WriteArena: A Writing Improvement Tool

EN 3 min read

A deliberate-practice tool for becoming a better reader and writer.

app writing
Table of Contents

Use the tool: https://writearena.renatobritto.com

The problem

“Become a better writer” is the wrong unit of practice for the same reason “get better at programming” is. It is a bundle of separable sub-skills. The reason a humanities degree works is not the lectures. It is a loop, repeated for years: read excellent models, attempt to produce something, get specific diagnostic feedback against criteria, revise. The teacher’s real contribution is in three of those four beats: choosing good models, diagnosing precisely where your piece failed, and forcing the revision.

Ordinary practice fails because people do all the sub-skills fused together in one undifferentiated act and then get a single blurry grade. WriteArena breaks that bundle open.

The solution

I created WriteArena, a sibling of RhetoricArena, built on the same idea: isolate a sub-skill, work just past your current ability, get immediate correction, repeat. RhetoricArena does this for debate. WriteArena does it for writing.

The skill map

The app is organized around a 23-node taxonomy of writing craft, built from six faculties that run from the writer’s mind outward to the page and back:

  • Apprehension (I): reading, observation, research, invention. What you take in and generate before any words.
  • Conception (II): substance. The controlling claim in an essay, conflict and character in fiction, the governing image in a poem.
  • Disposition (III): arrangement. Whole-work shape, form-specific architecture (plot, argument structure, the volta), and the paragraph and scene as units.
  • Expression (IV): the line. Syntax, diction, figuration, sound, dialogue, subtext, mechanics.
  • Address (V): rhetoric. Audience, persuasion, framing, voice, tone, restraint.
  • Execution (VI): the meta-loop. Drafting, revision, self-diagnosis, editing, taste.

Each faculty is tagged for how trustworthy automated feedback is: faculties I, III, and IV (clarity and structure) can be graded against rubrics reliably. Voice, music, and taste (V, VI) are trained by comparison to real exemplars, not by asking a model “is this good.”

Two surfaces

Analyze. Paste a text or load one from the built-in library of 750+ public-domain classics. Select a window inside it. The engine names what each paragraph is doing, identifies load-bearing words, names the techniques in play, and traces the author’s strategy. The surrounding text is sent as context; the selected window is what gets annotated. Every label comes from the taxonomy and carries a reliability tag.

Drills. Short, single-skill reps over real passages, graded against a gold reference so the feedback is trustworthy. All 23 nodes are covered across five drill kinds:

  • Rewrite: clarity, compression, show-this
  • Reconstruct: the Franklin game. Given a target move and scaffold words, build the sentence, then see the original.
  • Identify: find the volta, the broken seam, the error
  • Diagnose: does the scene turn? what is the unstated warrant? who is the audience?
  • Compare: which figure earns it?

Feedback follows a four-beat contract: name, label, locate, fix. No clock, because writing rewards revision, not speed.

Stack

FastAPI backend, vanilla JS frontend with no build step, JSON session storage, Dockerfile for Coolify. Analysis runs through Claude (local CLI, no API key needed), OpenAI-compatible APIs, or DeepSeek. Keys are bring-your-own in the UI, held in memory for the request only, never written to disk.

More Posts