AI Skills for Claude Code

Build reliable AI features. Powered by DSPy — a framework that lets you program language models with composable modules instead of hand-writing prompts.

Quick start

The only skill you need out the repo is /ai-do. Describe what you want to build and it tells you which skill to use next.

npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill ai-do

Then in Claude Code:

/ai-do I want to build a support ticket classifier

It picks the right skill, generates a ready-to-run prompt, and tells you what to install.

What problem are you solving?

| Your problem | Skill | What it does | |-------------|-------|--------------| | "I want to build something with AI but not sure where to start" | /ai-do | Describe your problem, get routed to the right skill with a ready-to-use prompt | | "I'm starting a new AI feature" / "DSPy quickstart" | /ai-kickoff | Scaffold a complete AI project with the right structure | | "I need a multi-phase plan for my AI feature" / "what order should I build this in" | /ai-planning | Create a phased roadmap before you start building | | "Which DSPy module should I use?" / "Predict vs ChainOfThought" | /ai-choosing-architecture | Pick the right DSPy pattern and architecture for your task | | "I need to auto-sort/tag/categorize content" / "classification results are inconsistent" | /ai-sorting | Build AI that sorts tickets, tags emails, detects sentiment | | "I need to search docs and answer questions" / "retrieval returns irrelevant results" / "RAG pipeline tutorial" | /ai-searching-docs | Build AI-powered knowledge base, help center, or doc Q&A | | "I need AI to answer questions about our database" / "chat with your Postgres" | /ai-querying-databases | Text-to-SQL: plain English questions over Postgres, MySQL, Snowflake | | "I need to condense long content into summaries" | /ai-summarizing | Summarize meetings, articles, threads — with length control | | "I need to pull structured data from messy text" / "the emails are messy and lack structure" | /ai-parsing-data | Parse invoices, extract entities, convert text to JSON | | "I need AI to take actions and call APIs" / "LLM function calling" | /ai-taking-actions | Build AI that calls APIs, uses tools, and completes tasks | | "I need AI to write articles, reports, or copy" | /ai-writing-content | Generate blog posts, product descriptions, newsletters | | "My AI fails on hard problems that need planning" / "LLM can't do multi-step logic" | /ai-reasoning | Add multi-step reasoning, Self-Discovery, chain-of-thought | | "My task needs multiple AI steps" / "LangChain LCEL alternative" | /ai-building-pipelines | Chain classify, retrieve, generate, verify into one pipeline | | "I need to verify AI output before users see it" / "LLMs invent data points" | /ai-checking-outputs | Add guardrails, fact-checking, safety filters, and quality gates | | "My AI makes stuff up / hallucinates" / "LLM makes up facts" | /ai-stopping-hallucinations | Ground AI in facts with citations, verification, and source checking | | "My AI doesn't follow our rules" / "LLM JSON output is unreliable" | /ai-following-rules | Enforce content policies, format rules, and business constraints | | "Review my DSPy code" / "is my AI code correct" / "best practices check" | /ai-auditing-code | Audit DSPy code for correctness, anti-patterns, and best practices | | "My AI gives wrong answers" / "I spent hours tweaking prompts" | /ai-improving-accuracy | Measure quality, then systematically improve it | | "My AI gives different answers every time" / "same prompt, different results every run" | /ai-making-consistent | Lock down outputs so they're predictable and reliable | | "My AI is too expensive" / "LLM API costs too high" / "GPT-4 costs too much" | /ai-cutting-costs | Reduce API costs with smart routing, caching, fine-tuning | | "Let's fine-tune on our data" / "prompt optimization hit a ceiling" | /ai-fine-tuning | Train models on your data for max quality or cost savings | | "Can we switch to a different model?" / "prompt broke after model update" | /ai-switching-models | Switch providers, compare models, re-optimize automatically | | "We don't have enough training data" / "no labeled data, need to bootstrap" | /ai-generating-data | Generate synthetic examples, fill data gaps, bootstrap from scratch | | "How do I put my AI behind an API?" / "deploy LLM as API" / "productionize my AI" | /ai-serving-apis | Wrap your AI in FastAPI endpoints for production serving | | "Is our AI safe to launch?" / "prevent prompt injection" | /ai-testing-safety | Automatically find vulnerabilities with adversarial testing | | "We need to moderate user content" | /ai-moderating-content | Build AI content moderation with severity levels and routing | | "Is our AI still working in production?" / "silent quality drops, prompt drift" | /ai-monitoring | Monitor quality, safety, and cost — catch degradation early | | "Why did my AI give that wrong answer?" | /ai-tracing-requests | Trace individual requests — see every LM call, retrieval, and step | | "I want to see what's happening during optimization" / "is my optimizer stuck?" | /ai-watching-optimization | Watch optimizer progress, interpret convergence, pick the right monitoring tool | | "Which of our optimization experiments was best?" | /ai-tracking-experiments | Log, compare, and promote optimization runs | | "I need AI to score, grade, or evaluate things" / "LLM as a judge" | /ai-scoring | Score essays, audit support quality, rate code reviews against rubrics | | "My AI works on simple inputs but fails on complex ones" / "works on simple inputs but fails on complex ones" | /ai-decomposing-tasks | Break unreliable single-step tasks into reliable subtasks | | "I need a conversational AI assistant" / "how do I build a chatbot" / "Intercom bot alternative" | /ai-building-chatbots | Build chatbots with memory, state, and doc-grounded responses | | "I need multiple AI agents working together" / "CrewAI alternative" | /ai-coordinating-agents | Supervisor agents, specialist handoff, parallel research teams | | "I need to translate content to other languages" / "localize our app" / "i18n with AI" | /ai-translating-content | Translate with glossary enforcement, batch i18n, brand voice | | "I need product recommendations" / "you might also like" / "personalize the feed" | /ai-recommending | Retrieval + LM re-ranking for personalized recommendations | | "I need to strip PII before sending to an LLM" / "GDPR compliance" / "anonymize data" | /ai-redacting-data | Detect and replace PII with regex + LM pipeline | | "I need to deduplicate contacts" / "entity resolution" / "merge records" | /ai-matching-records | Block, score pairs, merge duplicates across datasets | | "I need to normalize messy data" / "standardize company names" / "fix inconsistent formats" | /ai-cleaning-data | AI-powered data normalization with rule inference | | "I need to detect fraud or anomalies" / "flag suspicious transactions" / "abuse detection" | /ai-detecting-anomalies | Score events against baselines with severity and explanations | | "I need smart notification messages" / "weekly digest" / "incident alerts from logs" | /ai-generating-notifications | Event-driven notifications with channel constraints and digests | | "I need to analyze images" / "extract text from screenshots" / "generate alt text" | /ai-understanding-images | Vision model pipelines with dspy.Image for structured extraction | | "I need to rewrite text in a different tone" / "simplify legal language" / "adapt for different audience" | /ai-rewriting-text | Tone, reading level, and audience adaptation with fidelity checks | | "My AI is broken/erroring" / "Could not parse LLM output" | /ai-fixing-errors | Diagnose and fix crashes, wrong outputs, and weird behavior | | "DSPy can do X but there's no skill for it" | /ai-request-skill | Build a missing skill and submit a PR, or file a GitHub issue requesting it |

