TrainingBoard and Trust · A strategic framework · 2026

Advanced AI
training.

How to think about and instruct LLMs. Prompt engineering has become complex and technical — but the real gains come from a handful of mental models, not from technical syntax. This is the framework.

SessionBoard and Trust

Advanced AI training 01 / NN

The session

What we'll cover.

01

The current landscape

LLMs have radically changed and improved over the last three years. What are the substantive changes, and where are we today?

02

How to think about LLMs

Too much training focuses on tactics and tricks. It's more important to have mental models that can be applied to a range of applications.

03

Practical examples

Identifying possible use cases is a constant challenge for a general-purpose tool with jagged intelligence. We'll cover three concrete examples.

04

Getting started

LLMs can't be taught. How do you make space in your work to get practical experience, and what kind of project could you start with?

Advanced AI training 02 / NN
The state of AI

The current landscape.

LLMs have radically changed and improved over the last three years. What are the substantive changes, and where are we today?

Advanced AI training 04 / NN

The state of AI · capability

Beyond chat: the new AI stack.

01

The reasoning brain

Modern models have moved beyond word prediction. They plan approaches before generating a response. Recent models have blown through existing evaluations.

Diagram of the reasoning brain.
02

The harness

The harness is the execution environment. By giving models access to tools (search, terminal, CMS, browser, apps), AI moves from talking about work to doing it.

Diagram of the harness — the model's execution environment.
03

The agentic loop

The synergy of brain + harness creates an "agent". It doesn't wait for your next prompt; it creates a plan, uses tools, and works in a loop until the goal is achieved.

Diagram of the agentic loop.
Advanced AI training 03 / NN
Fable: an editorial collage of butterflies and moths arranged in a spiral, in the Guardian palette.

The state of AI · pace

Why this is already out of date.

Anthropic have released Fable today. This is the guardrailed version of Mythos. Two things distinguish it. It is more capable of completing highly complex end-to-end projects from a single prompt. And it represents a new level of safeguarding (specifically to stop the model from being used to identify digital vulnerabilities). Early signs suggest this is as big a leap as the one that occurred in November 2025 with Claude Opus 4.5.

Advanced AI training 04 / NN

The state of AI · agency

AI is not just something that happens to us.

We are professionals with clear codes and practices that govern behaviour. And we can make judicious decisions about how we use AI in that frame.

01

Should I use AI?

This is entirely in your hands. But too often it's presented as the only element of control, and one that should be imposed on everyone.

02

Where should I use AI?

Your job is very likely an amalgam of a lot of jobs. Some of those are appropriate for AI and some absolutely depend on your personal control and expertise.

03

How should I use AI?

AI shouldn't — and doesn't have to — be used to execute a whole task. Instead of asking it to create something you can use it to give feedback and suggestions.

Advanced AI training 04 / NN

The state of AI · adoption

AI has already arrived… and quicker than anything else.

Adoption in the first three years after launch — same window for each technology.

Internet at work (1995–98) Smartphone adoption (2007–10) Generative AI at work (2022–25)
50% 40% 30% 20% 10% Launch +1 yr +2 yrs +3 yrs 20% 25% 41%

Within the same three-year window from launch, generative AI is at 41% workforce adoption — roughly double internet-at-work and smartphone adoption at the equivalent point. Sources: US Census / Pew (internet, 1995–98), Pew Research (smartphone, 2007–10), Federal Reserve (GenAI, 2022–25). Y-axis is % of relevant population (workforce / adults).

Advanced AI training 06 / NN

The state of AI · adoption

The enterprise gap.

Firms that have formally adopted

18%

Of US businesses report adopting AI in any business function (Fed BTOS, end 2025).

Workers using it anyway

41%

Of US workers use generative AI at work — often via personal accounts, ahead of their employer (Fed RPS).

Why deploying large-scale projects is hard.

