How an experienced developer teamed up with Claude to create Elo programming language
(2026/01/24)
- Reference: 1769264049
- News link: https://www.theregister.co.uk/2026/01/24/human_ai_pair_programming_elo/
- Source link:
feature Bernard Lambeau, a Belgium-based software developer and founder of several technology companies, created a programming language called Elo with the help of Anthropic's Claude Code.
Starting on December 25, 2025, he published a series of posts about the project. The first post names Claude as a co-author.
"In roughly 24 hours of collaboration, we built a complete expression language with a parser, type system, three compilers, a standard library, a CLI tool, and a documentation website. Not bad for a day's work,” Lambeau and Claude [1]wrote .
[2]
" [3]Elo isn't just a demonstration that AI can write code. It's a demonstration that humans and AI can build together – each contributing what they do best,” they added.
[4]
[5]
As an expression language that compiles to JavaScript, Ruby, and SQL, Elo is intended as a portable way to handle form validation, e-commerce order processing, and subscription logic.
Lambeau, founder and CTO of [6]Klaro Cards and CEO of app consultancy [7]Enspirit , is not the first to develop a programming language with the help of AI.
[8]
Steve Klabnik [9]performed a similar feat last year with the Rue programming language. In September 2025, Geoffrey Huntley [10]enlisted Claude to write a programming language called [11]Cursed . And before that, Avital Tamir published a Claude-authored repo for the [12]Server programming language, with the caveat that the code is not intended for actual use.
Claude Code [13]isn't the only AI-assisted programming method having a moment . AI biz Cursor [14]created a rudimentary browser using OpenAI's GPT-5.2. And developer Ola Prøis used Cursor, powered by Claude, to create a Rust-based text editor called [15]Ferrite .
Claude users generally acknowledge that their pair partner makes mistakes. But those committed to AI assistance find it worthwhile to clean up after their helper.
[16]
"Claude Code knows almost every tech stack (and can search the web), knows the Linux commands that matter (search code, search & replace, compile, test, etc.), and does that 10x faster than I can do myself," Lambeau told The Register in an email interview.
Claude, he said, allows him to use technology he hasn't mastered.
"I was already a full-stack developer (on languages, frameworks & reusable libraries I knew); I'm now a full-stack++ dev because I can also use languages, frameworks, and reusable libraries I barely know, if at all," he explained.
"Claude Code falls short if you don't have a great methodology. It needs feedback loops to work fine; otherwise, it derails. One possible feedback loop is a human reviewing code and testing manually. But there's a better/complementary approach if you want it to work autonomously. On both Elo and Bmg.js, I've started by making sure the testing methodology was effective and scientifically sound. Claude writes the tests, executes them, discovers where it's wrong, and corrects itself. Impressive."
[17]AI hasn't delivered the profits it was hyped for, says Deloitte
[18]Anthropic writes 23,000-word 'constitution' for Claude, suggests it may have feelings
[19]AI conference's papers contaminated by AI hallucinations
[20]Cursor used agents to write a browser, proving AI can write shoddy code at scale
Lambeau said he still needs to review some of Claude's output.
"But if I read the tests, agree with them, and can check myself that they run fine, I'm 95 percent sure it's already correct as a black box (not even reading the code)," he explained. "Then I can check the architecture and code quality as a white box by having a general look at the code, but I don't have to understand every detail."
Notably, Lambeau documented the series of prompts he used to create the language. The repo includes [21]more than 100 tasks used to direct the AI model. In addition, Lambeau has published [22]a video that describes his AI pair programming process.
"I started in a setting where Claude Code asked for permissions every 20 seconds and I was checking everything it did," Lambeau explained. "After a few successes, I quickly set up safe environments to be able to let Claude Code run in full autonomy (isolated computer & isolated Linux user, or running in a Docker image)."
Lambeau said he still uses plan mode for complex tasks that require conversation with Claude.
"I review the plan, make sure we have a test strategy that's sound, then switch Claude to autonomous mode and look at the tests, code & results afterward," he said. "That's very similar to a lead-dev/CTO + QA role, btw; it's just much faster than with human devs."
Lambeau, who has a PhD in software engineering and 30 years of experience as a developer, said both experts and novices can benefit from Claude Code, though he added that a service like [23]Lovable might be more approachable for those not already acclimated to the command line.
"Now, when it comes to real software/product engineering, I think Claude Code requires experts (so far)," he said. "You still need to guide it a lot to keep the quality high enough. You need very strong expertise to do it effectively. Currently (Claude will still improve a lot), if you don't have the expertise, you certainly end up with a big mess of unmaintainable code."
