Free local tool

Experiment Protocol Builder

Build an exportable LLM experiment protocol and results-table template for testing Claude prompts, variants, fixtures, and success criteria.

how to test llm behaviorprompt experiment designClaude experiment protocol

Key facts

Cite facts
#

Fact

Browser-only protocol

The builder runs entirely in the browser. The prompt variants and criteria you type are not uploaded.

#

Fact

Evidence table included

The export includes fixture, variant, model, timestamp, score, notes, and raw-output fields.

#

Fact

Controls before claims

The generated card asks you to lock model, source state, tools, fixtures, and scoring before reading outputs.

#

Fact

Current model IDs

The model picker lists Claude Fable 5.1, Opus 5, Sonnet 5, and Haiku 4.5 by API model ID, plus a slot for legacy baselines.

Interactive builder

Generate an experiment card

Controls to lock

Protocol output

Results table template
fixture_idvariantrunmodelscorenotesraw_output_ref
How to Use This Tool
  1. Define the behavior question. Write the narrow Claude behavior you want to test, then choose the surface, the model under test, and the effort setting.
  2. Add prompt variants and criteria. Paste two or more variants, set the number of runs per fixture, and define measurable success criteria.
  3. Export the protocol. Copy or download the generated Markdown protocol and results table before running the experiment.
Tool FAQ
Does the builder run Claude?
No. It creates the protocol and results table. You run the experiment in your chosen Claude surface and paste raw evidence into your own records.
Which models can I select?
The September 2026 generally available lineup — Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5 — by API model ID, plus a legacy slot for regression baselines such as Fable 5 or an Opus 4.x model.
Can I use it for non-Claude LLMs?
Yes. The tool is written for Curious Claude readers, but the protocol fields work for any LLM if you record the model, surface, and source state.
What should I do before publishing a result?
Preserve the exact prompt, fixtures, model ID, date, source state, scoring rule, raw output, and scorer notes. Without those fields, publish it as a protocol note, not a result.
Internal links
Related pages
All free tools
The signature desk
The Bench
Flip six ready-made cards before you generate a new one.
Core method
The Curious Claude Method
The evidence standard behind every generated experiment card.
Protocol
Prompt Sensitivity
A protocol for testing small wording and structure changes.
Tool
Sampling Simulator
A toy visualizer for explaining temperature and top-p effects.