Fact
Six test surfaces
The protocol library covers prompts, long context, tool use, citations, safety boundaries, and vision reading.
Protocol index
Claude experiments, sorted by surface.
Each protocol is source-grounded, designed to be rerun, and careful about what it does not claim. Pick the surface you need to test, then preserve the raw evidence — including which of Claude Fable 5.1, Opus 5, Sonnet 5, or Haiku 4.5 answered, at which effort, on which surface.
Experiment 01
Prompt sensitivity
A reproducible way to test whether small wording, ordering, and formatting changes materially change Claude outputs.
Experiment 02
Context stress
A method for testing long-context tasks without mistaking token capacity for retrieval quality, reasoning quality, or cost control.
Experiment 03
Tool-use boundaries
A structured protocol for testing when Claude calls tools, asks clarifying questions, refuses unsafe action, or proceeds with incomplete parameters.
Experiment 04
Citation grounding
A protocol for testing whether Claude answers from supplied sources, cites the right passages, and avoids unsupported additions.
Experiment 05
Safety boundaries
A protocol for evaluating safe refusal, clarification, and prompt-injection handling without publishing jailbreak instructions.
Experiment 06
Vision reading
A protocol for testing Claude image and document reading with controlled fixtures, known answers, and coordinate-aware scoring.
Key facts
Cite factsFact
Rerunnable by design
Every protocol asks for enough evidence to rerun after model, tool, or source changes.
Fact
Source-grounded
Claims about Claude capabilities route back to Anthropic documentation, system cards, transparency pages, or documented runs.
Fact
Current lineup
Protocols were reviewed in September 2026 against Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5.
"Specific. Measurable. Achievable. Relevant."
Internal links
Related experiment resources
The signature desk
The Bench
Flip the six cards these protocols are cut from.
Tool
Protocol Builder
Turn a question and variants into a results-table template.
Core method
Curious Claude Method
The shared experiment card behind every protocol.
Freshness
Experiment Ledger
Track source changes and future documented runs.