Dylan Harbour · Technology Director · Ringier SA

Claude Better.

Months of learning in minutes

Mindset, resources and advice from agentic harness rollouts across teams.

?

Before we start · be honest with yourself

Where are you today?

If we mapped the global developer community’s AI fluency

LOW FLUENCY HIGH FLUENCY how many people → AI fluency → 0% 50% 100% LOW FLUENCY HIGH FLUENCY how many people → AI fluency → 0% 50% 100% Aware of change coming ChatGPT as an assistant Vibe coding 100% agent-written code Solid validation & feedback strategies Trained a frontier model

Where humans fit in

What humans continue to own.

Taste

What good looks like. Quality, style, the standard.

Judgement

The decisions that matter. The risks and trade-offs we accept.

Accountability

A human backs every outcome. “The agent did it” is never an excuse.

Strategy

Who we are and why we exist. Where we go, and why.

Relationship

The connections that are uniquely ours.

Trust

Given by people to people. Earned, never delegated to a machine.

As execution gets cheap, this is the signal that gets scarcer, and worth more. The agent can do the task; only you can own whether it is any good.

The fear underneath all of this

As your empowerment goes up, your fear goes down.

The job fear. The too-late fear.

Fears reduce as you build confidence, skills and understanding.

FEAR EMPOWERMENT the more capable, the less afraid

How long does it take to flip?

ONE MONTH.

Mindset shifts with one month of dedicated grind. Empowerment takes 1-3 months. Learning never stops.

Before the tips, the reason for them

This moment in time is too critical to leave in someone else’s hands.

So the hour goes where it counts

Things we won’t cover today.

The basics of Claude

What an LLM is, prompting 101, getting started. You use this daily, and earlier sessions have it covered.

A feature tour, or my setup

Features change monthly, and copying my config is not copying my results.

Claude vs Codex

I run both and the skills transfer cleanly. It is how you drive, not what you drive.

The deep dives

Context engineering and RPI get named, not taught. RPI is the next talk in this room.

None of it is worthless. It is just not where the value is now. A fool with a tool is still a fool, so we spend the hour on what still is: judgment, ownership, staying in the seat.

Say it out loud

Claude is just a tool.

01

The daily drivers

Claude is a Mercedes, Codex a BMW.
Different engines, different brand loyalty, a different feel when you climb in.
Real capability gaps by use case, not worth dwelling on.

02

Built, not bought

The enthusiasts who assemble their own, off an open-source base like Pi.
Rolling your own harness, tuned exactly how they like it.

Others belong here too, like Cursor, OpenCode, Amp and Factory.
I have not driven them, so I will not pretend to review them.

The move: they are all cars, and they all do the same job.
Pick a competent one and learn to drive it.
The gap between harnesses is small next to the gap between drivers.

What good looks like right now

The floor is higher than you think.

✅ 100% of your code, written by your agent

✅ Write your own PRD

✅ Pitch ideas as POCs

✅ Know deterministic from non-deterministic

✅ Build your own internal tools

✅ Understand MCP and how to leverage it

✅ Write tests and automation, with Playwright

✅ Run adversarial reviews

✅ Build your own skills, not just use them

✅ Think in workflows, repeatable, replayable, outcomes-focused

✅ Drive non-software work the same way

And who is already doing all of it?

Not who you would expect.

01

The QA

Writes the tests and the automation, not just logs the bugs.

02

The product manager

Builds a clickable prototype instead of writing the spec.

03

The designer

Vibe-codes their own tools, and hands structure to the developers’ agents.

04

The HR manager

Ships a Slack bot instead of a survey. Wasn’t writing code last year.

What should I be paying attention to?

(besides the lists previously mentioned)

Specific concepts to keep top of mind · Leverage, not volume

Slow down.

2–3x, not 10x. Quality rising, not volume.

One primary agent, one secondary. Parallelism has a human attention ceiling.

Don’t dump your speed on the team. Faster than they can absorb is offloading.

Specific concepts to keep top of mind · Participation, not verification

Stay in the loop.

You review, and you come away changed. Understanding is the bottleneck now.

Adversarial review Quizzes

Specific concepts to keep top of mind · Trust, by design

The art of back pressure.

tests types linters assertions approval gates definitions of done
Trust agents by giving them checks. Not by lowering the bar.

Specific concepts to keep top of mind · Curate your inputs

Avoid the politics, drama and hype.

Unless it’s your job.

A few quick wins · steal these today

Three tricks worth stealing.

01

Rewind and fork

Double-tap Esc to jump back and branch a session. The feature nobody uses.

02

Run it as you, anywhere

claude setup-token puts your subscription in CI, cron or a server. Headless, still on your seat.

03

Start your window early

A tiny 6am scheduled job opens your 5-hour window before you sit down. Fit three windows into a 9-hour day.

Not a tour, just a few wins you can steal. If the clock allows I will show them; if not, they are yours to try tonight.

My resources

Steal their workflows.

Someone has already mapped this terrain.

Horthy · Pocock · AI Engineer · The AI Daily Brief  + one who isn't an AI voice at all

Nathaniel Whittemore

The AI Daily Brief

Daily AI news · Nathaniel Whittemore (NLW)

Your daily 20-minute filter on the AI firehose. Signal over noise, every weekday.

WATCH & LISTEN

Spotify · Apple · every podcast app

MULTI-WEEK CHALLENGES

10-Week AI New Year Challenge
CampClaw · focused on OpenClaw

SITE & FOLLOW

Don't go it alone · Stay current

Headlines → One deep dive · ~20 min

Every weekday: a fast rundown of what happened, then one topic done properly. Stop after the headlines, or stay for the analysis.

