Salesforce development with Claude Code: the powerful, cost-effective next step
We started with a visual editor, but moved on to Claude Code (in VS Code) β more powerful and more cost-effective. You delegate entire tasks and keep control yourself.
Where we grew to
We started with Cursor, but in early 2026 we moved on to Claude Code β for us, via the Claude Code extension in VS Code. Why? More power: instead of watching along step by step, you describe a whole job and have it carried out. In our experience it is also more cost-effective.
Claude Code feels less like 'typing together' and more like 'delegating to a skilled colleague'. You give an instruction, the work gets done, and you assess the result. With the same condition as always: do it safely.
What makes Claude Code so powerful
Claude Code is an AI assistant that picks up entire tasks on its own. You work with your own Salesforce structure, but instead of following every action, you delegate: you describe the goal and Claude Code carries it out β several steps in a row. In practice you use it to:
- βHave larger, connected jobs carried out in one go
- βHave existing solutions explained and improved in the same motion
- βHave complete documentation and overviews drawn up
- βHave recurring, error-prone work handled consistently
You remain the one in charge: you set the boundaries, Claude Code works within them, and you assess and decide what becomes final. That delegating is exactly where the power lies.
What it feels like in practice
This is what you learn in the training platform
The CRM Force training platform takes you through it step by step, with your own org connected and expert support within reach.
Go to the training platformThe golden rule: stay away from customer data
However powerful Claude Code is β never let it loose on your production org or real customer data. Precisely because it works independently, an isolated practice environment is essential: that's how you harness the power without the risk.
"Give a powerful assistant a safe playing field. An isolated practice environment without customer data is where you let Claude Code work."
- Only let Claude Code work in an isolated practice environment (scratch org, developer sandbox or Trailhead Playground)
- Never give it access to real customer or personal data
- Keep production entirely separate from what you have it carry out
- Have it propose first, assess it, and only then deliver deliberately
Because you give instructions within clear boundaries, you keep the power manageable β and the impact of a mistake stays within your practice environment.
What to watch when you put Claude Code to work
Delegating powerfully calls for a few deliberate choices up front:
- Choose an isolated environment without customer data to work in
- Set clear boundaries: what may and may not be touched
- Have larger jobs worked out as a proposal first, before anything becomes final
- Check the result before you adopt it β you remain ultimately responsible
We deliberately keep this functional. Setting Claude Code up concretely (including the Claude Code extension in VS Code) and building it safely into your work, you learn step by step in the CRM Force training platform.
Officially supported by Salesforce
Good news: this isn't a workaround. Since 2026, Salesforce offers official, secure connections (Salesforce calls them 'hosted MCP servers') that let Claude, including Claude Code, work directly and safely with your Salesforce environment. It's a supported part of the platform.
"Connecting Claude to your Salesforce is no longer an experiment, but an officially supported part of the platform."
And exactly as we recommend: Salesforce supports this connection for isolated practice environments (sandboxes) too. You or your admin decide which data and actions are available.
Privacy and security in practice
Especially with a tool that works on its own, security is a fixed method. Stick to these principles:
- No customer data β only let Claude Code loose on test or anonymized data
- Separate environment β production and your experiments stay strictly apart
- Always review β you assess the result before you apply it
- Protect your access β never share credentials or keys and keep them out of sight
- Limited access β Claude Code can never do more than you're allowed to; your permissions stay in charge
- You stay in charge β your admin decides up front which data and actions Claude Code may use; nothing is open by default
What it brings you
Those who work safely and purposefully with Claude Code notice the power quickly:
Delegate whole jobs
Describe a larger task and have it carried out in one go, while you assess the result.
Handle things consistently
Recurring, error-prone work gets done neatly and consistently β time after time.
Document completely
Have full documentation and overviews drawn up in the same motion.
More value for your money
In our experience you get a lot of power with Claude Code for a friendlier price.
What this looks like at a client
At one client we use Claude Code on their Salesforce environment: hundreds of flows, several systems exchanging data and a team that wants to know exactly what is running. Two working methods from that project can be adopted in any org.
The living data model
The first step is understanding how the objects fit together. For that we have heavyweight models (Claude Opus and Claude Fable) analyse the org: which objects exist, how they refer to each other, which fields form the key and which flows write to them. The result lands in a data model document that grows along with the project. Every statement in it has been checked against the org, or is marked as 'to be verified'.
The document is only useful if it stays current. That is why one fixed rule applies to Claude Code in this project:
- If an object or field turns up that is not yet in the data model, Claude asks questions first: what it is, how it relates to the rest and which system is leading for it
- Then Claude checks the answer against the org, because the org is leading and old notes are not
- Only then is it recorded in the data model and does the work continue
"Every question Claude asks about an unknown object is a gap in the documentation you would otherwise have found at delivery."
Client deliverables and project memory
The data model is not the only document that comes out of the project. The functional overview for the client (which processes there are and which automations belong to them) and the reference guide for administrators come from the same source. If something changes, the source is updated and the overview and the reference guide are generated again. That way they never drift apart.
The project also has a CLAUDE.md: a short file that Claude Code reads at the start of every session and that serves as project memory. It holds:
- Which environment is used and where the data model and the reference guide live
- The decisions that are settled, such as which old logic may be cleaned up and which integration is left alone
- The conventions, such as the agreement that every flow and every field gets a description in plain language, readable for non-technical colleagues too
- The rule that Claude asks questions first when it runs into an unknown object or field
A new session therefore starts with the same knowledge as the previous one. Nobody has to explain the context again, and a colleague taking over the project reads in one file how the work is done.
