing tips: the things she wished someone had told her on day one.


Part 03

Claude Code for marketers: three myths, one live demo

Title slide reading '30-minute training · Claude Code for marketers. A no-panic, no-jargon, no-prerequisites crash course in the tool your technical friends won't shut up about. Speaker: Carolina Posma, Company: Postiv.ai.'
From Carolina's deck

The brief. Thirty minutes. No jargon. No panic. No prerequisites.

Carolina Posma, founder of Postiv.ai, runs the marketing and commercial side of her company. She uses Claude Code every single day. Her segment was billed as a thirty-minute crash course, and she opened, with considerable theatricality, by asking the room to raise their hand if reading the phrase "Claude Code" made them want to close their laptop.

Several hands went up. Then she spent the next twenty-five minutes destroying every reason those hands were raised.

Slide titled 'Wait, three different Claudes? Same brain underneath. Totally different jobs.' showing three columns: Claude (Think Together), Claude Cowork (Get It Done, No Code), Claude Code (Build Real Things) with descriptions of each.
From Carolina's deck

The clearest single-slide explanation of the three Claudes and their distinct jobs.

Before the myths, a moment of orientation: the same comparison Anna used, now in Carolina's framing. Three Claudes, one model, three completely different applications.

Think together

Claude (chat)

The web interface. Draft, research, brainstorm, summarise. You type, it replies, you act on what it says. Your daily thinking partner. No autonomy.

Get it done · no code

Claude Cowork

Autonomous agent. Reads your email, scrapes competitors, writes the report, formats it, hands it back done. Connects to Gmail, Calendar, Notion, Slack, Stripe.

Build real things

Claude Code

Writes and runs real code on your files. Landing pages, dashboards, scrapers, custom tools: shippable artifacts. Desktop app with no terminal needed.

Carolina's rule: roughly 80 percent of a non-technical founder's Claude time belongs in Cowork, roughly 20 percent in Code. The 20 percent is for when you need something built, not just done. They share the same underlying model, the same context folders, and increasingly, the same skill files.

Slide titled '80% Cowork. 20% Code. Here's the split.' Two columns: Claude Cowork (80%) for email/calendar/research/reports/scheduled tasks, and Claude Code (20%) for landing pages/dashboards/custom tools/batch file work/API integrations.
From Carolina's deck

The split that makes the toolkit less overwhelming: most of your work is Cowork, occasionally you need Code.

Myth #1: Claude Code is only for developers

The myth
"Claude Code is only for developers."
The reality
It turns plain English into software. If you can brief a junior, a freelancer, : clearly, with context, with an outcome in mind: you already have the skill that matters. Claude Code just needs a clear request. Writing clear requests is literally your job.
Slide titled Myth One: they say Claude Code is only for developers. Reality: It turns plain English into software. If you can brief a junior, a freelancer, or a new hire, you already have the hard skill.
From Carolina's deck

Myth one. The reframe: briefing is the skill, and you already do it.

Myth #2: I'm going to break something

The myth
"I'm going to break something catastrophic."
The reality
Claude Code asks permission before every destructive action. Work in a dedicated practice folder : that is your safety net. Read the yes/no prompts before clicking. That's the whole protocol. There's nothing else.
Slide titled Myth Two: they say they will break something. Reality: It asks permission before anything risky.
From Carolina's deck

Myth two. The permission model means the floor is higher than most people assume.

Myth #3: I need to learn to code first

The myth
"I need to learn to code before I can use this."
The reality
You need to learn to describe. Briefs, positioning, tone, audience, constraints, outcomes. The skill set you already use every day when working with agencies, contractors, or freelancers is the skill set that now translates directly into working software.
Slide titled Myth Three: they say they need to learn to code first. Reality: You need to learn to describe, which is the literal job of a marketer.
From Carolina's deck

Myth three, and the one that hit hardest in a room of marketers and founders.

The context folder: writing your instructions once

Slide titled 'Meet your context folder' showing a project folder on a laptop with subfolders: context/ (brand.md, audience.md, about-me.md), skills/ (landing-page.md, brief-ads.md), references/ (brand-guidelines.pdf, competitor screenshots, past work). Below: 'Instead of a wall of text every time, you just say: Write the hero copy.'
From Carolina's deck

The context folder: the architectural shift that turns one-off prompts into repeatable workflows.

Carolina's most practically useful section was about the underlying architecture that makes both Cowork and Code genuinely different from chatting with Claude on the web. The context folder is a project directory on your laptop that Claude reads permanently after the first setup. Inside it live three types of file: context (who you are, who you write for, what sounds like you), skills (one markdown file per repeatable workflow), and references (brand guidelines, competitor screenshots, past work, anything Claude should know).

Once this folder exists, you stop explaining yourself in every session. Instead of opening a chat with a paragraph about your brand voice, your audience, and your product positioning, you just say, "Write the hero copy." Claude reads the folder, finds everything it needs, and applies all of it automatically. The long prompt is dead. The folder killed it.

Slide titled 'Prompt skill: the difference: A skill is a recipe for your workflow.' Showing a before (a wall of instructional text retyped in every chat) and an after (the slash command /morning-report, with all context automatically applied).
From Carolina's deck

Before and after. Write the skill file once. Call it in two words forever after.

