rite.

A sentence editing drill. Practice brevity.

0.0s

Loading…

0 out of 100

Closest reference rewrite

What this is

A drill for one specific skill: turning verbose, hard-to-read prose into something concise and clear, fast. You're shown a sentence in the manner of a Victorian novelist, a regulator, an academic, or a corporate memo. You rewrite it. You're scored.

Where the sentences come from

45 hand-curated sentences in corpus.json, tagged by style:

Victorian — long subordinate clauses, qualifying asides, throat-clearing.
Bureaucratic — nominalizations, passive stacks, hedged conclusions.
Academic — "It must be acknowledged that there exists…"
Legal — "Pursuant to…", "Notwithstanding the foregoing…"
Corporate / TOS — PR-speak and click-through filler.

Each sentence ships with three reference rewrites — typically a literal paraphrase, an idiomatic version, and a third variant. Grading takes the highest similarity across all three, so idioms ("the proof's in the pudding", "back to square one", "doomed to repeat it") score as well as literal rewrites.

How grading works

Every submission gets four sub-scores, combined with a speed multiplier:

score = (0.5 · fidelity + 0.3 · brevity + 0.2 · clarity) × speed × 100

Component	Weight	Measured by
Fidelity	50%	Maximum cosine similarity (BGE embeddings) between your edit and any of the reference rewrites. Falls back to Jaccard token overlap if the AI binding is unavailable.
Brevity	30%	Peaks when your word count ≈ the median reference word count. Capped near zero if you didn't shorten the original at all.
Clarity	20%	Flesch reading-ease delta vs. the original. A 40-point jump caps the score.
Speed	multiplier	1.0 at ≤10s, decays linearly to 0.5 at 60s.

What it costs

Runs on a Cloudflare Worker with the Workers AI binding. Each grade does one batched embedding call (~1 neuron). The $5 Workers Paid plan includes 10,000 free neurons/day — roughly 5,000 free grades/day before pay-per-use kicks in.

Tips

Strip nominalizations: "make a determination" → "decide".
Cut hedges: "it would seem to be the case that…"
Active voice when you can; the Flesch score rewards it indirectly via shorter sentences.
If you're stuck, type anything short and submit — the reference rewrite is shown afterwards.
Cmd/Ctrl+Enter submits.

Source

The drill is a single Worker plus static assets. Corpus, scoring, and worker code live in the rite/ directory of minormobius/agent01.

Want to contribute sentences?

The corpus is crowdsourced via /fodder/ — a Tinder-style swipe interface where you triage verbose sentences mined nightly from Project Gutenberg. Right-swipes feed back into this corpus.

Sibling tools

/atlas/ — multi-view analytics over a Bluesky archive's prose threads.
/lexicon/ — word-level lenses (frequency, distinctiveness, NRC emotion, AFINN sentiment, concreteness).
/redact/ — a Redactle-style game over someone's longest Bluesky threads.
/ask/ — semantic search across a Bluesky archive's prose threads.
/list/ — semantic analysis over a Bluesky list (themes that bind the group).
/names/ — procedural name-set generator (300 names from one invented culture, with an API).