repo github.com View on GitHub ↗

Anthropic's Claude Cookbooks — the canonical recipe index

TL;DR

anthropics/claude-cookbooks is Anthropic's official, MIT-licensed collection of runnable Jupyter notebooks — "copy-able code snippets" for building with the Claude API (the capture reports 46,373★, updated 2026-07-02, README renamed from the older anthropic-cookbook). Its value to this vault is not a new argument but a canonical, first-party embodiment of a broad swath of concepts the graph already holds — the five Agentic Workflow Patterns, Pure Agent Applications via the Claude Agent SDK, Managed Agents (the managed_agents/ "CMA_*" recipes), Agent-Computer Interface (ACI) tool-use, RAG, Eval-Driven Development, and sub-agents (Reasoning Effort Control's Haiku-as-sub-agent-of-Opus recipe). So the dominant stance is corroborates: it is the runnable reference behind concepts sourced elsewhere, and it raises confidence without duplicating them. It earns its keep by surfacing two techniques the graph did not yet have a page for — Prompt Caching (misc/prompt_caching.ipynb, speculative_prompt_caching.ipynb) and Contextual Retrieval (capabilities/contextual-embeddings/) — which are spun out below. Should you care? Yes, as a lookup index, not a read-through: when a capture demands a concrete recipe (batch, citations, JSON mode, PDF, memory tool, programmatic tool calling, extended thinking, fine-tuning on Bedrock), this repo is the first-party place to fetch it — a deep-read of any one notebook is an on-demand deep-research burst, not this triage pass.

What it is %%ADAPTER_NOTE%% how it is organized

A flat set of topic directories, each holding self-contained .ipynb recipes plus a few support skills/scripts. The load-bearing top-level groupings:

capabilities/ — task recipes: classification, summarization, retrieval_augmented_generation/, contextual-embeddings/, knowledge_graph/, text_to_sql/.
claude_agent_sdk/ — eight numbered notebooks building end-to-end agents (research, chief-of-staff, observability, SRE, vulnerability-detection) + a 07_Hosting_the_agent; a Pure Agent Application / SDK track, incl. 04_migrating_from_openai_agents_sdk.
managed_agents/ — the "CMA_*" (Managed Agent) recipes: coordinate a specialist team, human-in-the-loop gate, iterate-fix-failing-tests, operate-in-production, orchestrate-issue-to-PR, prompt versioning/rollback, remember-user-preferences, verify_with_outcome_grader (LLM-as-Judge).
tool_use/ — calculator, customer-service, structured-JSON extraction, memory_tool, parallel_tools, programmatic_tool_calling_ptc, tool_choice, two tool_search_* recipes, automatic-context-compaction, context_engineering/.
multimodal/ — vision getting-started + best-practices, charts/graphs/PowerPoint reading, transcription, using_sub_agents (Haiku sub-agent of Opus).
extended_thinking/, evals/ + misc/building_evals.ipynb + generate_test_cases.ipynb, skills/ (custom-skills track), observability/usage_cost_api, patterns/agents/, third_party/ (Pinecone, VoyageAI, MongoDB, Deepgram, ElevenLabs, WolframAlpha, LlamaIndex…), and grab-bag misc/ (batch processing, citations, JSON mode, prompt caching, metaprompt, PDF summarization, session memory compaction).
Infra: Makefile, pyproject.toml/uv.lock, scripts/validate_*_notebooks.py + tests/ (notebooks are CI-validated), authors.yaml/registry.yaml, .claude/ (agents, commands, skills).

Concepts introduced

Prompt Caching — new page. Reuse a cached prompt prefix across calls to cut cost and latency on repeated long contexts; the repo ships both a base and a speculative recipe.
Contextual Retrieval — new page. Prepend chunk-situating context before embedding (and indexing) so retrieved chunks don't lose the document context that makes them findable; attaches to the existing retrieval spine (Hybrid Retrieval, Semantic Retrieval, Reranking).