"The gap between the people who try this and the people who don't is about to get very, very big."

Carolina Posma, Founder · Postiv.ai

The live demo

Carolina Posma presents from a podium, pointing at a large projected screen showing Claude Code building a branded landing page for a single-origin Amsterdam coffee roaster in real time.
A landing page building itself on a 4-metre screen while Carolina narrated. The room stopped typing.

The last ten minutes were a live build. Carolina opened a fresh Claude Code project, dropped in a one-page brand brief for a fictional single-origin coffee brand, the kind of brief a marketer would write for any client, and asked Claude Code to build a landing page. Within sixty seconds, code was writing itself on the four-metre projected screen. She iterated twice: once to adjust the hero headline, once to add an email capture form. The result was deployed before her segment ended. The audience, which had largely stopped taking notes and started watching very closely, was quiet in the way that rooms get quiet when something lands exactly as promised.

The audience watches Carolina's live coding session projected on screen, showing Claude Code's agent panel alongside the HTML file being edited in real time.
The live build on screen. Code editing itself in real time while Carolina talked the audience through what was happening and why.

Part 04

Five things to take from this into this week

Women seated on tiered cushioned benches inside TomTom, most of them with laptops open, attending the Ladies Building With AI event.
The audience mid-session. Laptops open. Notes getting taken. A few people already installing things.

Sitting in the room, the thesis that emerged across all three talks was the same one framed three different ways for three different roles. Engineers, founders, and marketers all arrived at the same operational conclusion: AI is not a tool you use. It is a teammate you brief. The quality of your delegation determines the quality of your output. And the skill that matters most right now is not prompt engineering. It is the skill of describing what you want clearly enough that someone else, human or AI, can go away and do it.

For founders specifically, the implications are unusually concrete. The talks collectively amounted to a usable playbook: not "here are all the ways AI could theoretically help your business," but "here are the specific things you can delegate, the exact format of the workflow, and the real numbers from someone who has already done it." That's rare enough to be worth translating into explicit next steps.

01

Audit the repeat work first. Before building anything, list the five tasks you repeat every week that follow a consistent structure. Those are your first skill files: meetings, reports, content research, inbox triage, and anything similar.

02

Build a context folder, not a habit of re-prompting. The compounding starts the moment you write down your brand voice, your audience, and your workflow once. From that point, every conversation with Claude carries the context automatically.

03

Start with Cowork, not Code. Eighty percent of the work that founders and marketers do is connectable without writing a single line of code. Get your Cowork stack running (Gmail, Calendar, Notion) before you build anything with Code.

04

Run expensive pipelines overnight. Deep research, SEO pipelines, anything heavy: schedule them as routines outside your working hours. This keeps your daytime token budget for chat and iteration, where the work is more sensitive to latency.

05

Delegate volume, hold onto judgement. The pattern from all three speakers was identical: AI accelerates the rate of execution. The architecture, the trade-offs, the final call on what ships: those stay with you. That's not a limitation. That's your job.

06

Minimum viable skill: one workflow this week. Pick one meeting you prep for regularly. Write a skill file for it. Run it once. Fix what's wrong. That's the milestone: one working automation and the confidence to build the next one.


Part 05

The community that makes the session stick

The half-hour after the formal programme ended was worth staying for. The pizza came out, people stopped taking notes, and the actual exchange started. The conversations were almost aggressively practical: exactly how Anna's meeting brief Notion schema is structured, whether to set Gmail permissions to read-only or read-and-write, what happens if you accidentally trigger a deep-research pipeline three times in one session, which Claude model runs which step in a chained pipeline (Opus for the thinking, Sonnet where it's cheaper and fast enough for a pure API call).

A large group of event attendees networking in TomTom's office break room after the Ladies Building With AI session, drinks in hand, animated conversation visible throughout.
The post-session in the TomTom break room. The most useful part of any event like this is the thirty minutes when the slides go dark.
Three attendees sit in low chairs by a floor-to-ceiling window overlooking the IJ river in Amsterdam, deep in conversation with coffee and notes after the event.
By the IJ after the session. The view behind them was almost distractingly good.

Nobody at these tables was speculating about artificial general intelligence or the future of white-collar work. They were doing the immediate, practical calculation: which of my recurring tasks match this pattern, what's the first connector I need to install, how do I structure a skill file if I don't know where to start? That is the right set of questions to be asking, and the fact that a room full of people was asking them is the sign that this particular moment is something other than hype.

Marta runs these sessions monthly in Amsterdam, with a Slack community for the weeks in between and a Luma calendar for upcoming events. The session directly following this one was a beginner-level walkthrough: a slower, question-driven version for people who left the TomTom evening interested but slightly overwhelmed. That's exactly the right programming sequence: one session to open the aperture, another to close it on something actionable.

If you weren't there and are reading this: the community is at ladiesbuildwithai.com. The slides from Anna and Carolina are available to members. Carolina is also recording the full Claude Code training. Connect with her on LinkedIn and include the message "Claude Training" to receive it.


AI Workflows Claude Claude Cowork Claude Code Women in AI Amsterdam Founders Delegation Event Recap SEO Automation Content Strategy