dspy-agent-skills

DSPy Agent Skills

Production-grade DSPy 3.2.x skills for coding agents. A synthesized, spec-compliant pack of five agent skills that turns Claude Code, Codex CLI, and any other agentskills.io-compatible agent into a DSPy expert.

• ✅ Validated against DSPy 3.2.1 (the real API, not inferred from stale docs)
• ✅ Single source of truth for both Claude Code and Codex CLI
• ✅ Progressive disclosure (short SKILL.md + deep reference.md)
• ✅ Runnable example_*.py scripts with offline --dry-run
• ✅ Includes a DSPy 3.2.x BetterTogether chaining example
• ✅ Plugin manifest + marketplace manifest for one-click install
• ✅ Validation tests for frontmatter spec, JSON schema, Python AST, skill-doc correctness, and version alignment

What's inside

Skill	When it auto-invokes
dspy-fundamentals	Any new DSPy code: Signatures, Modules, Predict/ChainOfThought/ReAct, save/load
dspy-evaluation-harness	Writing metrics, splitting dev/val sets, calling dspy.Evaluate
dspy-gepa-optimizer	Optimizing/compiling DSPy programs with dspy.GEPA
dspy-rlm-module	Long context, codebase QA, recursive exploration via dspy.RLM
dspy-advanced-workflow	End-to-end builds — orchestrates the other four

Install

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

The Vercel skills CLI currently expects a GitHub owner/repo, URL, well-known HTTPS endpoint, or local path as its source. The bare form npx skills add dspy-agent-skills is not resolvable unless the upstream CLI adds a source alias, so use intertwine/dspy-agent-skills.

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/

Flags: --claude-only, --codex-only, --copy (copy instead of symlink), --uninstall, --dry-run.

Manual

Drop skills/* into ~/.claude/skills/ (Claude Code) or ~/.agents/skills/ (Codex CLI). See docs/installation.md for all options.

Five-second demo

In your agent, say:

"Build a DSPy sentiment classifier, optimize it with GEPA, and save the artifact."

The agent auto-loads dspy-advanced-workflow, which chains the other skills and outputs a full baseline → GEPA → export pipeline. No further prompting needed.

End-to-end examples (current committed artifacts)

Three runnable demos under examples/ exercise every skill against real LMs and ship with committed baseline vs. GEPA-optimized numbers plus explicit 3.1.3 vs. 3.2.0 comparison notes.

Example	Artifact DSPy	Task LM	Baseline	Optimized	Δ	Status
01-rag-qa	3.2.0	Ministral 3B 2512	80.47	100.00	+19.53	Clean comparison refreshed on 2026-04-28
02-math-reasoning	3.2.0	Ministral 3B 2512	85.00	93.33	+8.33	Refreshed on 2026-04-21
03-invoice-extraction	3.1.3	Liquid LFM 2.5 1.2B (free)	0.833	0.931	+0.098	Historical artifact retained

The refreshed 01 and 02 artifacts use the paid pair openrouter/mistralai/ministral-3b-2512 + openrouter/qwen/qwen3-30b-a3b-instruct-2507. 03 stays on its historical DSPy 3.1.3 artifact because a clean DSPy 3.2.0 baseline on the same Liquid/Nemotron pair already reached 0.944, leaving little useful headroom for a replacement GEPA artifact. See examples/README.md and each example's version_comparison.md for the exact commands and caveats.

Grounding

Every API claim is grounded in:

• https://dspy.ai/ (official docs, DSPy 3.2.x)
• https://code.claude.com/docs/en/skills.md (Claude Code skill spec)
• https://developers.openai.com/codex/skills (Codex skill spec)

Development

# Run validation suite
uv run --with pytest python -m pytest tests/ -v

# Smoke-test every example offline (no API key needed)
for f in skills/*/example_*.py; do uv run --with dspy python "$f" --dry-run; done

# Validate the current DSPy API surface used by these skills
env -u UV_EXCLUDE_NEWER uv run --with dspy==3.2.1 python scripts/check_dspy_surface.py