| _version_ | 1866902025882566656 |
|---|---|
| author | Gomes, Renato Aparecido |
| author_facet | Gomes, Renato Aparecido |
| contents | In May 2026, an estimated four million software developers routinely configure the behavior of AI coding assistants through text files: CLAUDE.md, .cursorrules, AGENTS.md, copilot-instructions.md. They specify how the agent should communicate, what architectural patterns it should prefer, what ethical constraints it must respect, and what domain knowledge it should treat as authoritative. None of these files contain a single line of code. None require compilation. None require technical expertise beyond writing prose.<br><br>This phenomenon has no consensual name. The communities around Claude Code, Cursor IDE, and GitHub Copilot have each developed conventions independently, without a shared theoretical framework. We propose the name Cognitive Metaprogramming (CMP) for this phenomenon, and we offer its first systematic theoretical treatment.<br><br>In classical computing, metaprogramming denotes programs that generate or modify other programs — code that operates on code. CMP generalizes this: artifacts in natural language that configure the reasoning space of AI agents. The "meta" level is no longer code; it is prose. The "compilation" is no longer syntactic; it is cognitive. And crucially, the entry barrier has dropped from "can write code" to "can write" — a democratization without historical precedent in the history of programming.<br><br>The thesis of this paper: CMP is not merely a collection of useful tricks for AI practitioners. It is an emergent layer of computational infrastructure — the cognitive equivalent of what Infrastructure as Code was for server configuration in 2011. The field is at an inflection point: fragmented practice is accumulating, but theory, vocabulary, and shared infrastructure are absent. This paper provides the theoretical foundation.<br><br>Claims (pre-stated, falsifiable):<br>(1) A four-category taxonomy — Categories A (Agent Instruction), B (Product Specification), C (Skill Artifact), and D (CMP Primitive) — exhaustively covers the artifact landscape of mid-2026.<br>(2) A seven-dimension grammar covering audience, lifecycle, content_type, propagation, compression, version, and distribution provides complete descriptive coverage of structural variation.<br>(3) A five-level CMP Maturity Scale, analogous to CMMI, provides a falsifiable assessment tool with observable, non-overlapping criteria. |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_20355372 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Cognitive Metaprogramming: A Grammar for Agent-Readable Artifacts Gomes, Renato Aparecido Cognitive Metaprogramming CMP AGENTS.md CLAUDE.md agent configuration natural language programming AI governance soft programming metaprogramming agent-readable artifacts AIOX OSF pre-registration CMP Maturity Scale seven-dimension grammar In May 2026, an estimated four million software developers routinely configure the behavior of AI coding assistants through text files: CLAUDE.md, .cursorrules, AGENTS.md, copilot-instructions.md. They specify how the agent should communicate, what architectural patterns it should prefer, what ethical constraints it must respect, and what domain knowledge it should treat as authoritative. None of these files contain a single line of code. None require compilation. None require technical expertise beyond writing prose.<br><br>This phenomenon has no consensual name. The communities around Claude Code, Cursor IDE, and GitHub Copilot have each developed conventions independently, without a shared theoretical framework. We propose the name Cognitive Metaprogramming (CMP) for this phenomenon, and we offer its first systematic theoretical treatment.<br><br>In classical computing, metaprogramming denotes programs that generate or modify other programs — code that operates on code. CMP generalizes this: artifacts in natural language that configure the reasoning space of AI agents. The "meta" level is no longer code; it is prose. The "compilation" is no longer syntactic; it is cognitive. And crucially, the entry barrier has dropped from "can write code" to "can write" — a democratization without historical precedent in the history of programming.<br><br>The thesis of this paper: CMP is not merely a collection of useful tricks for AI practitioners. It is an emergent layer of computational infrastructure — the cognitive equivalent of what Infrastructure as Code was for server configuration in 2011. The field is at an inflection point: fragmented practice is accumulating, but theory, vocabulary, and shared infrastructure are absent. This paper provides the theoretical foundation.<br><br>Claims (pre-stated, falsifiable):<br>(1) A four-category taxonomy — Categories A (Agent Instruction), B (Product Specification), C (Skill Artifact), and D (CMP Primitive) — exhaustively covers the artifact landscape of mid-2026.<br>(2) A seven-dimension grammar covering audience, lifecycle, content_type, propagation, compression, version, and distribution provides complete descriptive coverage of structural variation.<br>(3) A five-level CMP Maturity Scale, analogous to CMMI, provides a falsifiable assessment tool with observable, non-overlapping criteria. |
| title | Cognitive Metaprogramming: A Grammar for Agent-Readable Artifacts |
| topic | Cognitive Metaprogramming CMP AGENTS.md CLAUDE.md agent configuration natural language programming AI governance soft programming metaprogramming agent-readable artifacts AIOX OSF pre-registration CMP Maturity Scale seven-dimension grammar |
| url | https://doi.org/10.5281/zenodo.20355372 |