Using a specific tool with DSPy?

| Tool | Skill | What it covers | |------|-------|----------------| | VizPy (prompt optimizer) | /dspy-vizpy | Drop-in ContraPrompt/PromptGrad optimizers as alternative to GEPA/MIPROv2 | | Langtrace | /dspy-langtrace | Auto-instrument DSPy with one line, cloud + self-hosted tracing | | Arize Phoenix | /dspy-phoenix | Open-source trace viewer with built-in evals, local UI at localhost:6006 | | W&B Weave | /dspy-weave | Cloud experiment tracking and team dashboards via @weave.op() decorator | | MLflow | /dspy-mlflow | Auto-tracing, experiment tracking, and model registry for DSPy | | LangWatch | /dspy-langwatch | Auto-tracing + real-time optimizer progress dashboard | | Langfuse | /dspy-langfuse | Tracing + scoring + annotation queues + experiment tracking | | Ragas | /dspy-ragas | Decomposed RAG evaluation: faithfulness, context precision/recall | | Qdrant | /dspy-qdrant | Official vector DB integration + custom retriever pattern for any DB | | Ollama | /dspy-ollama | Run DSPy with local models, no API key needed | | vLLM | /dspy-vllm | High-throughput production serving for self-hosted models |

Know which DSPy concept you need?

If you already know DSPy and think in its vocabulary, use these API-first skills instead:

