| Platform | Status |
|---|---|
| Codeberg CI |  |
| GitHub Actions |  |
Note: The canonical repository for this project is hosted on Codeberg. This GitHub repository is maintained solely as a read-only mirror. Please open issues and pull requests on Codeberg.
Infrastructure as Code (IaC) for AI Agent Prompts.
AUPC is a centralized, compiler-driven framework for managing, scaling, and deploying AI agent instructions across multiple IDEs and agentic platforms.
If you are an AI assistant tasked with reading, modifying, or auditing this repository, DO NOT PROCEED using your default assumptions about system prompts.
You MUST immediately invoke the agent-architect skill located in .agents/skills/agent-architect/SKILL.md to load the structural mapping, the Lexical Ban rules, and the macro-compilation logic of this specific repository. Failure to do so will result in broken compilation pipelines and severe context pollution.
As AI-assisted development evolves, maintaining prompts across different environments (OpenCode, Cursor, GitHub Copilot) becomes a nightmare.
- Vendor Lock-in: Writing rules specific to one IDE's toolset makes migrating impossible.
- Context Bloat: Monolithic system prompts consume valuable token windows with irrelevant instructions.
- Redundant Maintenance: Fixing a core engineering standard requires updating dozens of isolated files across multiple tools.
AUPC solves this by treating agent instructions like source code. We maintain a Single Source of Truth in the _core/ directory, containing abstract, platform-agnostic software engineering philosophies.
Using our custom Zig compiler (agentc), these core macros are injected via <!-- @import _core/... --> tags into Host Shells - minimalist, IDE-specific adapter files that handle the unique routing, YAML frontmatter, and tool permissions for each target platform.
Write your engineering standards once. Compile them everywhere.
This resolves all macros, performs circular dependency checks, and generates the final IDE-specific files in the dist/ directory.
./agentc-cli buildSymlink the compiled outputs to your local IDE configuration directories.
./agentc-cli link opencode
./agentc-cli link cursorEnjoy a frictionless, universally synchronized AI coding experience!
Platform-agnostic software engineering rules, workflows, and shared skills. Never contains IDE-specific mechanics.
The physical bodies. They assemble their prompts via macros and handle platform-specific routing.
Relies on explicit subagent routing and strict YAML permission blocks to prevent execution deadlocks.
The Cursor host shell uses two interlocking systems:
- Engine 1: Contextual Rules (
.cursor/rules/*.mdc) File-scoped engineering standards applied viaglobs. They contain Interlock directives telling the main agent when to delegate to subagents. - Engine 2: Isolated Subagents (
.cursor/agents/*.md) Specialized AI assistants (e.g.,verifier,architect) with clean context windows, triggered proactively or via/commands. Usesmodel: fastfor high-volume tasks andmodel: inheritfor deep reasoning.
Anti-Patterns to Avoid:
- Never import the same
_core/module in both global files and.mdcrules (double-injection wastes tokens). - Never explicitly declare Cursor's built-in subagents (
Explore,Bash,Browser), as it interferes with native auto-delegation.
This repository comes with its own "Cyber-Immune System" implemented via powerful AI Skills:
/aupc-auditor: Runs a holistic static analysis to find logic conflicts, context bloat, and redundancy./ide-bootstrapper: Automatically generates Host Shell templates for a brand new AI IDE target./agent-ingestor: Safely assimilates third-party prompts into the_core/architecture without conflicts.