  • Identifying use cases What's flexible for one person becomes hard to define for a thousand. Personal use bends to the user; organisational use has to nail the use case first.
  • Reproducibility and audit trail A sole trader can let AI draft an invoice. A 1,000-person firm needs an output that survives an audit.
  • Error consequences at scale A hallucinated paragraph is a shrug in a personal email; a correction in a 200,000-circulation paper.
  • Workforce concerns about displacement Real, live conversations inside most large organisations — including ours.
  • Negative customer reaction AI-produced news is seen as less trustworthy than human-produced news, even with human oversight.

Sources: Federal Reserve, April 2026; RISJ Digital News Report 2025. The Fed also notes adoption among the smallest firms is "stronger than would be expected based on size alone".

Advanced AI training 07 / NN
The thesis

Why most prompt advice misses

How to think about LLMs.

Prompt engineering has become complex and opaque. But the whole value of LLMs is that you can talk to them. Our training focuses less on tips and more on mental models that are better guides to good results.

Advanced AI training 07 / NN

Before we begin · Non-negotiables

Four key rules.

Rule 01

Check everything

You are responsible for anything you produce, either for internal or external use. Any input or sources used to create text or multimedia content, whether generated by AI or taken from other sources, must be checked. This is enshrined in our editorial code but applies to all our staff.

Rule 02

Use GenAI within your area of expertise

Generative AI is particularly risky in situations where it is used outside your field of expertise. It is tempting to use it as a shortcut to avoid resource constraints in other departments or functions. But a lack of expertise means a lack of ability to accurately judge potentially unreliable output. If you do use the technology in this way you must check your work with someone qualified to evaluate it.

Rule 03

Use GenAI as a co-pilot

GenAI can be used to completely automate a piece of work. But this approach gives too much responsibility to an unreliable machine and makes it tricky to check the output accurately. Instead, work with these tools to help you complete tasks step-by-step. You should still use your own creativity, thought and judgment. Outputs from LLMs may be owned by the LLM provider or not protected by copyright if they are not the product of human effort.

Rule 04

Don't ask GenAI to make decisions or form opinions about people

GenAI is not a thinking tool and it does not replace your role as an accountable decision maker. Do not use it to make decisions about individuals. It can be used to review and summarise content but only you can make decisions about another person.

Advanced AI training 03 / NN

How the machine thinks

The quantum leap principle.

A collage portrait in the Guardian palette illustrating the Quantum Leap idea — a mind that arrives brilliant but unoriented.
01

The Swiss cheese brain

The model pops into existence each time with no memory. It "leaps" in with massive intelligence but zero orientation. Your job is to fill the holes.

02

The fundamental anchor

If the model has amnesia by design, context is your most fundamental tool. Without it, the machine's brilliance is unanchored and prone to drift.

Advanced AI training 03 / NN

What you bring

Irreplaceable expertise.

01

The problem

Articulating the specific issue you want to address with absolute clarity and precision.

"I need to build a slide deck to help train the board on advanced AI."

02

Support context

Any supporting data, internal logic, or unique expertise that anchors the solution.

"Here are five recent training decks I've used. Absorb the style of writing and also the content. I will want to adapt many of them. Also ingest this academic text on proven presentation techniques."

03

Ideal output

The context for format (e.g. slides), tone of voice, intended audience, and specific constraints.

"Please take the design language, fonts and styling from the attached presentation and create a prompt that will reliably style my slide requests consistently in this form."

04

Governance

The magic of LLMs is their ability to infer intent from language. This is also the biggest cause of errors. Don't expect a prompt to work in one shot. Sculpt your way to success through conversation.

"I like the slide showing the graph of adoption but to make it land it needs to animate the lines from left to right."

Advanced AI training 04 / NN

What you bring

The magpie habit.

01

Collect everything

Maintain a digital scrapbook: designs, tweets, essays, YouTube videos, screengrabs, audio notes, academic papers, UIs … and anything that strikes you as useful collateral information.

A blue line-drawn illustration of Regent Street, London — a collected visual style reference.
02

Feed the machine

Images, voice notes, scribbled thoughts, screen grabs, videos and inputs that reduce the amount of time you need to prompt through text and can influence a project's creative direction.

