Overview
Consultants who already use AI every day are increasingly asked to build with it: to design multi-step agents and automations, take them into client environments and show that they can be trusted. This advanced one-day workshop moves straight past basic prompting. Delegates learn to design, build and evaluate agents in Claude, and to take them to clients with strict IT, data-protection and validation requirements.
The day is organised around squads. Each squad works on a realistic, fully synthetic set of project documents (requirements, specifications, test and training documentation, with errors deliberately seeded in) so that agent output can be checked against known answers. Every module ends in something working, and no client data is used at any point.
Who it's for
Consultants who already use AI daily and want to learn to build with it, working on implementation and change projects in regulated settings. It is especially suited to:
- IT/OT implementation consultants
- Programme and project managers
- Business analysts and requirements managers
- Quality, validation and process-improvement specialists in pharmaceutical, manufacturing and other regulated sectors
Prerequisites
- Daily use of a large language model and structured prompting; basic prompting is not taught
- Hands-on experience of implementation, project, analysis or validation work, ideally in a regulated setting
- Comfort with logical workflows and structured documents; no coding or software engineering background is needed
By the end of this course you will be able to
- Design a multi-step agent or workflow with a repeatable canvas covering goal, inputs, tools, context, guardrails, human checkpoints and evaluation
- Test an existing AI accelerator on realistic data, find where it fails and improve its instructions, structure and checkpoints
- Plan how to package and deploy an accelerator in a client’s approved tools, with handover, adoption and support
- Build a working prototype or mock-up tool, with its own evaluation set, using Claude
- Evaluate agents for drift, hallucination and reliability, and design human-in-the-loop checks and audit trails suited to regulated work
Course outline
Times show the standard running order and can be adjusted when we tailor the course to your team. The trainer stays available during breaks to help anyone finish an exercise.
Welcome & ways of working 10 min
The shape of the day, how squads work, and the ground rules: only synthetic material goes into the exercises, and nothing built is treated as validated.
Tools and methods for regulated delivery 30 min
A short tour of the Claude capabilities most relevant to IT/OT and regulated project work: Projects with custom instructions, Skills, connectors, document analysis, interactive pages and small apps built in the chat, and Claude Code for fuller builds. Each is shown on a concrete example from requirements, specification, test or training work. The design method that follows applies to any tool your organisation uses.
From prompts to agents 30 min
Workflows versus agents, tool calling, context and memory, and coordination between several agents. Just as important: recognising when an agent is the wrong choice and a simple, well-written prompt or a fixed workflow is safer and cheaper to validate.
The agent design canvas 35 min
One repeatable method for every agent: goal, inputs, tools, context, guardrails, human checkpoints and evaluation, captured on a single page that a reviewer or quality lead can read.
Testing and improving an AI accelerator 40 min
Each squad takes one flagship offering and its AI accelerators, supplied by the client or by you, and tests them on the synthetic data to see where they work and where they fail. Then tighten the instructions, tools, structured outputs and human checkpoints, building on the canvas.
Getting accelerators out to clients 30 min
Packaging an accelerator as a repeatable offer: deployment in the client’s approved tools, handover, training and adoption, support and versioning, and the questions client IT and quality teams will ask.
Rapid prototyping lab 80 min
Squads build and test one working prototype or mock-up on the synthetic data. Choose one option or propose your own for trainer approval. A: a specification-to-test traceability agent that checks or generates test cases and flags gaps and change impacts. B: a test documentation assistant that drafts and reviews scripts and turns results and incident logs into structured reports with review gates. C: an interactive project cockpit mock-up, such as a data-flow overview or a risk and change dashboard, for a client demonstration. D: your own idea, confirmed in a scoping call beforehand. Every prototype includes a completed canvas, defined human checkpoints and a small evaluation set drawn from the seeded errors. The trainer coaches throughout.
Evaluating agents 30 min
Test sets with known answers, scoring rubrics, checks for drift and hallucination, and regression testing when models or instructions change. Using AI as a reviewer, always under human oversight.
Governance for regulated clients 25 min
Human-in-the-loop validation gates, schema-enforced outputs and audit trails. What FDA 21 CFR Part 11 and EU GMP Annex 11 mean for tools used near GMP systems, and where the draft revised Annex 11 and proposed Annex 22 on AI stand; these were consulted on and their status should be confirmed before you rely on them. This is awareness for delivery teams, not validation or legal advice.
Squad demonstrations & action planning 35 min
Each squad demonstrates its prototype in about five minutes and receives peer and instructor feedback. Then each squad names the next steps for taking its prototype or accelerator to a client.
What you'll need
Equipment and accounts
- A laptop with a modern browser and a working Claude account on a plan that supports Projects, Skills and file creation. Claude Code is shown in the tour; running it is optional.
- The plans and features needed are confirmed with your IT team during preparation and re-checked shortly before the day.
Preparation with the trainer
- A fully synthetic data pack, agreed beforehand, covering a short brief for each flagship offering and its accelerator, a user requirements specification and interview notes, functional and design specifications for a fictional system, test scripts, results and incident logs with seeded errors, and work instructions and role-based training material. No client data or confidential content is used.
- A kick-off call with the trainer about six weeks before the day, followed by review of the data pack and a set-up checklist one week before.
All exercises use synthetic material prepared for the course. Nothing built on the day has been validated for GxP use.
Practice files are provided for every exercise. You are welcome to use your own everyday documents where your organisation's policy allows.