The Problem
Coding agents (Claude Code, Codex, Cursor) default to short-horizon, task-by-task behavior. They lack a persistent architecture for long-running, multi-month work across software engineering, research, and operations. Users get competent but shallow results — no memory model, no verification loop, no compounding improvement.
What This Does
The repository is a single, unusually long system prompt (README.md) plus an architecture diagram (most_capable_agent_system_architecture.svg). Pasting the prompt into a coding agent instructs it to build a self-improving "agentic operating system" rather than answer as an assistant.
The prompt enforces a reader contract: read design principles first, create a local operating summary, ask only blocking questions, and bias toward writing files over producing strategy essays. The first milestone is a closed loop — task in, verified result out — before any breadth.
How It Is Wired
There is no code to trace. The repository is a prompt and a diagram; the "wiring" is the prompt's internal structure:
- Entry point: paste the prompt into any agent's system prompt,
CLAUDE.md, or first message. - Control flow: the prompt names a specific reading order —
NON-NEGOTIABLE DESIGN BETS,RELIABILITY MATH AND HARNESS ENGINEERING,RECOMMENDED DEFAULT IMPLEMENTATION CHOICES,BUILD ORDER,FIRST MILESTONE DEFINITION,NON-NEGOTIABLE RULES,INITIAL ACTIONS YOU MUST TAKE NOW— then instructs the agent to scaffold immediately. - Effects: the only file system effect is the agent writing a local operating summary file for itself, per the prompt's instructions. The runtime effects (what the agent builds) are determined by the agent, not by this repo.
- File map:
README.mdcontains the prompt and quick-start;most_capable_agent_system_architecture.svgvisualizes the target architecture.
There is no module graph, no call graph, no hub-and-spoke structure to analyze — the repository's entire surface is the prompt text itself.
How To Use It
Setup: none. No dependencies, no build step, no configuration files.
Running it: copy the prompt from README.md and paste it into your agent. The README documents this verbatim:
- Copy the prompt below
- Paste it into your agent (system prompt,
CLAUDE.md, or first message) - It starts building immediately
The README lists supported targets: Claude Code, OpenAI Codex, Cursor, Antigravity, OpenClaw, OpenCode, OpenHands, Claude Agent SDK.
Real-World Use
A team standardizing agent behavior across its engineering staff pastes this prompt into each developer's CLAUDE.md. Every agent then scaffolds the same architecture: a transparent state model, a verification harness, and a build order. The team gets consistent agent behavior and a shared vocabulary for agent capability — without maintaining a codebase of their own.
Code Health & Issues
Static analysis found the following (heuristic, verify before acting):
- Medium/SDLC – No test files detected – repository-wide. There is no code to test; the artifact is a prompt.
- Medium/SDLC – No CI/CD pipeline detected – no
.github/or CI config. Again, nothing to build or gate. - Medium/SDLC – No LICENSE file – root. Usage and redistribution rights are unclear. The source repo (
fainir/most-capable-agent-system-prompt, 866 stars) may carry a license that applies here, but it is not present in this clone.
These findings are structural, not defects. The repo is a prompt, not a program; the meaningful risks are licensing and prompt maintenance, not test coverage.
The Bottom Line
A well-structured, opinionated prompt that fills a real gap: agents that build systems rather than answer questions. The trade-off is that value depends entirely on the target agent's capabilities — this repo provides instructions, not guarantees. Use it if you want a starting point for agentic scaffolding; verify the license before commercial use.