The Curious Claude Method

The Experiment Card

Every Curious Claude protocol starts with an experiment card. The card names the question, the Claude surface, the model ID, the date, the prompt, the fixtures, the allowed tools or source documents, the scoring rule, and the evidence bundle. If any field is missing, the writeup can still be a note, but it is not a result. The live desk for these cards is The Bench.

The card prevents a common failure: changing many things at once and then narrating the outcome as if one variable caused it. A Claude experiment should make the variable visible before the run starts.

• Question: the narrow behavior being tested.
• Variable: the one thing that changes between conditions.
• Controls: model, fixture, prompt shell, tools, sources, and scoring.
• Evidence: raw output, logs, citations, tool calls, and scorer notes.

The Model Under Test

As of September 2026 the generally available Claude models are Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5. An experiment card names one of them by API model ID — claude-fable-5-1, claude-opus-5, claude-sonnet-5, or claude-haiku-4-5 — because "Claude" on its own is not a testable condition. Which model to choose for production is a different question, owned by Claude Central; this page covers how to record the one you tested.

Effort and thinking settings belong to the model condition, not to the prompt. Anthropic documents adaptive thinking as always on for Fable 5.1, and a default effort that differs by surface: high in Claude Code, medium on Claude.ai and Cowork. The same prompt run on two surfaces is therefore two conditions, not one repetition, and a card that records only "Fable 5.1" has already lost a variable.

Knowledge cutoffs differ across the lineup, from February 2025 on Haiku 4.5 to June 2026 on Fable 5.1. A fixture that asks about something after a given cutoff is testing retrieval or refusal behavior, not recall, and it will not mean the same thing on two models. Legacy models remain listed — Fable 5, the Opus 4.x line, and Sonnet 4.6 and 4.5 — and are fair regression baselines as long as the writeup labels them as legacy conditions rather than current behavior.

• Model ID: the exact string sent to the API, never "latest" or a family name.
• Effort and thinking: the effort level requested and whether thinking was enabled, since surface defaults differ.
• Surface: API, Workbench, Claude Code, Claude.ai, or Cowork.
• Served model: Anthropic documents automatic fallbacks when safety classifiers flag a request to Opus 5 or Fable 5, so record the model that answered, not only the one you asked for.
• Retirement window: reproducibility ends when a model retires. Haiku 4.5 is documented as retiring no sooner than 2026-10-15.

Source State

Claude experiments drift because the platform moves. Anthropic publishes model docs, release notes, system cards, and transparency summaries; those sources should be linked from the experiment when a claim depends on current model behavior.

Record source access dates and, when possible, quote only a short anchor phrase. The point is not to copy documentation. The point is to make the experiment inspectable by a future reader.

Fixtures

A fixture is the input case used to test the behavior. Good fixture sets include normal cases, edge cases, a boring control, and an impossible or unsupported case. The impossible case is essential because it tests whether Claude can decline to infer from missing evidence.

When fixtures contain private or copyrighted material, publish metadata, hashes, and substitute examples rather than the private content itself.

Scoring

Scoring should be written before the outputs are read. A rubric can be qualitative, but it must be consistent. Anthropic explicitly allows well-defined qualitative scales in evaluation guidance; the key is that scorers know what counts before seeing the answer.

Separate dimensions that fail independently: answer correctness, source support, format validity, safety behavior, tool choice, latency, and cost. A single pass/fail column is often too blunt for Claude behavior.

Publication

Published results should include enough detail to rerun the experiment and enough caution to avoid overclaiming. If a protocol has not been run, say so. If the run used mocked tools, say so. If the model changed, make a new dated ledger entry rather than editing away the older context.

Curious Claude is strongest when it treats non-results as useful. A prompt variant that fails to improve outcomes, a context pack that passes only the short control, or a citation run that abstains correctly can all prevent bad product decisions.