Hands-on Claude Code training so your engineers take a real piece of work from brief to reviewed, tested change in their own codebase, in half a day.
10,000+ people trained across 500+ organisations
Claude Code plans and makes changes across a whole codebase, not one file at a time. Used well, it changes how engineers find work, build it, review it and ship it. Used carelessly, it produces confident mess at speed.
This half-day course is built around your stack, your repositories and your tickets. Each engineer picks a small, real build at the start and takes it all the way to a reviewed, tested change by the end, with the guardrails that keep code, secrets and production safe.
What engineers said after the class
100%
said the class will help them get more value from Claude Code at work
94%
enjoyed the class
79%
would recommend it to a colleague
5 hrs
average time each engineer expects to save every week
Post-class surveys from the engineering cohorts at a 400-person UK consumer technology company, July 2026.
Using it much more for planning work before execution and setting up systems that stop me repeating myself or it making more basic mistakes
The project session was good as it allowed an opportunity to create something.
Every session is built from scratch around your work, but every session builds the same foundations.
Installing and updating Claude Code, the four permission modes and when to use each, and having Claude interview you to write a CLAUDE.md that records your conventions, tooling and access.
Scanning repos, tickets, logs, errors and usage data for opportunities with evidence behind them, then debating the backlog with a simulated product team before committing to one small build.
Extracting your coding style from the repo into CLAUDE.md, and building a Stop hook that makes Claude prove it is done: acceptance criteria met, tests run, no TODOs left.
Having Claude interview you for a one-page implementation brief, reviewing it sceptically, then planning before any edit with the plan approved by a human.
Choosing the model and context with cost in mind, handing off cleanly with /clear, and steering live by queuing follow-ups and interrupting when Claude drifts.
Running two Claude Code sessions on two independent tasks at once across git worktrees, and delegating to sub-agents when a task splits cleanly.
Building a scientific-debugging Skill, then a read-only sub-agent acting as a ruthless tech lead that ranks what to fix, with the top issue fixed and proved by a test.
A verification plan with unit and end-to-end tests, browser QA through Claude in Chrome and the Playwright MCP, and a final summary of the diff, evidence and risks before you commit, raise a PR or park it.
The capstone
Every engineer takes one real build from their own repo all the way to a reviewed, tested change, so the team leaves with shipped work and a repeatable way of working, not just notes.
We can usually map your highest-impact use cases in a single conversation.
Get in TouchThese came out of our Claude Code classes, built by engineers in their own repositories.
A read-only scan that maps models, jobs, queues and every external integration, then shortlists fragile paths, missing error handling and thin test coverage with file evidence.
Documents naming, structure, error handling, logging, retries and test patterns with file references, then saves it as a Skill that brings new code into line.
A read-only sub-agent reviews the code like a ruthless tech lead and returns a severity-ranked fix list, with the top finding fixed by a minimal diff and a proving test.
Runs tests and reads logs to shortlist issues, then diagnoses the root cause with code evidence, one hypothesis at a time, before proposing the smallest fix.
Drawn from our anonymised client use case library.
Three signals showing how fast agentic coding is becoming standard practice. Each links to its source.
The day runs in four parts. Each engineer works on their own small build throughout, so every exercise moves real work forward.
01
Getting set up and the permission modes. How LLMs generate code and why deterministic tests matter. A read-only first contact with git, gh and your tooling. Finding something worth building, a product-team debate, house style and a Stop hook.
02
A sharp one-page brief, reviewed sceptically. The plan-first loop with human approval. Model and context choices with cost in mind. Clean handoffs. Steering live. Two tasks in parallel across git worktrees.
03
A scientific-debugging Skill. Delegating to sub-agents. A read-only red-team review with a ranked fix list. Annotating a live interface so Claude makes the changes. Drafting a scheduled routine that checks dependencies and scans for secrets.
04
A verification plan with unit and end-to-end tests. Browser QA that tries to break the app. Guardrails for secrets and least privilege. Taking the build from zero to shipped, then a short build sprint on something real.
Delivered in person at your offices, or remotely.
Engineering leads rightly ask what happens to their code. Safety is built into the day rather than bolted on at the end.
Engineers start in the permission mode where Claude asks before every action, and never use the mode that skips permission checks
The first look at your tooling is read-only, and reviews by sub-agents are read-only by design
No .env files opened, no credentials printed and no live endpoints called during exercises
Secrets, least privilege and review before every commit are taught as a dedicated guardrail
Scheduled routines are drafted with a human review step, not switched on unattended
Anyone without a repo they can safely use works on our practice repository instead, so no client code is ever required
Going further
Once a team has the habits, the next step is making them a standard across every repo and engineer.
Teams who want to use Claude Code properly rather than by trial and error, including those sceptical about AI.
Developers working on real products and integrations who need changes that are tested, reviewed and safe to ship.
Leads and platform teams deciding conventions, guardrails and how agentic coding fits the review process.
Data engineers and BI developers who want the same rigour on pipelines, scripts and reporting code.
Other courses teams pair with this one.
Every Claude Code session is led by one of our instructors, with a dedicated tutor in the room so nobody gets left behind.
Founder + Chief AI Officer
AI Instructor
AI Instructor
AI Instructor
AI Instructor
AI Instructor
AI Instructor
AI Instructor
Yes. Each engineer works in a repository they can safely use, so the brief, the build and the review are all real. Anyone without one uses our practice repository instead.
Yes. The syllabus covers CLAUDE.md, the permission modes, plan mode, a Stop hook, model and context choices, git worktrees, sub-agents, Skills, the Playwright MCP and Claude in Chrome for browser testing, and drafting scheduled routines.
Engineers start with Claude asking before every action, first contact with your tooling is read-only, exercises never open .env files or call live endpoints, and secrets, least privilege and review before commit are taught as a dedicated guardrail. No client code is required if you would rather use our practice repository.
Claude Code installed and updated, a terminal with git and gh, and a repository they can safely work in. Claude in Chrome is needed for the browser testing exercises. We send a short setup checklist in advance.
Yes. Our Claude Code training for non-developers teaches analysts, finance and operations teams to build their own tools from plain English. Many firms run both.
In person at your offices by default, with an instructor and a dedicated tutor in the room. Having people on hand makes setup and the first real change much smoother. We can run it remotely if that suits your team better.
Yes. We are based in Covent Garden, London, and run most sessions in person at clients’ offices in London and across the UK. We also work with teams worldwide, from Paris and Berlin to New York and Singapore, and can run sessions remotely when that suits better.
From The Distillation, our newsletter on putting AI to work.
We analysed our last 30 client conversations to find the five AI challenges firms raise most, from uneven adoption to proving ROI. Plus field notes from Cowshed and Montagu, The AI Summit London 2026, and funded graduate training places.
Josh Page ·
Anthropic just built the most capable AI model ever and decided not to release it. The cyber capabilities are extraordinary, the coding benchmarks are unprecedented, and the implications for every firm using software are serious.
Tom Hewitson ·
Reasoning models do more than predict the next token. They break problems into steps, use tools, inspect evidence, and adapt their approach as they work. That makes them especially useful for analytical and multi-step business tasks.
General Purpose ·
Tell us about your stack and how your engineers work, and we will design a day around your repositories and the way your team ships.
Get in Touch