Practical AI capability through responsible, hands-on learning.

An independent, evidence-led research project on the UK's AI skills gap — and what a credible response to it would need to look like.

---

What this is

Grounded AI Practice is a research project examining how UK adults and small organisations actually build practical AI capability - and why the current national approach appears to be falling short.

It is also its own first case study. The file structure, working rules, verification tooling and logged decisions here are built so that an AI assistant can read them and work from them directly. That system is what produces the research in this repository — so how it is built, and why, is on show alongside what it found.

Whether the same practice can be taught to other people is a working hypothesis, not a finding, and the brief records it as one. Most people using AI today use it as a chatbot and stop there. The untested claim is that a practice like this one produces better results at lower cost.

It is currently at the research and pilot-design stage. This repository is the working record: the sources, the reasoning, the decisions, and the things still unresolved.

What it is not: a finished curriculum, a product, or a set of conclusions presented as settled. Where the evidence is thin, the log says so.

Why the repository is public

The research method is the output at this stage. Anyone can read every source consulted, every inference drawn, and every claim the project has retracted or failed to verify.

That is deliberate. A project arguing that the UK needs better-evidenced AI skills provision should be able to show its own evidence.

How the research works

Four rules the project holds itself to, visible throughout the logs:

Rule	What it means in practice
Evidence and inference are never blurred	Every entry separates what a source directly supports from what the project inferred. Opinion is labelled as opinion.
Sources are tagged by interest	Government, independent/academic, vendor/commercial, advocacy. Makes it visible when the evidence base is concentrated in parties with a stake in the answer.
Confirming and competing evidence are sought together	Competing viewpoints and alternative sources are considered to mitigate potential biases and ensure decisions are made objectively.
Nothing is silently altered	Corrections and retractions are added as new entries referencing what they supersede, so the reasoning stays traceable.

Where to look

File	Contents
project_brief.md	Problem statement, scope, working decisions, what is still open. Start here.
research_questions.md	The ten research priorities driving the work.
research_log.md	Source key and dated findings — source-backed research only. Ends with a running list of open threads.
project_log.md	Dated record of decisions, design and production work.
exports/	Finished, approved outputs.
drafts/	Work in progress. Nothing here is settled.
assets/	Visual identity: logo system and a 36-icon content set.
CLAUDE.md	Working rules and conventions, including how this repo separates public from internal material.

Current focus

• A research report on the UK's AI climate, government posture, and actually delivered results.

• A single pilot learning unit on effective prompting — deliberately small, designed to be tested with real learners before anything is scaled.

A note on completeness

Some working material is kept out of this repository: private contacts, funding strategy, and candid assessments of named parties. The logs record that such positions exist and what their evidential status is, without reproducing content that named individuals have not consented to seeing published. The rules governing that split are themselves public.

Tooling

Research and document production are AI-assisted, with human review at every decision point.

---

_{Independent research project · United Kingdom · 2026}