From ecosystem consultant to Claude

You consult in an enterprise ecosystem. Salesforce, ServiceNow, Google Cloud, Databricks, SAP, Workday, Dynamics. You run discovery workshops, map processes, write requirements, configure, design, and get a change landed so it sticks.

You have been watching AI from the side of the room, and you want to be in the middle of it. You do not know where to start, and the courses you have looked at teach the platform as though you were starting from nothing.

You are not starting from nothing. This pack covers what transfers, what to learn, and six exercises that take you from your first real Claude work to an agent design, each leaving you with an artefact you can show.

There is an installable Claude skill that runs all of it with you, above. It assesses where you are starting from, runs each exercise (including playing five difficult stakeholders in a workshop rehearsal), and reviews what you build against the bar a hiring team would use.

What you already have

The hard part of getting AI into production is not the model. It is understanding the process well enough to know what an agent should do, and landing the change so people use it. Consultants from other ecosystems already have that:

• Running a workshop where the stakeholders disagree with each other
• Mapping a process as it runs, not as the documentation says it runs
• Knowing when a requirement is a political constraint in disguise
• Being the outsider in someone else’s building and earning trust inside a day
• Landing a change so it sticks, rather than shipping software and leaving

An agent built without this understanding is a demo. An agent built with it is the thing people keep using at month six.

What you need to learn

• What Claude does well, where it stops, and how to tell the difference on real work
• How skills, plugins, connectors and agents fit together
• MCP, the standard that connects Claude to the systems a business already runs
• Writing instructions a model follows consistently, which is a craft rather than a knack
• Claude Code, which is where build work happens

What will be hard

Writing logic in prose. Configuration trains you to express logic as settings. A skill is logic in sentences, and ambiguity in a sentence becomes inconsistency in behaviour.

Probabilistic behaviour. The same input can produce different outputs. The job is making the variation acceptable and catching it when it is not.

The pace. A prototype takes an afternoon. If you come from an ecosystem with quarterly releases, the instinct is to over-plan the first build. Build it, test it, change it.

Being the builder. If your ecosystem has a separate dev team, you are used to handing specs over. Here the person who ran the workshop builds the first version.

The translation table

The fastest way into the Claude platform is through the one you know.

Where to learn the platform

Anthropic runs a free Claude Academy. It has courses on using Claude for everyday work, on what language models can and cannot do, on Claude Code, on building and sharing skills, on connecting Claude to other systems with MCP, and on building agents with the Claude API.

This pack does not repeat any of it. Each exercise below names the topic to study first, and then gives you something to build with it.

Keep the courses in proportion. They give you the language. What you build gives you something to talk about. Someone who has finished every course and built nothing has not moved, and an employer can tell within a minute.

The exercises

Exercises 2 to 6 run on one fictional client, Hartwell Freight: a UK road freight company, 2,400 people, four months into an Anthropic Enterprise agreement, with a CIO who has promised the board two AI processes in production. Her first choice is customer claims. 900 a month, 19 days to close, and a finance director who believes £400,000 a year goes on claims Hartwell was not liable for.

The skill holds the full brief, including five stakeholders who each know something they will not volunteer unless you ask the right question.

A fictional client matters for a reason beyond convenience. Your current clients’ data, documents and system details should never go into a personal Claude account, even anonymised. Your employer’s contracts almost certainly forbid it, and a portfolio built on a client’s material becomes a liability the moment an interviewer asks where it came from.

Each exercise’s output is the next one’s input.

1. Your own work, pushed as far as it goes

Learn first: the Academy’s introductory material on using Claude for everyday work, and on what models can and cannot do.

Take something you wrote yourself and own: a workshop agenda, a requirements template, a process guide. Remove anything that identifies a client. Hand it to Claude and keep pushing past the point where you would normally stop. Restructure it, rewrite it for a different audience, turn it into an interactive artifact, ask what is missing.

You end with: the before and after, and a note of what surprised you and what Claude got wrong that would have embarrassed you.

