The College · a proposal, opening soon

Access isn't skill.
This is where skill gets built.

Foundations, prompts and workflows for a newsroom using AI with its eyes open.

A programme of the Digital Transformation Department
Foundations · exercise 01 02:14
Three claims in the briefing. One of them isn't real.
Checked against the primary document Verified
Checked against a second source Verified
No source exists. It was invented, fluently Caught
Session one teaches this before it teaches anything useful.
◆ Teach failure first
✓ Found in 2m 14s
How it's built

Three layers.
Everyone, then every role.

Calibration first. Craft second. And a small cohort who keep the whole thing running without us.

See the role tracks

The three layers

01 FoundationsNot a tool demo — a calibration session Everyone · 90 min
02 CraftBecause a translator and a data reporter share almost nothing here By role · 2 × 90 min
03 MultipliersJournalists, not engineers, with time protected in writing 2–3 per service
Layer three is how this scales past one department — and the first thing that gets cut. It shouldn't be.

What's inside

Opening soon

Five strands, one place. Written by the desks that use them, not for them.

FoundationsWhat these tools are bad at, demonstrated on our own beats Everyone
Prompt libraryVetted prompts grouped by desk — 15–20 good ones, not 200 unchecked By desk
WorkflowsWhere a tool fits in a real task, and where it stops Diagrams
Role tracksSeven of them, each built around a specific way the tool fails By role
The Desk toolsThe same library, with the training sitting next to it Live now
Only the last row exists today. Everything above it is what this proposal is asking to build.

Role tracks

7 tracks

Each track is built around the specific way the tool fails for that role. That failure mode is the curriculum — the useful applications follow from understanding it.

Tap any track for what it covers and the failure it's built around.
The part that compounds

The tools help teach the tools.

Models are exceptional at producing output that's plausible and wrong. Used deliberately, that's an unlimited supply of training material.

Five mechanisms

Red-team drillsBriefings with three seeded errors. Cheap, endless, and it trains the one skill that matters Highest value
Practice partnersAn evasive spokesperson, or an editor demanding your sourcing Rehearsal
Exercises from our archiveMaterial drawn from pieces the room actually published In-house
A tutor on the libraryAsk in plain language, get the newsroom's own vetted prompt Discovery
InstrumentationWhich prompts get reused, edited, or abandoned tells us what to teach next Feedback
One hard limit: AI generates the exercise. A person reviews the answer. Nothing carrying a sign-off is machine-assessed.

Roadmap

24 months

Timings are starting proposals, not commitments. They should be revised against what the baseline survey actually finds.

00
Weeks 1–4
Before any training
Anonymous baseline survey — what is already being used, and for what Pick three pilot workflows that already hurt, chosen by the desks Write the policy one-pager: disclosure, prohibited inputs, sign-off Recruit four to six volunteers, including at least one open sceptic
Gate — policy signed off by editorial leadership. Nothing below starts without it.
01
Months 1–3 · short term
Prove it on real work
Foundations delivered once per language service, in-language Prompt library live with 15–20 vetted prompts Three pilot workflows running with before-and-after numbers Weekly 30-minute open clinic, drop-in, no agenda
Not doing — rolling out to everyone. The pressure to do this will be significant and should be resisted.
02
Months 4–9 · medium term
Make it role-shaped
Four role tracks live, starting where the pilots succeeded Workflow diagrams published AI leads named per service, with time protected in writing Red-team exercise library, refreshed monthly Foundations folded into new-hire induction
Milestone — every desk owns at least one workflow it maintains itself.
03
Months 10–24 · long term
Make it self-sustaining
The College is the default reference — checked before asking a person Light-touch per-track badging, human-reviewed Contribution model: journalists submit, AI leads vet Annual curriculum rewrite, because the tools will have moved
Exit criterion — DTD stops running the training and starts running the platform. Still delivering Foundations in month 24 means this didn't work.

What we hold to

7 commitments

Nothing publishes itself.

Every tool stops short of the send button, on purpose.

Every output is a draft.

Something to argue with, never an answer to accept.

The byline stays human.

Credit and accountability sit with the journalist.

Teach failure first.

Session one shows the tool being confidently wrong, before it shows anything useful.

Train on live work.

No toy exercises. People bring a real task from their own desk.

The trainer is a colleague.

Peer-led, in-language, in-house. Not a vendor deck.

Sceptics are quality control.

The person who distrusts the tool finds the failure modes. Recruit them.

What this needs to start.

Protected time — 90 minutes once for every journalist; ~4 hours a month per AI lead, defended Policy sign-off from editorial leadership before Phase 00 closes One named owner for College content. Not a committee Tool licences for the pilot cohort, then per service Permission to stay small for the first three months
The College — training, prompts and workflows for AI in the newsroom.
A project of the Digital Transformation Department (DTD), RFE/RL.
PilotDesk The Desk Role tracks Roadmap