Manuel Odendahl is a software engineer, open-source developer, and writer whose work gives programmers more expressive, controllable ways to build with language models. He created the GO GO GOLEMS family of developer tools and argues that AI systems work best when they generate executable code, transform information between useful formats, and leave consequential decisions open to human intervention.
Odendahl’s early work combined embedded programming and electronic music. He developed Mididuino, a framework for programmable MIDI devices, and documented how testing, rapid prototyping, automated builds, and deployment could be adapted to the constrained hardware of the MidiCommand controller. He subsequently worked across back-end development, databases, search, e-commerce, and web applications. His sustained interest in Common Lisp and interactive development shaped his preference for extensible languages and software that can construct its own abstractions.
By March 2023, Odendahl was arguing that language models would change software engineering beyond code completion, influencing architecture, documentation, collaboration, and team workflows. He developed those ideas through GO GO GOLEMS: Geppetto, a Go framework for language-model applications; Pinocchio, a command-line tool for reusable prompting workflows; and Glazed, a framework for structured-data command-line applications. His guide to custom AI command-line tools shows how templates, declarative commands, and project-specific context can turn repeated engineering tasks into reusable utilities. In 2024, he was a principal engineer affiliated with The Tree Center.
- Language models as translation engines. Odendahl breaks development into transformations between useful representations: meeting transcripts become technical summaries, action items, GitHub issues, and scripts; existing code becomes documentation, onboarding material, or targeted context. Each intermediate artifact remains inspectable and useful to people.
- Domain-specific languages for sharper context. He uses compact, purpose-built formats to expose the structure models need: network captures can become YAML descriptions of relevant traffic, while large repositories can be reduced to function signatures, documentation, and selected interfaces. His AI-assisted programming workshop also shows how simulated application interfaces and adversarial fictional code reviews can uncover requirements and security problems before implementation.
- Executable code over rigid tool calling. Odendahl argues that predefined agent tools can proliferate unnecessarily, repeat existing context, and return oversized results. In his critique of conventional MCP integrations, a Go-hosted JavaScript sandbox with SQLite and a web-server interface lets a model inspect schemas, filter data, save reusable functions, register REST endpoints, and assemble an interactive CRM.
- Human-editable agent interactions. He advocates interfaces that let users correct tool arguments, trim returned data, adjust filters, and rerun individual operations. Those controls preserve human judgment while avoiding expensive, opaque agent workflows that offer only approval or rejection.