An engraving-style portrait montage — an example image fed into a project.
03

Sculpt and capture

Refine the inputs you've given the machine by nudging it towards a more personalised vision, enriched by your expertise. Then turn the output into a reproducible prompt.

The same montage reworked with colour and texture — a sculpted, adapted output.
Advanced AI training 05 / NN

How to instruct

The mentor & the manager.

Prompt level

Think like a teacher

Teaching is a skillset, and it's exactly what the machine needs. Like a child, it's capable but forgetful — easily distracted, easily led astray. Clear, patient instruction and a steady process matter far more than clever syntax.

Project level

Manage the team

Lead a team on individual tasks that add up to a whole. Break big projects into small, achievable, and verifiable chunks.

Advanced AI training 06 / NN

The workflow

The strategic expedition.

Start wide, then narrow down: four waypoints from aim to a verified result.

01
02
03
04

Define aim

Set the ultimate goal. Start wide — tell the machine you are planning a project.

Map steps

Deconstruct it into small, modular chunks.

Specific context

Feed the data and expertise each step needs.

Loop & refine

Iterate and verify at every waypoint.

Advanced AI training 07 / NN
In the newsroom

Practical examples.

All this can sound hypothetical and vague. Here are three concrete examples of what we've been building with Claude Code in the newsroom, responding quickly and effectively to requests that would otherwise remain unrealised.

Advanced AI training 14 / NN

A beginner's deep research prompt

“How might the agentic web affect the Guardian's journalism and revenue model in the next five years?

Think about your own deep research question that might be more relevant.

Advanced AI training 17 / NN
Screenshot of a UK fuel-price tracker dashboard: average prices, regional breakdowns and anomaly detection.

Fuel tracker · A worked example

Building a fuel-price dashboard.

The opening prompt

The head of Consumer wants an easy way to keep track of and monitor fuel prices across the UK. We have a set of reliable sources for the data below. Identify any other official sources we may have missed and then suggest a framework for a dashboard to allow the editor to apply their expertise to search the data in multiple ways and through different visualisations with the intent of uncovering significant disparities or trends in the context of the impact of the Iran war on UK consumers.

Advanced AI training 18 / NN
Screenshot of the Institute for the Study of War's daily Iran Update report.

Iran Watcher · A worked example

Building a bespoke newsfeed.

The opening prompt

I want to build a daily Iran briefing for the Guardian's defence and security editor. His most authoritative read right now is an American think tank's report — three thousand words a day, written by six named analysts. I don't want to clone it. I want something a UK newsroom can sustain: faster, weighted to primary sources we can actually get at, and honest about what won't be possible. Let's scope it properly first — what's realistic, what isn't, and where we'd add something they don't. I'll feed you sources and examples that the editor values as we go.

Advanced AI training 15 / NN

Iran Watcher · Context in

Feed it material. Keep feeding it material.

The project was born from one thing, and stayed sharp because of one thing: material being fed into the conversation.

What we fed in What it triggered
The opening brief: what the editor needed, and the report to measure against An honest plan for what we could realistically build
A screenshot of a UN approval email A humanitarian-reports feed, added within the hour
A screenshot of the site looking broken on a phone A full redesign for mobile
An off-hand remark that we'd used up our free social-media feeds Two more papers found that we could already pull in for free
A screenshot of a shipping-incident post we couldn't reliably read One line of inquiry dropped, a sharper one started
Five of the think-tank's reports, shared over six weeks Five side-by-side comparisons, each exposing the next gap
Advanced AI training 12 / NN

Iran Watcher · Suggestions out

Once it has the context, it starts proposing things you didn't ask for.

Three moments where the model suggested something Chris hadn't asked for:

Chris asked

"We've used both our free social-media-feed slots on IDF and CENTCOM. The question is whether to pay for more, or write a custom Twitter adapter."

