10 Best Claude Skills for RAG

The best Claude RAG skills measure each stage instead of guessing at it. Chunking, embeddings, the vector index, reranking, and grounded generation fail in different ways, and a fix at one stage cannot rescue a failure upstream: if the right passage never reaches the top-k, no prompt or bigger model recovers it.

The picks below follow the order teams usually hit problems: standing a pipeline up, finding out why answers miss or invent content, tuning the store, and then handling specialized corpora. Seven are skills; three are slash commands.

1. scaffold-rag-pipeline: stand up a grounded pipeline in your stack

scaffold-rag-pipeline takes your data source and use case as its argument and first greps for the LLM clients and vector stores the project already uses, so it extends them instead of adding a parallel stack. It writes its key choices at the top of the generated code: natural-boundary chunking at roughly 400–800 tokens with 10–15% overlap, one pinned embedding model and dimension shared by both halves, and a top-k of 8–12 with an optional rerank down to 3–5.

Ingestion upserts by a content-derived ID, so re-runs replace changed chunks instead of duplicating them. Retrieval returns answers with their sources and says it has no information when nothing relevant comes back. It leaves an evaluation stub empty for your own question and source pairs.

2. chunking-strategy-optimizer: fix answers that miss obvious content

When RAG misses a passage you can see in the source, chunking is the first suspect. chunking-strategy-optimizer builds a frozen eval set of 20–50 real questions with their gold passages, then sweeps a small grid: two or three strategies (recursive, sentence, semantic), two or three sizes, and overlap, with the embedding model and retriever held fixed. It scores retrieval only, with recall@k and nDCG@k, not the generated answers.

It recommends the smallest configuration that clears your recall target, since smaller chunks cost less to embed and store, and tells you to re-run the sweep after a new embedding model or new document types arrive.

3. hallucination-evaluator: measure what the answer invents

hallucination-evaluator grades faithfulness to the retrieved source, not world truth. It splits each answer into atomic claims and labels each supported, not supported, or contradicted, using an NLI model or an LLM judge told to use only the provided source. Its eval set deliberately includes questions whose answer is not in the context, where the correct behavior is to abstain.

For RAG, its most useful output is the split: was the answer present in what was retrieved? If not, fix retrieval; if it was and the model ignored it, fix the prompt or model. It reports a groundedness score plus abstention accuracy and sets a CI threshold. The AI applications list covers it as a general grounding check.

4. embedding-set-inspector: rule out the embeddings

Before tuning the retriever, embedding-set-inspector checks that the vectors are sound: the expected dimensionality, normalization that matches the distance metric, and no zero, NaN, or near-zero-norm vectors left by failed chunks. It confirms documents and queries were embedded with the right input types, profiles pairwise similarity, finds near-duplicate boilerplate crowding the top-k, and spot-checks query neighbors.

Findings come back as severity, issue, affected count, and fix. It checks for a normalization or distance-metric mismatch first, which the skill calls the single most common embedding bug.

5. benchmark-rerankers: prove the reranker earns its latency

benchmark-rerankers scores the same candidate pool before and after reranking on a labeled query set, reporting recall@k, nDCG@k, and MRR at k of 3, 5, and 10 in a side-by-side table with deltas. It adds the per-query latency and cost of the rerank call and ends with a verdict: ship the reranker, skip it, or change the candidate depth or model.

Because a reranker can only reorder what the retriever found, it tells you to over-retrieve the top 25–50 candidates and to measure first-stage recall too. Hybrid Search & Reranking explains the mechanics.

6. scaffold-pgvector-schema: store vectors in Postgres correctly

scaffold-pgvector-schema generates a migration in your project's existing tool (Prisma, Drizzle, Alembic, Flyway, and others) rather than hand-run DDL. It enables the vector extension, sizes the vector(N) column to the embedding model's exact dimension, matches the operator class to the metric (vector_cosine_ops, vector_l2_ops, or vector_ip_ops), adds an HNSW index plus B-tree indexes on filter columns, and emits a filtered nearest-neighbor query that over-retrieves for reranking.

