Jens Marions Conway · Stockholm · November 2026
How I use AI, day to day

Intent,
shipped.

I lead digital workplace and AI adoption in my day job. Outside it, I use AI to close the gap between deciding something and having it done: emails sent, sites shipped, a renovation kept in order, training logged and coached. This page shows how that works, including what doesn't.

A typical eveningElapsed
  1. Reply to the installer: Thursday works, 8 amCosmo · drafted, approved, sentseconds
  2. Ship a new practice set to pluggClaude · commit pushed, site redeployedminutes
  3. Log today: legs, 22 min, felt heavyTracker · logged, next load adjustedseconds
  4. Did the return match what I sent back?Checked line by line against the confirmationminutes
The setup

The context is mine. The models are interchangeable.

I don't treat any one assistant as the centre. The part that matters, what the AI knows about me and my projects, lives in a folder I own. Models plug into it and get swapped as they improve. Actions run through connections I've approved.

Context

PersonalOS: notes on me, my projects and how I like to work, kept in my own Google Drive.

Google DriveModel-agnosticI own it

Models

Whichever is best at the job this month. I compare them on real work, not benchmarks.

ChatGPTClaudeCodex

Actions

The places where things actually happen, each behind a connection I approved.

GmailCalendarGitHubCloudflareScheduled tasks
Cosmo

Email at the speed of intent.

Cosmo is my assistant for email and calendar. I say what I want in one line. It finds the thread, pulls the context, writes the reply in my voice and sends it once I say yes. A reply that used to take me a few minutes to write is now out in seconds. One email saves little. Dozens a week add up to real hours.

Minutes
Seconds
IntentOne line, typed or spoken.
ContextCosmo reads the thread and the calendar.
DraftWritten the way I'd write it.
Approve, sendOne yes. Then it's gone.
What it adds up to
30
4 min
30 s
7.6
hours back per month
Example inputs. Move the sliders to your own numbers.
Nothing leaves without my yesNo payments, everConsequential changes need approvalDrafts stay editable
The workbench

What I build, what runs on a schedule, and what I just get help with.

hunches.dev

Live

My lab for experiments. I describe a site and Claude creates the repo, wires GitHub Actions and Cloudflare, and puts it on its own subdomain. A stats page finds every project through the Cloudflare API.

hunches.dev

plugg

Live

Maths test practice for school, used on an iPad. Results save on the device, so no logins. New subjects go in as needed.

Private, on hunches.dev

Training tracker

Live

My own tracker for a four-day split of 20–25 minute sessions. I log energy and difficulty; the AI makes the progression call for next time.

Built for me

Home map

Next

A floor plan of our home that controls the lights, for phone and tablet. Lights first, more later.

Planned for hunches.dev

Tribera

Paused

A SaaS platform for Swedish housing associations (BRFs). Got as far as a full design system overhaul.

Parked by choice

Ballflow

Paused

A pickup basketball scoreboard that keeps score from live video: a stream server, FFmpeg and a vision model calling the baskets.

Parked by choice

Cosmo

Running

Inbox and calendar triage, follow-ups, and replies sent in seconds after one approval.

Continuous

Market briefings

Running

A pre-market and a post-close briefing for my own investing. Long-term signal over the day's noise.

Twice each trading day

Sweden in Context

Running

Weekly commentary on the Swedish economy, as a newsletter and an audio script, timed to catch the morning's data releases.

Fridays

The AI Stack, Swedish Edition

Prototype

A weekly scorecard on the AI trade, through a transparent basket and a Swedish investor's eyes. Fully automated; no column for me to write.

Saturdays, once it's ready

Today's session

Running

The day's workout arrives on the day, not a week in advance. Legs, pull, push, athletic.

Monday to Thursday

Digital workplace and AI adoption

Ongoing

At work: strategy drafts, stakeholder narratives, adoption measurement, and pressure-testing ideas before they reach a room.

Day job

Kitchen renovation

Wrapping up

Orders, appliance specs, contractors and returns, each reconciled against the retailer's own confirmation. Bathroom is next.

2026

Decisions, big and small

Ongoing

Comparing family EVs for later, judging a bike deal, moving a phone between platforms, school and travel logistics.

Whenever it comes up

Play

Ongoing

Riddles one at a time, bedtime stories, sports recaps without spoilers.

With the kids
Habits

How I work with it.

Questions before answers

For anything that matters I ask the AI to interview me first. A recommendation built on a guess is worse than no recommendation.

Critique over agreement

I ask to be told where I'm wrong. An assistant that agrees with everything is a liability.

Check it against the source

A neat table isn't done. The kitchen return was finished when the retailer's confirmation matched my count, line by line.

Delegate, keep the keys

The AI can draft, deploy and schedule. Anything that leaves my hands or can't be undone waits for my yes.

The trade-off

Starting is cheap now. Finishing still isn't.

AI makes it easy to start things, so I start a lot of them. My building time is evenings, after the kids are asleep, and that doesn't grow. So every project gets one of three answers, and I try to give it early.

Ship ithunches.dev, plugg, the tracker.
Put it on a scheduleBriefings, newsletters, workouts. They run without me.
Pause it on purposeTribera and Ballflow. Parked, not abandoned.