Claude suggested

"Before either — let me check whether Times of Israel and Jerusalem Post are on Bluesky's free public API. They are. About twenty extra items a day, from a layer you didn't ask about."

Chris asked

(after a benchmark) "We're missing the vessel-level maritime tracking the think tank cites — sources like Lloyd's List and Windward AI."

Claude suggested

"Both paid, enterprise tier — out of reach. But the OSINT outfit the think tank actually cites in its footnotes — Tanker Trackers — is on Bluesky and posts publicly. One line to add."

Chris asked

"Should the Twitter and Bluesky cards be treated the same way as the Iranian state Telegram ones?"

Claude suggested

"Partly. The Israeli army's Twitter is a broadcast — same band as the regime channels. The Israeli newspapers on Bluesky are different — they're newspapers that happen to deliver via Bluesky. Suggesting we split on editorial unit type, not just provenance."

Advanced AI training 13 / NN

Iran Watcher · Benchmark and repeat

Verify against something external. Repeatedly.

Our benchmark was the original ISW report. Across six weeks, I asked Claude to compare our report to theirs. Each comparison was published as its own page — what they covered, what we covered, what each side missed.

Every comparison identified the next gap to close.

Comparison date What the comparison identified as missing What got built next
21 April First benchmark — multiple gaps Editorial design overhaul; threading and filters
24 April Better presentation of UK Parliament questions Rich question-and-answer layout
30 April Direct access to Iranian state messaging Three Iranian regime channels added
1 May Military operational broadcasts (Israeli army, US Central Command) Both social-media feeds added
14 May Vessel-level maritime tracking Tanker Trackers added

Without the comparisons, the project would have drifted on instinct. With them, every round of work had a measurable target.

Advanced AI training 14 / NN

In the newsroom · And more

But that's not all…

Could you build me an equivalent of Google Trends for the Guardian? What are the data sources in Hansard, and what could I build with them for journalists? How might this work for Congress? Can you create an interactive map that builds on a Guardian article about derelict churches?
Advanced AI training 23 / NN
Over to you

Getting started.

LLMs can't be taught. How do you make space in your work to get practical experience, and what kind of project could you start with?

Advanced AI training 20 / NN

Getting started · No shortcuts

First, you have to use it.

You can't get your hands around what LLMs can do — or grasp their potential at scale — without committing real time to using them. It's non-negotiable. Skip it, and you're the person who tried AI Overviews once, decided GenAI was broken, and walked away.

  • Put time aside, and protect it. Be disciplined about actually turning up.
  • Expect not to finish. The project may never land — but the hours still teach you how they do and don't work.

Don't know where to begin? You have a magic genie in front of you. Ask it. Here's how I started:

My very first prompt

Can you build me a simple app on my machine? I can't code. I will need you to explain everything to me in simple language. Don't assume knowledge on my part. Redirect me if I ask for something stupid or there's a better way.

Buy Claude, and use the desktop app.

Advanced AI training 24 / NN

Getting started · Three ways in

Pick a first project.

The best first project is something you already know and care about. Three ways in:

  1. 01

    A hobby

    Start where you're passionate. You know the territory, and you know where it's painful.

    "I'm a keen photographer, so I asked Claude to build me a small Mac app that reads the forecast and tells me, the night before, whether tomorrow's golden hour is worth shooting."

  2. 02

    A child's homework

    Use it as assistant and tutor. You'll learn from your kids, learn to steer them past just getting the answer, and see what a teacher it can be.

    "My daughter couldn't get expanding double brackets — honestly, neither could I. So I asked Claude to explain it in plain English, then build us interactive examples and quizzes."

  3. 03

    A week at work

    Take one real task and use AI at every step. You're finding where you want the machine — and where you don't. Not a long-term project; prep for one big thing.

    "Before a meeting to settle a long-running project, I asked Claude to pull together three months of Chat, Mail, Docs and Slides — where we'd got to, what's still open, who owes what, and what I need to tie off myself."

Advanced AI training 25 / NN