| DSPy concept | Skill | What it covers | |-------------|-------|----------------| | Signature, InputField, OutputField | /dspy-signatures | Inline and class-based signatures, typed fields, Pydantic models | | dspy.LM, dspy.configure | /dspy-lm | Provider strings, temperature/max_tokens, per-module LM assignment | | dspy.Assert, dspy.Suggest | /dspy-assertions | REMOVED in DSPy 3.x — legacy docs only. Use /dspy-refine or /dspy-best-of-n instead | | dspy.Module, forward() | /dspy-modules | Custom modules, composing sub-modules, save/load state | | dspy.Example, Prediction | /dspy-data | with_inputs(), train/dev splits, loading from CSV/JSON/HuggingFace | | dspy.Evaluate, metrics | /dspy-evaluate | SemanticF1, exact match, LM-as-judge, composite metrics | | dspy.Predict | /dspy-predict | Direct LM calls, simplest inference module | | dspy.ChainOfThought | /dspy-chain-of-thought | Step-by-step reasoning, reasoning field | | dspy.ProgramOfThought | /dspy-program-of-thought | Code generation + execution for math/computation | | dspy.ReAct | /dspy-react | Tool-using agents, Reasoning-Action-Observation loop | | dspy.CodeAct | /dspy-codeact | Agents that write and execute code to act | | dspy.MultiChainComparison | /dspy-multi-chain-comparison | Multiple reasoning chains, pick the best | | dspy.BestOfN | /dspy-best-of-n | Rejection sampling with a reward function | | dspy.Parallel | /dspy-parallel | Concurrent LM calls, batch processing | | dspy.Refine | /dspy-refine | Iterative self-improvement with feedback | | dspy.RLM | /dspy-rlm | Reinforcement-learning-style refinement | | dspy.BootstrapFewShot | /dspy-bootstrap-few-shot | Auto-generate few-shot demos, first optimizer to try | | dspy.BootstrapFewShotWithRandomSearch | /dspy-bootstrap-rs | Random search over candidate demo sets | | dspy.MIPROv2 | /dspy-miprov2 | Best prompt optimizer, instructions + demos jointly | | dspy.GEPA | /dspy-gepa | Instruction generation and selection | | dspy.BetterTogether | /dspy-better-together | Combined prompt + weight tuning | | dspy.BootstrapFinetune | /dspy-bootstrap-finetune | Fine-tune weights from bootstrapped data | | dspy.COPRO | /dspy-copro | Instruction candidates with breadth search | | dspy.Ensemble | /dspy-ensemble | Combine multiple optimized programs | | dspy.InferRules | /dspy-infer-rules | Extract decision logic from examples | | dspy.KNN, dspy.KNNFewShot | /dspy-knn-few-shot | Embedding-based demo retrieval | | dspy.LabeledFewShot | /dspy-labeled-few-shot | Hand-picked demonstrations | | dspy.SIMBA | /dspy-simba | Small-step incremental optimization | | ChatAdapter, JSONAdapter, TwoStepAdapter | /dspy-adapters | Prompt formatting, structured output (see also /dspy-two-step-adapter) | | dspy.ChatAdapter (deep dive) | /dspy-chatadapter | Field delimiters, parse/format internals, JSON fallback, fine-tuning data | | dspy.Tool, PythonInterpreter | /dspy-tools | Wrapping functions as tools, code execution | | dspy.Retrieve, ColBERTv2, Embedder | /dspy-retrieval | Search, RAG pipelines, embeddings | | dspy.Image, dspy.Audio, dspy.Code, dspy.History | /dspy-primitives | Multimodal inputs, conversation history | | dspy.streamify, StreamListener, StreamResponse | /dspy-streaming | Real-time token streaming to frontends, SSE, WebSocket | | dspy.Tool.from_mcp_tool, MCP servers | /dspy-mcp | Connect agents to MCP tool servers | | dspy.experimental.Citations, Document | /dspy-citations | Structured source attribution for RAG | | dspy.TwoStepAdapter | /dspy-two-step-adapter | Reasoning models (o1, o3, DeepSeek-R1) structured output | | aforward(), acall(), async patterns | /dspy-async | Async execution, FastAPI, concurrent calls | | inspect_history, save/load, configure_cache | /dspy-utils | Caching, debugging, persistence | | VizPy (ContraPromptOptimizer, PromptGradOptimizer) | /dspy-vizpy | Commercial drop-in prompt optimizer, alternative to GEPA | | Langtrace (langtrace.init) | /dspy-langtrace | Auto-instrument DSPy, cloud + self-hosted LLM observability | | Arize Phoenix (DSPyInstrumentor) | /dspy-phoenix | Open-source trace viewer with evals, local UI | | W&B Weave (@weave.op()) | /dspy-weave | Cloud experiment tracking and team dashboards | | MLflow (mlflow.dspy.autolog()) | /dspy-mlflow | Auto-tracing, experiment tracking, model registry | | LangWatch (langwatch.dspy.init) | /dspy-langwatch | Auto-tracing and real-time optimizer progress | | Langfuse (DSPyInstrumentor, @observe) | /dspy-langfuse | Tracing + scoring + annotation queues + experiments | | Ragas (ragas.evaluate) | /dspy-ragas | Decomposed RAG evaluation with LLM-as-judge metrics | | Qdrant (QdrantRM) | /dspy-qdrant | Vector DB retriever with hybrid search | | Ollama (ollama_chat/) | /dspy-ollama | Local model serving for development | | vLLM (openai/ + local server) | /dspy-vllm | Production self-hosted model serving |

