Wordcraft

v0.1

Source Text

Parameters

Tone Neutral

Casual Formal

Length Same

Shorter Longer

Complexity Moderate

Simple Elaborate

Audience

Intent

Scroll to zoom · Drag to pan · Click cards to critique

Your canvas awaits

Generate variants to begin

Card added to canvas

Analysis

Analyzing your text...

Suggestions

Critique

Analyzing...

Agent Swarm

Idle 0 steps · 0 tokens

Quick Compare

2 variants

Show Diff

100%

Enter your API Key

Your key is stored locally in your browser and never sent to any server except Anthropic's API. Get your key here

Remember key on this device

On shared machines, uncheck this — the key is kept only for this browser tab's session. To remove a stored key, clear the field and save.

Model

Launch Agent Swarm

An orchestrator decomposes your goal, then dispatches a swarm of agents — researcher, writers, a critic that presses and probes, an editor, and a judge — onto the canvas.

Goal / Brief

Writers

Revise rounds

Enable web research agent

Ablation Lab

Run the same brief through two pipeline variants N times each, judge the outputs blind in randomized pairs, and see whether the extra stage actually earns its cost — not process theater, a win rate.

Goal / Brief

Factor under test

Trials per arm

Controlled Experiment

Signal vs. noise: this runs your current settings once and re-runs a neutral baseline — so you can see how much of the change is your edit versus the model's randomness. Baselines are cached across runs to build a real noise distribution, and a cheap judge scores how much the meaning moved, not just the wording (2–3 generations + 1–2 cheap judge calls per run).

Test

Noise · same config, re-run

How much the output changes from randomness alone

Signal · your settings

How much the output changes with your edit

Running the experiment…

Generating rewrites in parallel to separate your change from the model's random variation.

Sensitivity Map

Which knobs actually move the output for this text? This sweeps every generation knob against the neutral baseline and charts each one’s impact — meaning (model-judged) and wording — against the model’s own noise floor.

Mapping sensitivity…

Sweeping knob settings against the neutral baseline.