lebsral

Request or contribute a new AI skill that does not exist yet. Use when DSPy supports something but there is no skill for it — helps you build the skill and submit a PR, or file an issue requesting it. Also use when the user says there should be a skill for this, can we make a skill, I want to contribute a skill, none of the ai- skills cover my use case, how do I add a new skill, submit a skill, or

Build AI that searches your documents and answers questions. Use when building a knowledge base, help center Q&A, chatting with documents, answering questions from a database, search-and-answer over internal docs, customer support bot, or FAQ system. Also use when embedding search loses critical context, retrieval returns irrelevant results, the right document is buried deep in search results, RAG

Use when working with reasoning models (o1, o3, o3-mini, DeepSeek-R1, Claude extended thinking) that reject system prompts or ignore formatting instructions. Common scenarios - using o1 or o3 with DSPy, getting structured output from reasoning models, two-phase prompting where a reasoning model generates freely then an extraction model parses the output, or fixing format errors from thinking model

Use when you have independent LM calls that can run concurrently — batch processing, fan-out patterns, or speeding up pipelines with no data dependencies between steps. Common scenarios - processing a batch of inputs through a DSPy module concurrently, fan-out patterns where multiple independent LM calls run at once, speeding up evaluation by parallelizing predictions, or reducing wall-clock time

Use when the task requires calling external tools or APIs to gather information — multi-step tool use with reasoning, like searching databases, calling APIs, or combining multiple data sources. Common scenarios - building agents that search the web and synthesize results, multi-step information gathering from APIs, chatbots that look up data before answering, question answering that requires exter

Analyze images and extract structured data from visual content using AI. Use when analyzing product photos, extracting text from screenshots, generating alt text for accessibility, visual question answering, categorizing images by content, reading receipts and invoices from photos, OCR with AI, describing images for search indexing, product photo categorization, document image processing, chart an

Use when you need DSPy infrastructure - caching control, debugging with inspect_history, saving/loading optimized programs, or runtime validation with Refine/BestOfN. Common scenarios - controlling the cache to avoid stale results, debugging with inspect_history to see raw prompts, saving and loading optimized programs, or validating outputs with reward functions. For streaming see /dspy-streaming

Use when you have multiple optimized versions of a program and want to combine them — voting, averaging, or routing across program variants for more robust outputs. Common scenarios - you have optimized several versions of a program and want to combine the best ones, using majority voting across multiple programs for higher accuracy, building a robust system by routing to different specialized pro

Use when output quality varies across runs and you want to sample multiple completions and pick the best — trading latency for reliability on high-stakes outputs. Common scenarios - generating multiple candidate answers and picking the highest-scoring one, improving reliability on high-stakes classification, reducing variance in creative generation, getting better summaries by sampling several and

Break a failing complex AI task into reliable subtasks. Use when your AI works on simple inputs but fails on complex ones, extraction misses items in long documents, accuracy degrades as input grows, AI conflates multiple things at once, results are inconsistent across input types, you need to chunk long text for processing, or you want to split one unreliable AI step into multiple reliable ones.

Verify and validate AI output before it reaches users. Use when you need guardrails, output validation, safety checks, content filtering, fact-checking AI responses, catching hallucinations, preventing bad outputs, or quality gates. Also used for - AI output looks right but is wrong, how to validate JSON from LLM, LLM returns invalid data, catch bad AI outputs before users see them, output quality

Use when you have 50+ labeled examples and want a quick accuracy boost as your first optimization step — the simplest and fastest DSPy optimizer. Common scenarios - your first optimization attempt on a new DSPy program, adding few-shot examples automatically from labeled data, quick accuracy boost before trying heavier optimizers, bootstrapping demonstrations from a teacher model, or getting start