10
Claude Code Principles
What the Research
Actually Says.
Most AI coding advice is measurably wrong. This series distills 17 peer-reviewed papers into 10 actionable principles for production-grade agentic workflows — plus the two open-source tools that make them automatic.
17
Research Papers
10
Principles
2
Open-Source Tools
The 10 Principles
Read the Series Overview →
Act I
Foundations
01
The Hardening Principle
Every fuzzy LLM step that must behave identically every time must eventually be replaced by a deterministic tool.
02
The Context Hygiene Principle
Context is your scarcest resource. Treat it like memory in an embedded system, not disk space on a server.
03
The Living Documentation Principle
Documentation is context. Stale documentation is poisoned context.
Act II
Execution Discipline
04
The Disposable Blueprint Principle
Never implement without a saved, versioned plan artifact. And never fall in love with one.
05
The Institutional Memory Principle
When an agent makes a mistake, don't just correct it — codify it forever.
06
The Specialized Review Principle
A generalist reviewer trends toward the median. Specialists find what generalists can't.
Act III
Governance & Safety
07
The Observability Imperative
If you can't see inside your pipeline, you're trusting it on faith.
08
The Strategic Human Gate Principle
Rubber-stamp approval is the single most common quality failure in multi-agent systems.
09
The Token Economy Principle
Tokens are money. Most people are burning it.
The Tools
Forge
Science-backed agent assembly. Vocabulary routing, PRISM identities,
MAST anti-patterns, the 45% threshold, and the cascade pattern —
all encoded into a tool that builds specialist agent teams from a goal description.
claude plugin marketplace add jdforsythe/forge
claude plugin install forge
View on GitHub →
claude plugin install forge
jig
Intelligent Context Utilization. Define session profiles that declare
exactly which tools, MCP servers, and skills activate per project.
Check profiles into git for reproducible context across the team.
brew tap jdforsythe/jig
brew install jig
View on GitHub →
brew install jig
The Research
17 Sources. 9 Years of Science.
From “Attention Is All You Need” (Vaswani et al., 2017) to DeepMind’s multi-agent scaling laws (2025). Every principle traces to published research. Every design decision in the tools traces to published research. The author’s contribution is synthesis and implementation. The science belongs to the researchers.
View Complete Bibliography →