Many developers have said as much about AI tools. They're more useful as an amplifier of expertise than as a replacement for it. The situation is analogous to the introduction of sequencing software, digital synthesizers, and drum machines half a century ago. These tools enabled a lot of people who weren't great musicians to make music. But they didn't instill musical skill, and they produced the most interesting work in the hands of practiced musicians.
The cost to do this, Lambeau said, has been a Claude Max subscription that he purchased in December for €180 a month. In that time, he says, he wrote Elo ( [24]https://elo-lang.org ), completed Bmg.js ( [25]https://github.com/enspirit/bmg.js ), completed Bmg's documentation ( [26]https://www.relational-algebra.dev ), and created the first version of the Try page ( [27]https://www.relational-algebra.dev/try ).
"It's all personal research and open-source projects," he said. "It would have required several weeks to do the same manually myself, and several months to ask another developer to do it. The cost would be mostly because of the scientific & technical knowledge transfer about the data language I envision. Strangely enough, it's very cheap with Claude Code. There's something true about the fact that those LLMs have a PhD."
Lambeau explained that Elo isn't just a way to test Claude Code. He also sees it as an extension of his academic work in software engineering and his personal interest in the Relational Model – he's served as a lecturer for database courses at Belgium’s UCLouvain.
"I'm absolutely convinced that we need better/safer/simpler programming languages inside no-code tools and when interconnecting them (e.g. Zapier, Make, n8n, etc.)," he said. "Mainstream programming languages are very complex, error-prone, sometimes dangerous, and the programs are difficult to review for non-experts."
"More importantly, they are cumbersome to use for even simple data tasks. I mean, even validating the schema and constraints of a data file at runtime tends to be a nightmare in existing languages. It's not built-in in any mainstream language; you immediately need validation libraries; most of them are limited in what they can easily check, so you need to add dedicated boilerplate code."
In a world where non-technical people will have the opportunity to write untrustworthy code with the help of AI, he said, we need to be able to run that code safely.
"Elo aims at providing a safe & simple alternative," he said. "It will be a limited language (non-Turing-complete, as we say) but super safe & simple, and usable in 80 percent of common data use cases. The very first no-code tool to integrate it will be Klaro Cards, of course." ®
Get our [28]Tech Resources
[1] https://elo-lang.org/blog/20241225-building-elo-with-claude/
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://github.com/enspirit/elo
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://www.klaro.cards/en/
[7] https://enspirit.co/
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[9] https://www.theregister.com/2026/01/03/claude_copilot_rue_steve_klabnik/
[10] https://ghuntley.com/cursed/
[11] https://cursed-lang.org/
[12] https://github.com/AvitalTamir/sever
[13] https://build.ms/2026/1/19/the-claude-code-moment/
[14] https://www.theregister.com/2026/01/22/cursor_ai_wrote_a_browser/
[15] https://github.com/OlaProeis/Ferrite
[16] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[17] https://www.theregister.com/2026/01/21/deloitte_enterprises_adopting_ai_revenue_lift/
[18] https://www.theregister.com/2026/01/22/anthropic_claude_constitution/
[19] https://www.theregister.com/2026/01/22/neurips_papers_contaiminated_ai_hallucinations/
[20] https://www.theregister.com/2026/01/22/cursor_ai_wrote_a_browser/
[21] https://github.com/enspirit/elo/blob/main/.claude/tasks/done/0000-first-conversation.md
[22] https://www.loom.com/share/ef65c6ebfc35491783ff0a6067447dd4
[23] https://lovable.dev/
[24] https://elo-lang.org
[25] https://github.com/enspirit/bmg.js
[26] https://www.relational-algebra.dev
[27] https://www.relational-algebra.dev/try
[28] https://whitepapers.theregister.com/
Starting on December 25, 2025, he published a series of posts about the project. The first post names Claude as a co-author.
"In roughly 24 hours of collaboration, we built a complete expression language with a parser, type system, three compilers, a standard library, a CLI tool, and a documentation website. Not bad for a day's work,” Lambeau and Claude [1]wrote .
[2]
" [3]Elo isn't just a demonstration that AI can write code. It's a demonstration that humans and AI can build together – each contributing what they do best,” they added.
[4]
[5]
As an expression language that compiles to JavaScript, Ruby, and SQL, Elo is intended as a portable way to handle form validation, e-commerce order processing, and subscription logic.
Lambeau, founder and CTO of [6]Klaro Cards and CEO of app consultancy [7]Enspirit , is not the first to develop a programming language with the help of AI.
