What did Charles build with Claude — and how fast?
Charles scaffolded a full CRM — contacts, leads, transactions, and pipelines — using Claude in a single afternoon in September 2025. He used no developer background, no syntax, and no configuration files. Claude Artifacts handled all code generation from a plain-language prompt, as Charles reported on iCharles.
That same build had taken him $50,000 and five developers across multiple years when attempted with Salesforce.
How does Claude generate a CRM without coding?
Claude Artifacts is Anthropic's code-generation feature that produces a working interface from a plain-language description. You describe what you need — contacts, deals, pipelines — and Claude generates the entire front end in a single prompt.
No developer background is required. No syntax. No configuration files. The user describes the product; Claude writes the code.
This is the core mechanic behind what Charles calls building a Salesforce replacement with Claude.
What was the real problem Charles ran into?
Generating the interface and wiring it to a real database are two completely different problems. Claude solves the first one by default. It only solves the second if you ask specifically.
You might also like
Charles discovered this the hard way. He threw away three full builds over roughly five months after finding that nothing was persisting — the UI looked right, but no data was actually being saved.
By Day 307 of his live coding run, he was still rebuilding, working 12 hours a day.
How does this compare to the Salesforce build?
Here is the before-and-after from Charles's own account:
| Factor | Salesforce Build | Claude Build |
|---|---|---|
| Cost | ~$50,000 | $0 (code generation) |
| Developers involved | 5 | 0 |
| Time to scaffold | Multiple years | One afternoon (September 2025) |
| Data persistence | Working | Required separate, explicit prompting |
| Builds thrown away | 0 reported | 3 over ~5 months |
The $50,000 figure is Charles's personal estimate based on what he paid across five developers over multiple years, per his own account.
Why did Charles keep rebuilding after the first working version?
The first build looked functional. The interface showed contacts and pipelines. But the data was not saving to any real database. When Charles discovered this, he started over.
He repeated this cycle three times across roughly five months. Each time, the UI appeared correct. Each time, persistence was missing until he asked Claude to solve that problem explicitly.
As of Day 307 of his live coding run, he was still in an active rebuild at 12 hours a day.
What does this tell builders about AI-generated software?
Here's what we know so far: the gap between a working interface and a working application is exactly where AI-assisted builds break down. Claude can scaffold the front end fast. The database layer requires a separate, deliberate prompt.
This pattern matters for anyone using Claude or similar tools to replace SaaS products. The Anthropic Pentagon Claude injunction dispute and Microsoft's $2.5B AI implementation unit both point to the same underlying question enterprises are asking: where does AI-generated software actually hold up under production conditions?
Charles's five-month rebuild log is one of the most detailed public answers to that question from a non-developer builder.
What tools and approach did Charles use?
- Claude (Anthropic's AI model) — used to generate all code via plain-language prompts
- Claude Artifacts — the specific feature that handles code generation and renders the interface
- No external developers — Charles built alone after the initial scaffold
- Live coding sessions — documented publicly, reaching at least Day 307
- Iterative rebuilds — three full builds discarded before a stable version was in progress
The Salesforce replacement build is documented in detail on iCharles.
What is the current status of the build?
As of Day 307 of Charles's live coding run, the CRM was still being rebuilt. Charles was working 12 hours a day on the project. No completion date is stated in the source material.
The three prior builds were discarded entirely. The current build is the fourth attempt.
0 Comments
Log in to comment
Not a member yet? Join the community
Pick a meme
KlipyHave a great take?
Drop your email — we'll send a magic link so you can post it. No password.
Not a member of the community? Join today.
Join the community →