recursive-agentic-improvements: Instructions file for Claude Code

CLAUDE.md — Recursive Agentic Improvements

Behavioral guidelines for Claude Code and human contributors working in this repository. Adapted from the Karpathy CLAUDE.md and extended with project-specific rules.

Priority order when rules conflict: Core Invariants > Skill File Contract > Docs Contract > Style > Judgment.

---

Strict Token Optimization Protocols

1. Response Verbosity

• Moderate Explanations: Provide minimal direct solution, fix, or code block.
• Moderate Chat: Concise all pleasantries, introductions, summaries, and concluding remarks.
• Inline Comments Only: Place critical context strictly inside code blocks as short comments.
• Diffs Only: Output only the modified lines of code. Never print the entire file.

3. Tool Execution Constraints

• Lazy Execution: Execute tools only when explicitly necessary for the immediate task.
• Fail Fast: If a file path is ambiguous, stop and ask the user instead of searching the codebase.

---

Project Snapshot

What this is: A collection of three Claude Code skills (slash commands) that let any developer create, improve, and extend AI agents across Agno, CrewAI, LangGraph, and Google ADK — for any domain, without preset templates.

What this is NOT: A framework, a library, or an agent runtime. The skills install into other projects. This repo is the source of truth for those skill files.

Primary artifacts (in priority order):

1. .claude/commands/*.md — The three skills. These ship to users.
2. docs/**/*.md — Per-framework reference guides that inform the skills.
3. installer/ — The npx installer package that must stay in sync with .claude/commands/.

Design goal: A developer who runs /create-agent agno for a legal document summariser and one who runs it for a DevOps pipeline must both get a correct, production-quality scaffold — without the skill knowing about either use case in advance. The research phase does the discovery.