Held, not dropped (themes the repo embodies that don't warrant their own page yet — spin out on demand when a second witness or a deep-read lands):

Programmatic tool calling (PTC) (tool_use/programmatic_tool_calling_ptc.ipynb) — the model emits code that calls tools rather than one-shot tool JSON. Held pending a focused source.
Tool search (tool_use/tool_search_alternate_approaches, ..._with_embeddings) — selecting from a large tool set; related to Search-Then-Get. Held.
Memory tool (tool_use/memory_tool.py, memory_cookbook.ipynb) — first-party agent-memory recipe; relates to AI Second Brain / Context Substrate. Held.
Context compaction (tool_use/automatic-context-compaction, misc/session_memory_compaction) — managing a growing context window. Held.
Extended thinking (extended_thinking/) — a recipe surface for Reasoning Effort Control; corroborates, no new page.
Batch processing, citations, JSON mode, moderation filter, metaprompt, PDF summarization, fine-tuning on Bedrock, knowledge-graph, text-to-SQL, custom skills, usage/cost API — each a discrete recipe; held as an index entry, not a concept, until a capture needs it.

Key claims

The Claude Cookbooks are copy-able code recipes for building with the Claude API; concepts are Python-first but portable to any language that can call the API. (observation — Golden Templates / Reusable Workflow Library) — the repo's self-description; it is a first-party recipe catalog, the official analogue of the personal loop-library pattern.
Notebooks are CI-validated (scripts/validate_all_notebooks.py, tests/notebook_tests/, lychee.toml link-check). observation — recipes are tested artifacts, not stale gists.
The repo treats "official recipes" as a maintained, contribution-driven catalog (author schema, registry.yaml, issue/PR templates). (best_practice — Reusable Workflow Library) — context: a vendor curating a shared reference; the anatomy (registry + validation + authorship) is what keeps a public recipe catalog trustworthy, contingent on wanting canonical status.
Sub-agent recipe: use Haiku as a sub-agent in combination with Opus. (observation — Reasoning Effort Control) — the source's stated pattern in multimodal/using_sub_agents.ipynb; a concrete cheap-model-under-strong-model instance.

Check-worthy source claims (attributed to the repo, not adjudicated — a later grounding pass can verify; named models/dates/counts are groundable):

The repo reports 46,373★, MIT license, default branch main, updated 2026-07-02, and a README renamed from anthropic-cookbook to "Claude Cookbooks." observation — capture metadata.
A fable_5_fallback_billing/guide.ipynb directory exists (a recipe framed around a model named "Fable 5" and fallback billing). observation — named-model/feature claim; attributed as the repo's, flagged for verification, not adjudicated here.
A claude_agent_sdk/04_migrating_from_openai_agents_sdk.ipynb recipe exists. observation — named cross-vendor migration recipe; groundable.

Why this `corroborates` (and where it adds)

The repo attaches to a large, already-populated region of the graph — agent architecture, tool use, evals, retrieval, managed agents — so it is not novel: its primary graph role is to be the canonical first-party embodiment of concepts sourced elsewhere (Building Effective Agents, the video distillates on loops/managed-agents/evals). It corroborates rather than duplicates: each "CMA_*" notebook is an independent witness that Managed Agent recipes exist and are maintained; patterns/agents/ + the SDK notebooks witness Agentic Workflow Patterns and Pure Agent Application in runnable form; evals/ + building_evals + verify_with_outcome_grader witness Eval-Driven Development and LLM-as-Judge. Note this is first-party corroboration (same vendor as building-effective-agents), not a fully independent second source — it raises availability confidence (the recipe is real and tested) more than consensus confidence.

Two secondary novel threads justify the two new pages: Prompt Caching and Contextual Retrieval attach to no prior concept (the retrieval spine had hybrid/semantic/rerank but not the situate-before-embed move). No contradicts tension surfaced — the repo is a reference index, complementary to the graph's argumentative distillates.