For consultancy clients, the knowledge gathered about their IT landscape also lands as a visual blueprint in the training platform. That way everyone in the organisation sees how systems, integrations and records fit together, and that knowledge does not stay with one person.
What that looks like
The knowledge becomes a picture everyone can look at
What Claude Code records along the way does not stay stuck in a text file. For consultancy clients it lands as a blueprint in the training platform: which systems exist, what runs between them, how a record comes into being and which automation drives the process. Everyone in the organisation can open it, including people who never touch Salesforce.
- The map above: systems, integrations and what is being built.
- Per record type the route, as it runs now and as it will run later.
- The data model in plain language, with field names in the side panel.
- The key flows, with a link through to the explanation per flow.
Example prompts to get started
These are the prompts from the working method above, stripped of client details. Copy one and adapt it to your org, in an isolated practice environment.
Map out the data model
Set up the project memory
One source, two documents
Check the data model against the org
DIY with Claude Code or guided? An honest comparison
Claude Code is powerful and cost-effective β but what does it take to set it up well and safely yourself, and what does guidance add?
DIY with Claude Code | CRM Force training platform | |
|---|---|---|
| Who it's for | Powerful if you dare to delegate, but you set everything up yourself | Admins, key users and consultants β guided, no figuring out |
| Getting started | Install Claude Code (in VS Code), connect it and arrange a practice org yourself | Connect your org; the starter pack and the data model are ready for you |
| Cost | More affordable than a premium editor, but you pay and manage it yourself | Included, with guidance and tools that genuinely move you forward |
| What you get back | The work gets done; assessing and making sense of it is up to you | Plain-language explanations that tell you what matters (Flow Coach, Case Builder) |
| When you get stuck | Figure it out and try things yourself | Personal expert support whenever you need it |
| Learning | Figuring it out yourself, by trial and error | A structured learning path with real-world cases |
Claude Code is a powerful, cost-effective choice β certainly if you want to delegate. But if you're an admin or key user who wants fast and safe results without figuring it all out yourself? Then guided is the shortest route.
Premium: this way of working, ready for your org
What we build by hand at a client, the Premium subscription of the training platform sets up for you as soon as your org is connected. The package is made for this way of working: you start with a filled project memory and a data model from your own org, instead of an empty file.
For working with Claude Code
- βA Claude Code starter pack per org: CLAUDE.md, skills and MCP configuration, filled with the objects, flows and conventions from your own org
- βThe Implementation Guide with Claude Code as the main route, step by step and functional
- βReadable flow descriptions via Flow Coach: what a flow does and why, in plain language
- βA convention score in Flow Overview, so you see per object where naming and descriptions are lagging behind
- βA data model with relationships from the org scan, as a starting point for your own living data model
- βAn overview of the differences since the previous sync: which flows, objects and fields were added or changed
What else Premium holds
- βAI explanations in Flow Coach for all your flows
- βApex Analyzer for existing code
- βCreating new cases with AI in the Case Builder
- β4 hours of expert support per month
The full comparison of Basis and Premium is on the training platform page.
Frequently asked questions
No. You describe in plain language what you want done; Claude Code carries out the job. You assess the result and decide what becomes final β delegating, not programming.
We started with Cursor ourselves β pleasant and visual. But Claude Code gives you more power (delegating entire tasks) and in our experience is more cost-effective. For anyone wanting to go a step further, that's worth the switch.
Yes, as long as you only let Claude Code loose on an isolated practice environment without customer data β never on production. Precisely because it works on its own, that's extra important.
An isolated practice environment, such as a scratch org, developer sandbox or Trailhead Playground. There you can let Claude Code do its work freely.
This article shows the why and the what. Learning to use Claude Code powerfully and safely you do with guidance, step by step, in the CRM Force training platform.
A document that describes which objects your org has, how they fit together and which system is leading per field, and that grows along with every change. You maintain it because Claude Code draws its context from it: if an object is in there, you do not have to explain it again; if it is not, Claude asks questions first. That way the documentation stays correct without separate effort.
Premium sets up the starting position we build by hand at a client: a CLAUDE.md, skills and MCP configuration filled from your own org, a data model with relationships from the org scan and an overview of what has changed since the previous sync. On top of that you get Flow Coach, Apex Analyzer, AI cases and 4 hours of expert support per month.
Related topics
Go deeper into AI, automation and the CRM Force training platform:
- Salesforce development with Cursor
- Flow Coach: understand every Salesforce Flow
- Cases and the Case Builder
- Connect your Salesforce org
- Automating with Salesforce Flow
- Apex for admins and analysts
- Agentforce Vibes: opportunities and risks
- AI and ChatGPT in Salesforce
- The role of AI for the Salesforce admin
- Expert support for Salesforce
- The CRM Force training platform
- Tailored Salesforce training
This is a preview β here's how you get the most out of it
You saw the why and the what. In the CRM Force training platform you do it yourself: powerful, guided and with your own Salesforce environment connected.
- βConnect your own org and practice safely with your real automations
- βFlow Coach & Flow Overview: understand and improve your flows in plain language
- βCase Builder: from a real-world question to concrete advice
- βPersonal guidance and expert support whenever you get stuck
Rather have it built than develop it yourself?
Salesforce in plain language. Finally.
- 14-day free trial
- No credit card required
- Expert support included
- Connect your Salesforce org instantly

DIY with Claude Code
CRM Force training platform