Source map
Primary Sources
Every protocol on Curious Claude is anchored to first-party Anthropic documentation or public policy/research pages where possible. Secondary commentary is useful for context, but it is not the evidence base here.
"Claude is a family"
Key facts
Cite factsFact
Primary first
The source map prioritizes Anthropic documentation, public system cards, transparency pages, and research or policy pages.
Fact
Record source state
Experiments should record the date and source state when a claim depends on current Claude behavior.
Fact
Sources are not results
A source can document a capability surface; it does not prove a local protocol outcome without a run.
Fact
Model pages included
The map includes the Claude Fable 5.1 and Claude Opus 5 model pages, because migration notes and effort defaults change how a run must be recorded.
Primary Sources
- Claude models overview
Current model family, modalities, platform surfaces, and model-selection context.
- Claude Fable 5.1 overview
Fable 5.1 effort levels, thinking behavior, and migration notes that change how a run must be recorded.
- Claude Opus 5 overview
Opus 5 capabilities, fast mode, and classifier fallback behavior relevant to run records.
- Choosing a model
Anthropic model-selection guidance used when declaring the model under test.
- Define success criteria and build evaluations
Evaluation criteria and measurement framing.
- Prompting best practices
Claude prompting guidance for clarity, examples, structured prompts, thinking, and agentic systems.
- Context windows
Context-window accounting, long-context limits, and what counts toward context.
- Token counting
Preflight token measurement and cost/rate-limit planning.
- Tool use with Claude
Client tools, server tools, and structured tool-call behavior.
- Define tools
Tool name, description, input schema, and example fields.
- Citations
Source-document grounding and passage-level citation behavior.
- Web search tool
Current web access, cited search results, and domain filtering.
- Vision
Image input surfaces and API image content blocks.
- Mitigate jailbreaks and prompt injections
Boundary between jailbreak and prompt-injection threat models.
- Prompt injection defenses
Anthropic research framing prompt injection as not solved.
- Claude text watermarking
What text watermarking is and is not, for experimenters asking whether it perturbs output fixtures.
- Alignment assessment and cybersecurity incidents
Anthropic disclosure of evaluation-containment failures; the public basis for sandboxing rules in boundary tests.
- Anthropic Transparency Hub
Public model summaries, access surfaces, safety summaries, and system-card links.
- Anthropic system cards
System-card index for capabilities, safety evaluations, and deployment decisions.
- Anthropic Responsible Scaling Policy
Public commitment framing around capability thresholds and safeguards.
Internal links
Related source pages
The signature desk
The Bench
Cards that cite these sources as the condition of a run.
Freshness
Experiment Ledger
Where source changes and protocol updates get dated.
Core method
Curious Claude Method
How source state fits into the evidence bundle.
Tool
Protocol Builder
Export source-state fields before running a test.