Carl Nordell - AI & AutomationBerlin · 2026 · Let's talk

Lovingly handprompted by Carl Nordell ↓

AI & Automation · Marketing Operations

I never set out to run an AI department. Here we are.

I'm a copywriter and content strategist. No CS degree, no team - just a nerd who kept building the tools he wished existed. Many in production, used by real teams.

Let's talk → Berlin-based

✓ shipped
≈20
internal AI tools, brief → deploy
✓ in production
10
client workspaces, automated
✓ unattended
5
run on a schedule, no babysitting
Keep scrolling. The toolbox is where it gets good.
01

How I work

i.

Remove the friction

The big creative problems aren't where the hours go. They leak out the small stuff you redo every week - the report, the reformat, the copy-paste. I turn that into software first. The boring parts get the love.

ii.

Evidence over vibes

Every output points back to its data. I build the fact-checking in, not bolted on - a citation per claim, an honest "weak signal" over a confident lie. If it can't show its work, it doesn't ship. Not from me.

iii.

Augment, never replace

Cutting the busywork is the whole point - it hands people their hours back for the parts only humans do well: judgement, taste, strategy. I build tools that make a team sharper. Not smaller.

iv.

Don't just make a website

AI put real problem-solving on tap, so rebuilding the same old thing, just faster, misses the point. A tool only earns its place if it does something that wasn't possible before, not an old workflow with a chatbot bolted on.

02

The toolbox

A working tool says more than a CV bullet. Twenty-ish of them - I stopped counting - each built solo, from brief to deployment. The little diagrams animate as you scroll. The same instinct runs off the clock, too - I keep my own life on a self-built OS.

✓ shipped & running

Automation · Agentic pipeline

Roster

✓ In production

New faces, every Monday.

web pages → curated profiles → Monday digest Diagram: pipeline from scanned web pages to curated profiles to the Monday digest.

Scouting bookable directors and photographers is slow, manual work. Every Monday Roster scans 100+ pages - production-company rosters and industry press, runs them through a cascade of AI gates, scores each for brand-fit, and curates the keepers into profiles. A digest lands with the team at 9am. Unattended, on a server.

PythonHaiku + SonnetPerplexityNotion API
100+
pages scanned every week, hands-off
8
people on the Monday 9am digest
Reporting · Creative briefing

Briefer

✓ In production

Never a blank page.

spreadsheet → ideas Diagram: pipeline from a content spreadsheet to ready creative ideas.

Turning a content plan into a creative briefing is manual, repetitive, per-client work. Briefer automates it: plan in, finished on-brand briefing out. It runs on a schedule across ten client workspaces - so a strategist starts from a draft, not a blank page.

Python 3.12Claude Sonnet 4.6rclonecron · systemd
10
client workspaces, on a schedule
0
blank pages - start from a draft
The one I'm proudest of
Research · Custom tooling

Probe

✓ In production

Just one more question.

Born from two frustrations: briefs too vague to start from, and waiting on meetings for feedback. It's a Claude Code skill - from inside a project it reads the work, then builds and deploys a custom interview agent with a bit of personality. I send a link and a password; it chats people through the questions, digs where it matters, and hands back a clean summary - one charming stand-in, talking to everyone at once. From there, adding a feature, fixing a problem, or kicking off a new build is often as simple as pasting those answers straight back into Claude Code.

Claude Code skillFastAPIClaude Sonnet 4.6Cloudflaresystemd
2min
from prompt to a shareable interview
0
meetings needed to get a usable brief
built by a Claude Code skill ↓ Diagram: a chat interview with Pip, summarized into insights for the build.
Audience intelligence

Composite

✓ Client-proven

Almost a person.

culture & data → living persona Diagram: pipeline from cultural and data sources to a living audience persona.

Built after a podcast on how the notorious Lily's Garden ads were engineered from a psychological map of their audience - Gilmore Girls dynamics and all. Composite builds that map from real data: the hard statistics, the media a group actually consumes, and how they talk online about what matters to them - into a portrait deep enough to interview.

PythonHaiku→Sonnet→OpusYouTube APIPerplexity
19
personas delivered, with evidence
1.5k
posts harvested per build
Personal · LifeOS

LifeOS Assistant

✓ In production

Runs my whole week.

Hevy in → AI → new program, back in ↺ Diagram: workout data flowing in, a new training program flowing back out.

The same instinct, off the clock - a Telegram bot wired to my own life. The part I love: it reads my Hevy workout history, spots where I'm progressing, and writes next week's program straight back into the app. It also runs my mornings - sleep, calendar, markets, family. Real daily use, not a demo.