This is the most important of the six. It shows you reach for AI on your own work before anyone asks you to, which is the first thing an employer in this space looks for.

2. Plan a discovery workshop, and rehearse it

Learn first: the Academy’s material on working with AI effectively: how to delegate, describe what you want, and judge what comes back.

Plan the half-day claims workshop at Hartwell with Claude: objectives, agenda, stakeholder map, question bank. Then rehearse it. The skill plays the CIO, the head of customer service, the finance controller, a depot manager and a senior claims handler, in character, and holds back what each of them would hold back in a real room.

Afterwards it steps out of character and tells you which of the hidden facts you found, which you missed, and the question that would have found each.

You end with: a workshop plan and a rehearsal debrief.

3. Map the process and find where AI changes it

Learn first: nothing new. This is your existing craft. Our guide to mapping an AI operating model covers the method we use, if you want it.

Map the claims process as it runs, from what the rehearsal uncovered: every step, who does it, how long it takes, where it waits. Then draw the future state and mark each step as done by a person, done by Claude with a person approving, or done by Claude alone. For each Claude step, note what it needs to reach, and whether security has approved that.

If you prefer, map it in the Kowalah Plugin in Claude. Installing it and signing in gives you your own organisation with an empty operating model, free. A document or a diagram works as well for the exercise.

You end with: a current-state map, a future-state map with each step classified, and the access each Claude step needs.

4. Build the client playback

Learn first: by doing. Claude Design builds decks, one-pagers and prototypes from a conversation.

Build what you would take back to Hartwell’s CIO a week after the workshop: what you found, the current and future state, the one first build you recommend, what has to be true for it to go ahead, and the next step. Write it for a CIO who will forward it to the finance director.

You end with: a deliverable you would be comfortable sending to a real client. It leads with what you found, not with what AI can do.

5. Package your method as a skill

Learn first: the Academy’s material on building and sharing skills, and its introduction to Claude Code.

Turn the way you plan discovery workshops into a Claude skill: a folder with a SKILL.md and, where they help, reference files and templates, so a colleague can plan a workshop for any process the way you planned Hartwell’s. Install it, test it on a process that is not claims, and revise it.

The description at the top of a skill decides when Claude uses it. Test that it fires on “help me prep a stakeholder session” and does not fire on “write an agenda for the team meeting”. Our Claude skill template shows the structure.

You end with: a working skill, tested on two cases, with a note of what you changed after testing.

6. Architect your first agent

Learn first: the Academy’s material on connecting Claude to other systems with MCP, and on building agents with the Claude API.

Design the agent for the highest-value step in your future-state map. At Hartwell that is gathering the evidence for a claim, checking it against the customer’s contract, and drafting a recommendation for a handler to approve. The skill includes a design template covering purpose, what the agent must never do and how that is enforced, tools and access, evaluation cases, cost per day, and where your build stops.

That last one matters more than it looks. A prototype proves the process works. A production agent needs authentication, monitoring and error handling at volume, and a good builder knows which of those to hand to a platform team rather than build alone.

Stretch: build a working prototype in Claude Code against a folder of made-up claims, contracts and delivery records.

You end with: an agent design, a set of evaluation cases, and optionally a prototype that passes them.

Turning it into a portfolio

Six exercises, six artefacts. Written up one page each, they show an employer the whole arc of the job: understanding a business, finding where AI changes a process, building the thing, and knowing its limits.

The skill includes a portfolio template. Three rules for it:

Label the fictional client as fictional. Presenting Hartwell as a real engagement is found out in the first follow-up question.

Show what went wrong. What Claude got wrong in exercise 1, what you missed in the rehearsal, what you changed after testing the skill. Candour about limits reads as experience.

Be ready to explain every decision. An interviewer will pick one choice in one artefact and ask why. If the answer is “Claude suggested it”, the artefact is working against you. The skill’s review mode gives you the questions an interviewer is likely to ask about each piece.