DSPy agent skills pack with GEPA optimization for Claude Code and Codex CLI
what the dspy-agent-skills repo implements
This repository provides a production-grade collection of five DSPy 3.2.x agent skills tailored for Claude Code and Codex CLI environments. DSPy (Declarative Skills for Python) is a specification-driven framework for defining, validating, and running agent skills, and this repo builds on version 3.2.x of that framework with a solid focus on practical use and optimization.
Under the hood, the repo delivers spec-compliant skill definitions, including detailed SKILL.md files and reference documentation, which collectively define the behavior and invocation logic for each skill. The skills cover core DSPy fundamentals, evaluation harnesses, GEPA (Gradient-based End-to-end Prompt Architecture) optimization modules, Reinforcement Learning Modules (RLM), and advanced workflows.
The stack is Python 3.10+, leveraging standard tools and idioms for validation and testing. The repo includes 80 validation tests that guard correctness at multiple levels — from frontmatter metadata and JSON schema to Python AST syntax and skill documentation consistency.
Architecturally, the skills are modular and designed for progressive disclosure, meaning they reveal themselves contextually and auto-invoke based on the agent’s current state and inputs. This design allows Claude Code or Codex CLI agents to evolve into DSPy experts without requiring additional user prompting or manual skill management.
how g.e.p.a. optimization and rigorous validation set this repo apart
The standout technical feature is the integration of GEPA optimization, which applies compile-time improvements to DSPy skills. This approach contrasts with typical prompt engineering by optimizing the skill definitions and workflows before runtime, yielding substantial accuracy gains.
The repo commits baseline and GEPA-optimized benchmark numbers across three end-to-end demo tasks: retrieval-augmented generation (RAG) question answering, math reasoning, and invoice extraction. For example, the RAG QA task sees a +19.53% accuracy jump from 80.47% to a perfect 100% using only a 3B parameter model, which is a concrete and impressive gain that illustrates the value of compile-time optimization over prompt tuning alone.
Similarly, math reasoning accuracy improves by +8.33% (from 85.00% to 93.33%), and invoice extraction improves by +0.098 (from 0.833 to 0.931). These numbers are fully reproducible thanks to the included validation tests and runnable example scripts, which support offline dry-run scenarios.
The codebase achieves a balance between modularity, validation rigor, and performance. The validation suite ensures specs and implementations stay in sync, preventing common pitfalls in skill development. Progressive disclosure and auto-invocation also reduce developer friction and runtime overhead, as skills only activate when contextually relevant.
The tradeoff is increased complexity: newcomers must understand DSPy 3.2.x specs, GEPA optimization concepts, and the agent skill architecture. The repo assumes Python 3.10+ and specific runtime environments (Claude Code, Codex CLI) which limits general applicability but fits well in its target niche.
quick start with the dspy-agent-skills pack
Installation and setup are straightforward with multiple options depending on your environment:
Claude Code (via marketplace)
/plugin marketplace add intertwine/dspy-agent-skills
/plugin install dspy-agent-skills@dspy-agent-skills
Agent Skills CLI (npx skills)
npx skills add intertwine/dspy-agent-skills --list
npx skills add intertwine/dspy-agent-skills --skill '*' -a codex -y
Note: Use the full intertwine/dspy-agent-skills notation as the CLI expects a GitHub owner/repo or URL.
Claude Code + Codex (repo checkout)
git clone https://github.com/intertwine/dspy-agent-skills
cd dspy-agent-skills
./scripts/install.sh # symlinks into ~/.claude/skills/ and ~/.agents/skills/
The install script supports flags such as --claude-only, --codex-only, --copy, --uninstall, and --dry-run for flexible setups.
Manual installation
Simply drop the skills/* directory into your ~/.claude/skills/ (Claude Code) or ~/.agents/skills/ (Codex CLI) folder. Full options and details are in docs/installation.md.
verdict: a solid toolkit for DSPy skill optimization in Claude Code and Codex
intertwine/dspy-agent-skills is a focused, well-engineered collection designed for users who are already invested in the DSPy framework and Claude Code or Codex CLI environments. The GEPA optimization results provide convincing evidence that compile-time skill tuning can outperform traditional prompt engineering for specific AI tasks.
The code is surprisingly clean and well-tested for the complexity it handles, and the progressive disclosure mechanism improves runtime efficiency and developer experience. However, the learning curve for newcomers is non-trivial, and the dependency on DSPy 3.2.x and Python 3.10+ restricts its direct applicability.
If your work involves building or optimizing AI agents in Claude Code or Codex CLI, and you want to explore systematic skill optimization beyond prompt engineering, this repo is worth serious exploration. For casual users or those outside this ecosystem, the specialized nature and technical overhead might be limiting.
All in all, this repo is a practical case study and toolkit for production-grade DSPy agent skills with measurable performance gains.
→ GitHub Repo: intertwine/dspy-agent-skills ⭐ 207 · Python