[8]
Steve Klabnik [9]performed a similar feat last year with the Rue programming language. In September 2025, Geoffrey Huntley [10]enlisted Claude to write a programming language called [11]Cursed . And before that, Avital Tamir published a Claude-authored repo for the [12]Server programming language, with the caveat that the code is not intended for actual use.
Claude Code [13]isn't the only AI-assisted programming method having a moment . AI biz Cursor [14]created a rudimentary browser using OpenAI's GPT-5.2. And developer Ola Prøis used Cursor, powered by Claude, to create a Rust-based text editor called [15]Ferrite .
Claude users generally acknowledge that their pair partner makes mistakes. But those committed to AI assistance find it worthwhile to clean up after their helper.
[16]
"Claude Code knows almost every tech stack (and can search the web), knows the Linux commands that matter (search code, search & replace, compile, test, etc.), and does that 10x faster than I can do myself," Lambeau told The Register in an email interview.
Claude, he said, allows him to use technology he hasn't mastered.
"I was already a full-stack developer (on languages, frameworks & reusable libraries I knew); I'm now a full-stack++ dev because I can also use languages, frameworks, and reusable libraries I barely know, if at all," he explained.
"Claude Code falls short if you don't have a great methodology. It needs feedback loops to work fine; otherwise, it derails. One possible feedback loop is a human reviewing code and testing manually. But there's a better/complementary approach if you want it to work autonomously. On both Elo and Bmg.js, I've started by making sure the testing methodology was effective and scientifically sound. Claude writes the tests, executes them, discovers where it's wrong, and corrects itself. Impressive."
[17]AI hasn't delivered the profits it was hyped for, says Deloitte
[18]Anthropic writes 23,000-word 'constitution' for Claude, suggests it may have feelings
[19]AI conference's papers contaminated by AI hallucinations
[20]Cursor used agents to write a browser, proving AI can write shoddy code at scale
Lambeau said he still needs to review some of Claude's output.
"But if I read the tests, agree with them, and can check myself that they run fine, I'm 95 percent sure it's already correct as a black box (not even reading the code)," he explained. "Then I can check the architecture and code quality as a white box by having a general look at the code, but I don't have to understand every detail."
Notably, Lambeau documented the series of prompts he used to create the language. The repo includes [21]more than 100 tasks used to direct the AI model. In addition, Lambeau has published [22]a video that describes his AI pair programming process.
"I started in a setting where Claude Code asked for permissions every 20 seconds and I was checking everything it did," Lambeau explained. "After a few successes, I quickly set up safe environments to be able to let Claude Code run in full autonomy (isolated computer & isolated Linux user, or running in a Docker image)."
Lambeau said he still uses plan mode for complex tasks that require conversation with Claude.
"I review the plan, make sure we have a test strategy that's sound, then switch Claude to autonomous mode and look at the tests, code & results afterward," he said. "That's very similar to a lead-dev/CTO + QA role, btw; it's just much faster than with human devs."
Lambeau, who has a PhD in software engineering and 30 years of experience as a developer, said both experts and novices can benefit from Claude Code, though he added that a service like [23]Lovable might be more approachable for those not already acclimated to the command line.
"Now, when it comes to real software/product engineering, I think Claude Code requires experts (so far)," he said. "You still need to guide it a lot to keep the quality high enough. You need very strong expertise to do it effectively. Currently (Claude will still improve a lot), if you don't have the expertise, you certainly end up with a big mess of unmaintainable code."
Many developers have said as much about AI tools. They're more useful as an amplifier of expertise than as a replacement for it. The situation is analogous to the introduction of sequencing software, digital synthesizers, and drum machines half a century ago. These tools enabled a lot of people who weren't great musicians to make music. But they didn't instill musical skill, and they produced the most interesting work in the hands of practiced musicians.
The cost to do this, Lambeau said, has been a Claude Max subscription that he purchased in December for €180 a month. In that time, he says, he wrote Elo ( [24]https://elo-lang.org ), completed Bmg.js ( [25]https://github.com/enspirit/bmg.js ), completed Bmg's documentation ( [26]https://www.relational-algebra.dev ), and created the first version of the Try page ( [27]https://www.relational-algebra.dev/try ).
"It's all personal research and open-source projects," he said. "It would have required several weeks to do the same manually myself, and several months to ask another developer to do it. The cost would be mostly because of the scientific & technical knowledge transfer about the data language I envision. Strangely enough, it's very cheap with Claude Code. There's something true about the fact that those LLMs have a PhD."
Lambeau explained that Elo isn't just a way to test Claude Code. He also sees it as an extension of his academic work in software engineering and his personal interest in the Relational Model – he's served as a lecturer for database courses at Belgium’s UCLouvain.