Stay current without drowning.

You can't read everything. NLW reads it for you, from the flood of tools to the disruption to the big alignment questions, and hands you the day in one listen.

What it's good for

Staying current. The signal-over-noise filter, so you hear what moved without living in twenty browser tabs.

How to use it

Headlines on the commute for the pulse; the deep dive when the topic is your problem. It's daily, so it compounds.

Dex Horthy

Dex Horthy

Founder, HumanLayer

The clearest honest voice on making agents ship in real, messy codebases, not demos.

TALKS & VIDEOS

“Why Software Factories Fail” NEW · World's Fair '26
“What We Got Wrong About RPI” NEW
“No Vibes Allowed” · AI Engineer
“Advanced Context Engineering”

PODCAST & FOLLOW

“AI that works” · weekly, live, w/ Vaibhav

Don't go it alone · Leapfrog

Research → Plan → Implement

No code until a reviewable plan exists. You review the plan, not the diff, because that's where the leverage is: a plan is ~10x cheaper to fix than a diff.

2-3x is credible. 10x is slop.

10x is real for a greenfield prototype. On serious brownfield systems 2-3x is the honest target, and the win is quality going up, not just volume.

Beat the slop

“The way to not loop slop into more slop is to read the thing coming out and make sure it's not slop.”

Coding ≠ shipping

The model writes code fast. Deciding what matters, aligning humans, verifying and owning it, that's still the job.

Matt Pocock

Matt Pocock

Founder, AI Hero · Total TypeScript

Teaches the software fundamentals that decide whether your agents ship or make a mess.

TALKS & VIDEOS

“It Ain't Broke: Software Fundamentals” NEW · AI Engineer '26
“Agentic Engineering Workflow (just copy him)” NEW
“Your codebase is NOT ready for AI”

READ & FOLLOW

“How To Make Codebases AI Agents Love”

Don't go it alone · Take control of your codebase

AI isn't a super-powered dev. It's a new starter with no memory.

You spawn 20+ of them a day. The codebase is their onboarding: bad structure means slow feedback, hard navigation, and you patching the mess by hand.

Skills

The leading meta-prompting framework, especially on TypeScript. /grill-with-docs is gold.

Not vibe coding

“What works for humans is also great for AI.” Deep modules, clear boundaries, tests as fast feedback. Good practice for 20 years, now with higher stakes.

AI Engineer

AI Engineer

Conference + community · swyx / Latent Space

Where the field's best practitioners show, in public, how they actually build with AI.

MY FAVOURITE TALKS

“The New Code” · Sean Grove, OpenAI
“Advanced Context Engineering for Agents” · Dex Horthy
“HTML Is All Agents Need” · James Russo, HeyGen

LISTEN & FOLLOW

Latent Space · the AI Engineer podcast

Don't go it alone · Watch how the best build

Frontier labs + Real practitioners → Free

The highest-signal AI engineering talks, published free. When a new pattern lands, MCP, agents, harnesses, this is where its authors explain it first.

The highest-demand engineering job of the decade.

swyx named the role in 2023 (“The Rise of the AI Engineer”). The conference became where the field compares notes, in the open, instead of behind vendor decks.

What it's good for

Breadth and recency. One place to see what frontier labs and top teams actually ship, told by the people who shipped it.

Start here

Don't binge it. Skim the World's Fair keynote titles, watch the two that hit your stack, keep the Latent Space pod for the rest.

Alex Hormozi

Alex Hormozi

Founder, Acquisition.com · not an AI voice

Teaches leverage, value and doing the reps.

BOOKS & PODCAST

“$100M Offers” & “$100M Leads”
“The Game” · podcast

RECENT

“The Barbell Strategy for the AI Shift” NEW

Don't go it alone · and one who isn't an AI voice

Direction (you) × AI (leverage) = Output

AI is leverage: a multiplier on the work. Multiply a good direction and you compound. Multiply a bad one and you just reach the wrong place faster. The direction is a thinking skill, not a tool skill.

Don't delegate the hardest thinking work you have. You'll just get so weak, so fast.

AI can 10x your reps. It can't choose the target or hold the standard, and it won't supply focus and patience, “the two enduring competitive advantages because they're so anti-human.”

Why he's here

Not an AI voice, and that's the point. The biggest lever on AI isn't more AI; it's the leverage, value and clear thinking you bring to it. Better direction beats a better model.

Own the decision

“Someone has to own the decision.” The model recommends; you hold the standard and carry the risk. Volume negates luck, but the reps are still yours.

Don’t go it alone · The best resource

Your community.

01

Your team

The people shipping on your stack, on your problems.
The closest feedback loop you have.

02

Your friends

The peers you trade workflows and war stories with.
Low ego, high signal.

03

Your meetup groups

Rooms like this one.
Where you test an idea against people outside your bubble.

You don’t get good by buying the tool. You get good through the reps, the mistakes, and the people around you.

A few quick wins · steal these today

Three tricks worth stealing.

Rewind and fork

Double-tap Esc to jump back and branch a session. The feature nobody uses, and it kills the fear of losing everything.

Run it as you, anywhere

claude setup-token puts your subscription in CI, cron or a server. Headless, still on your seat.

Start your window early

A tiny 6am scheduled job opens your 5-hour window before you sit down. Fit three windows into a 9-hour day.

Thank you

Thank you.

Email   dylanh@ringier.co.za

These slides   mscc-claude-better.dylanharbour.com

SPEAKER NOTES · N TO HIDE