PythonClaude HaikuWhispersystemd
3
input modes - voice, photo, text
2
scheduled daily/weekly broadcasts
Personal · Family OS

Family Hub

✓ Daily use

The household, on autopilot.

A Svelte touchscreen that runs the household - calendar, chores, lights, the robot vacuum, all from one panel. My favourite part is a little economy: quests earn stars, and stars buy screen time. It even got my son asking to do his chores. The boring stuff mostly runs itself now.

SvelteTouchscreen kioskHome AssistantHue · Roborock
0
reminders to switch the screen off
1
panel for the whole house
chores → stars → screen time Diagram: a quest list turning chores into stars that trade for screen time.
I notice the small, annoying, repeatable things - and quietly turn them into software.
- the whole job, basically
Twenty tools in. Still going. Let's talk →
03

Two closer looks

No meeting · one command

The interview becomes the next build

A research sprint takes a week. Probe takes a command. It's a Claude Code skill - you point it at a project and it reads the actual files, not some brief you hand it. Then it writes and deploys a custom interview agent: a link, a password, and off it goes. It runs the whole interview itself, asks its own follow-ups when an answer gets interesting, and sends back a clean summary.

And here is the part I like: research and building kind of stop being two separate jobs. You paste that summary back into Claude Code, and the next thing - a feature, a fix, sometimes a whole new tool - often starts from those exact words people gave you.

↳ It writes its own follow-up questions, and the answers come back as the next build - still a bit strange to watch, honestly.
›claude /probe
1ReadsReads the projectthe real files, not a brief
2DeploysBuilds the interviewcustom agent · link + password
3ListensHands back a summarythe actual words, cleaned up
4BuildsPaste back → it buildsfeature, fix, or next tool
↻loops
↺loops back to step 1
answers in → next build out · one skill
0
briefings needed
2min
prompt → shareable interview
Audience modelling with receipts

A persona built from evidence, deep enough to interview

Most audience personas are basically a workshop guess with nothing behind them. Composite builds one from what a group actually does - the hard numbers, the stuff they watch and listen to, and most of all how they talk online when nobody is asking them anything. Around 1,500 posts per build. Then three models split the work: Haiku pulls things out, Sonnet finds the patterns, Opus writes the actual portrait.

Every trait comes with its source - an actual quote, an actual number - so anyone can poke at a claim and follow it back to where it came from. That's the thing that got the strategists on board, actually: a map you can interrogate, not a vibe.

↳ Raw culture → signals → one person you could basically interview - every trait with its receipts.
YouTubeForumsLyricsStatsReviews
~1,500 posts harvested
▼Haiku extracts
clustered into signals
Wants low effortmost cited
Time-poor
Hates admin
▼Sonnet finds patterns · Opus writes
OutputPersona 01 / 19
Lena, 34“I want it easy - not another job.”
“wants low-effort wins”↳ 3 threads + a review
~1,500 posts · 3 models · 1 auditable portrait
19
personas delivered, with evidence
1.5k
posts harvested per build
This is the kind of thing I build. Let's talk →
04

What I'm good at

I'm at my best owning something end to end - with the autonomy and the proper tools to actually do it. In practice, that means:

i.
I architect AI workflows

I design and ship agentic pipelines end-to-end - multi-step, multi-model, running unattended on servers. Same orchestration as n8n or Zapier, just without the visual layer.

ii.
I run reporting at scale

I build a self-grading monthly intelligence report that checks its own past calls against what actually happened. Reporting frameworks are home turf.

iii.
I make the tools talk to each other

I've wired Claude, Perplexity, Apify, Notion, Slack, Matomo, Telegram and Google Drive into integrated pipelines that run unattended.

iv.
I've driven AI adoption org-wide

I run enablement sessions, built a tool a non-technical colleague rated on par with a commercial product, and wrote the org's AI-rollout plan. Adoption is the half I care about most.

05

The whole range

I'm not the best engineer in the room, or the best strategist, or the best copywriter. But I understand all three - and that's the rare bit. The tools get built by someone who has actually done the work they're meant to serve.

Words · craftSystems · code →
Content & strategyCopywriting, brand voice, storytelling, positioning, presentations
Audience & researchEvidence-based persona modelling, audience insight, interviews
AI engineeringPrompt & agent engineering, Python, LLM APIs, RAG, multi-model pipelines, evals
Ship & adoptDeployment (Linux, Cloudflare, cron), scraping/ETL, trust layers, Figma handoff, training & AI adoption
Svenska
native
English
fluent
Deutsch
understand most · working on speaking it