Install

Option 1: npx skills (recommended — works with any AI coding agent)

Install all 94 skills in one command. Works with Claude Code, Cursor, Codex, Cline, Windsurf, and 35+ other agents.

npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills

The CLI will prompt you to pick which skills and which agents to install. Or install everything non-interactively:

npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --all -y

To install globally (available in all your projects):

npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills -g

Option 2: Claude Code plugin marketplace

# In Claude Code, run:
/plugin marketplace add lebsral/DSPy-Programming-not-prompting-LMs-skills

Then install a skill group:

/plugin install dspy-build-skills@dspy-skills     # Building AI features (17 skills)
/plugin install dspy-quality-skills@dspy-skills    # Quality and reliability (8 skills)
/plugin install dspy-ops-skills@dspy-skills        # Production operations (7 skills)
/plugin install dspy-api-skills@dspy-skills        # DSPy API-first skills (37 skills)

Option 3: Manual (git clone)

git clone https://github.com/lebsral/DSPy-Programming-not-prompting-LMs-skills.git

Copy skills to your agent's skill directory:

# Claude Code
cp -r DSPy-Programming-not-prompting-LMs-skills/skills/* ~/.claude/skills/

# Cursor
cp -r DSPy-Programming-not-prompting-LMs-skills/skills/* ~/.cursor/skills/

Or symlink to stay in sync with updates:

ln -s "$(pwd)/DSPy-Programming-not-prompting-LMs-skills/skills/"* ~/.claude/skills/

Option 4: Slash command (just /ai-do, no dependencies)

If you only want the router and prefer Claude Code's built-in slash command system:

curl -fsSL https://raw.githubusercontent.com/lebsral/DSPy-Programming-not-prompting-LMs-skills/main/commands/ai-do.md \
  -o ~/.claude/commands/ai-do.md

This gives you /ai-do globally in Claude Code. It routes you to the right skill and tells you how to install it. No npx, no plugins — just one file.

Managing skills

npx skills list          # See what you have installed
npx skills check         # Check for updates
npx skills update        # Update all installed skills
npx skills remove        # Uninstall skills

Keeping skills up to date

These skills are actively improved. To get the latest versions:

npx skills check          # See what's changed
npx skills update         # Pull latest versions

For manual/symlink installs, run git pull in your cloned repo.

> Migrating from v1.12.1 or earlier? A YAML formatting change in v1.12.2 means npx skills update may fail on some skills. Reinstall to fix: > ``bash > npx skills remove dspy-skills > npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills -g -s '*' > ``

Adding new skills after initial install

npx skills update only updates skills you already have. When new skills are added to this repo (like the dspy- API-first skills), re-run the add command to pick them up:

# Interactive — choose which new skills to add
npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills

# Non-interactive — add all new skills automatically
npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --all -y

This won't duplicate skills you already have — it only adds the ones that are missing.

Use a skill

In Claude Code (or any agent that supports the Agent Skills standard):

• Invoke directly: /ai-sorting or /ai-kickoff my-project
• Ask naturally: "Help me sort support tickets into categories" — the agent picks the right skill

How It Works

Each skill is a directory under skills/ containing:

• SKILL.md — Main instructions Claude follows (YAML frontmatter + markdown)
• examples.md — Worked examples (loaded on demand)
• reference.md — Detailed reference material (loaded on demand)

Under the hood, skills use DSPy — a framework for building AI features with composable modules that compile into optimized prompts. You don't need to know DSPy to use these skills; they guide you through everything.

Skills follow the Claude Code skills format and the Agent Skills open standard.

Reference Docs

• docs/dspy-reference.md — DSPy API quick reference (modules, optimizers, patterns)
• docs/langchain-langgraph-reference.md — LangChain & LangGraph API quick reference (loaders, tools, StateGraph)
• docs/skills-spec.md — Claude Code skills specification (for contributors)

Contributing

Adding a new skill

1. Create skills/ai-<problem>/SKILL.md — name it after the problem, not the DSPy concept 2. Add YAML frontmatter with name and description (include phrases users would naturally say) 3. Write step-by-step instructions in the markdown body 4. Add examples.md and/or reference.md for supporting content 5. Update the problem catalog table in this README 6. Test with /ai-<problem> in Claude Code 7. Bump the patch version in .claude-plugin/marketplace.json

See docs/skills-spec.md for the full skill format specification.

Links

• DSPy Documentation
• DSPy GitHub
• Claude Code Skills
• Agent Skills Standard