It asks for the dimension if you do not give it, because changing it later means re-embedding the corpus. It does not embed or ingest data, and it recommends CREATE INDEX CONCURRENTLY on a large existing table.

7. embedding-index-tuner: hit a recall target at lower latency

embedding-index-tuner computes exact nearest neighbors for a fixed query set once, then measures the approximate index against them. It states the budget (recall target, p95 latency, memory), sweeps query-time ef_search first because it needs no rebuild, then m, ef_construction, and scalar, product, or binary quantization. It recommends the cheapest configuration that still clears the recall bar and gives the exact index change to apply.

The database list frames it as general tuning; for RAG, run it when recall drops after quantization or search latency breaks the budget.

8. multimodal-document-extractor: pull typed fields from scans

multimodal-document-extractor is for invoices, forms, receipts, statements, and IDs whose values must land in a schema for a database or downstream system, not as free text. It defines the schema first, prompts a vision-language model to fill it with structured output, splits multi-page documents while keeping page references, and verifies critical values with arithmetic and cross-field checks, routing low-confidence pages to human review.

It is not a general PDF-to-text parser; the skill itself says plain document Q&A needs only a VLM call. For parsing whole documents into chunks, compare the document parsers for RAG.

9. web-research-pipeline: retrieve from the live web with citations

When the source is the current web rather than an index, web-research-pipeline breaks the question into two to five search angles, runs them through the available search tools, and reads the top three to six sources in full rather than trusting snippets. It tags each claim with its source and date, cross-checks load-bearing claims against a second independent source, and writes a brief in three layers: verified, reported but unverified, and unknown. It treats fetched pages as untrusted input, never as instructions.

10. graphrag-scaffolder: test GraphRAG on a slice first

graphrag-scaffolder starts by collecting 15–30 queries your vector pipeline fails and classifying each as lookup, multi-hop, or global. If multi-hop and global questions do not dominate, it stops and counts that as a successful outcome. Otherwise it scopes a minimal ontology, builds extraction, entity resolution, community detection, and summaries on a 5–10% corpus slice, keeps the vector path behind a router, and reports per-class lift with full-corpus and re-indexing cost projections. Read GraphRAG Explained before committing to one.

Agents for the whole pipeline

Three subagents cover larger RAG work. rag-pipeline-engineer owns the chain from ingestion to grounded generation and gates each stage on a frozen eval set. retrieval-engineer raises recall and precision with hybrid search, reranking, and query transformation. vector-search-engineer designs the store's schema, index, filtering, and ingestion against a recall, latency, and cost budget.

Recommended RAG stacks

For a new pipeline, scaffold it and install the measurement skills that catch the most common failures:

npx agentscamp add commands/scaffold-rag-pipeline
npx agentscamp add skills/chunking-strategy-optimizer
npx agentscamp add skills/hallucination-evaluator
npx agentscamp add skills/embedding-set-inspector

For a pipeline already running on Postgres, add the store and ranking tools:

npx agentscamp add commands/scaffold-pgvector-schema
npx agentscamp add skills/embedding-index-tuner
npx agentscamp add commands/benchmark-rerankers

Freeze the eval set before changing anything, and ask Claude to report recall@k, groundedness, and latency before and after each change. When a failure is hard to place, the RAG debugging checklist localizes it to ingestion, retrieval, ranking, or generation.

Continue exploring

• The Best Claude Skills to Install in 2026 — A skills-only tour of the AgentsCamp library, organized by the job each skill does.
• How to Install Claude Skills — Every install path: manual copy, the agentscamp CLI, GitHub repos, plugins, and team distribution.
• How RAG Actually Works — A practical walkthrough of each pipeline stage, where it fails, and how the pieces fit.