"I'm absolutely convinced that we need better/safer/simpler programming languages inside no-code tools and when interconnecting them (e.g. Zapier, Make, n8n, etc.)," he said. "Mainstream programming languages are very complex, error-prone, sometimes dangerous, and the programs are difficult to review for non-experts."
"More importantly, they are cumbersome to use for even simple data tasks. I mean, even validating the schema and constraints of a data file at runtime tends to be a nightmare in existing languages. It's not built-in in any mainstream language; you immediately need validation libraries; most of them are limited in what they can easily check, so you need to add dedicated boilerplate code."
In a world where non-technical people will have the opportunity to write untrustworthy code with the help of AI, he said, we need to be able to run that code safely.
"Elo aims at providing a safe & simple alternative," he said. "It will be a limited language (non-Turing-complete, as we say) but super safe & simple, and usable in 80 percent of common data use cases. The very first no-code tool to integrate it will be Klaro Cards, of course." ®
Get our [28]Tech Resources
[1] https://elo-lang.org/blog/20241225-building-elo-with-claude/
[2] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[3] https://github.com/enspirit/elo
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[6] https://www.klaro.cards/en/
[7] https://enspirit.co/
[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[9] https://www.theregister.com/2026/01/03/claude_copilot_rue_steve_klabnik/
[10] https://ghuntley.com/cursed/
[11] https://cursed-lang.org/
[12] https://github.com/AvitalTamir/sever
[13] https://build.ms/2026/1/19/the-claude-code-moment/
[14] https://www.theregister.com/2026/01/22/cursor_ai_wrote_a_browser/
[15] https://github.com/OlaProeis/Ferrite
[16] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/devops&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aXT6rgAQanmuuJtwtrI9ugAAAYk&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[17] https://www.theregister.com/2026/01/21/deloitte_enterprises_adopting_ai_revenue_lift/
[18] https://www.theregister.com/2026/01/22/anthropic_claude_constitution/
[19] https://www.theregister.com/2026/01/22/neurips_papers_contaiminated_ai_hallucinations/
[20] https://www.theregister.com/2026/01/22/cursor_ai_wrote_a_browser/
[21] https://github.com/enspirit/elo/blob/main/.claude/tasks/done/0000-first-conversation.md
[22] https://www.loom.com/share/ef65c6ebfc35491783ff0a6067447dd4
[23] https://lovable.dev/
[24] https://elo-lang.org
[25] https://github.com/enspirit/bmg.js
[26] https://www.relational-algebra.dev
[27] https://www.relational-algebra.dev/try
[28] https://whitepapers.theregister.com/
Re: Someone with appropraite skills manages to use 'AI' in useful way !!!
LessWileyCoyote
Also if you use AI to drastically reduce the workforce now, you run the risk in future of not having a supply of people with enough experience to know how to use the tool safely and productively.
Cue more recruiters searching for 18-year-olds with 20 years' experience.
Paul Crawford
It seems a common thing for software engineers to try and "solve" today's problems by inventing yet another new language, rather than creating tools that can help fix/mitigate the risks in the already written code that so much of the world now depends upon.
Reflections on Trusting Trust
b0llchit
"In roughly 24 hours of collaboration, we built a complete expression language with a parser, type system, three compilers, a standard library, a CLI tool, and a documentation website. Not bad for a day's work,”
I think everybody must (re-)read: [1]Reflections on Trusting Trust by Ken Thompson (yeah, that Ken) and ask yourself who and what you could/would/should trust.
[1] https://dl.acm.org/doi/pdf/10.1145/358198.358210?download=true
@b0llchit - Recommended reading?
William Towle
> I think everybody must (re-)read: Reflections on Trusting Trust by Ken Thompson [...]
...not least because AI already has. [1]--> /o\
[1] https://www.theregister.com/2025/09/29/when_ai_is_trained_for/
Re: @b0llchit - Recommended reading?
b0llchit
I rest my case.
Someone with appropraite skills manages to use 'AI' in useful way !!!
"Lambeau, who has a PhD in software engineering and 30 years of experience as a developer"
Says everything you need to know !!!
'AI' in the hands of 'Joe Public' will not be creating any new language 'soon' ...
All these exercises do is to confirm what we already know ... 'AI' is not a tool that is useful to the general public, it requires skills to 'guide' it to give meaningful output. Those skills need to include some knowledge about the 'thing' you are querying the 'AI about.
'AI' is not a black box that gives 'right' answers to general questions.
It 'may' sometimes BUT as you cannot predict when this happens you need to have some method to cross-check the output against 'reality' and 'truth'.
:)