Windowpane

Prose you can see through.

"Good prose is like a windowpane." — George Orwell, "Why I Write" (1946)

---

Windowpane is a small, focused writing plugin for AI coding harnesses. It turns George Orwell's 1946 essay "Politics and the English Language" into a working editorial discipline — one that both makes sincere writing clearer and exposes evasion in writing that dodges meaning.

It ships as one skill and two agents:

Component	What it does
orwell-prose (skill)	The craft layer: Orwell's six rules, the four-fault taxonomy, two modes (editor + detector), worked examples. Everything else builds on it.
prose-editor (agent)	Edits sincere prose in a file for clarity — cuts padding, dead metaphors, jargon, needless passives; reports before→after. Read-write.
prose-critic (agent)	Critiques read-only — gives clarity notes, and runs detector mode to expose euphemism, spin, and bad-faith vagueness in political/corporate/PR/marketing text. Never edits.
/proofread, /unspin	Slash commands that invoke the two modes directly.

The split is deliberate: the editor that improves sincere prose must refuse to "improve" insincere prose, because a tighter lie is a better lie. Point prose-editor at a layoff memo and it stops and hands you to prose-critic.

Command or agent? Same discipline, different reach. Use the commands (/proofread, /unspin) for a quick pass in the current conversation, on text you've pasted or a file you're already looking at. Use the agents (prose-editor, prose-critic) when you want the work done in its own context — a larger doc, a batch, or a step inside a bigger task — so the analysis doesn't crowd your main thread. The agents also self-trigger on plain edit/review requests; the commands are explicit.

The discipline

Orwell's six rules (heuristics, not laws — rule 6 governs the rest):

1. Never use a metaphor/simile you're used to seeing in print.
2. Never use a long word where a short one will do.
3. If it's possible to cut a word out, cut it out.
4. Never use the passive where you can use the active.
5. Never use jargon/foreign/scientific words if an everyday equivalent exists.
6. Break any of these rules sooner than say anything outright barbarous.

The four faults: dying metaphors · operators / verbal false limbs · pretentious diction · meaningless words. Detector mode adds a 9-category evasion taxonomy (euphemism, actor-deleting passive, nominalization, abstraction, dead-metaphor-as-thought-stopper, inflated diction, question-begging, quantifier fog, and internal contradiction / fails-its-own-thesis).

The guiding test for any suspect phrase: can I picture it, measure it, or point to it?

Install

Claude Code (native)

/plugin marketplace add bflandev/windowpane
/plugin install windowpane@windowpane

Or clone and load locally:

git clone https://github.com/bflandev/windowpane.git
/plugin marketplace add ./windowpane
/plugin install windowpane@windowpane

Then just write naturally — "tighten this README", "what is this press release hiding?" — or use /proofread and /unspin. The agents trigger on review/edit requests; the skill loads its own reference files on demand.

Other harnesses (Codex, Cursor, etc.)

There is no universal cross-harness plugin standard — Claude Code's plugin format is the one this repo targets natively. But the substance here is plain Markdown prompts, so it ports:

• OpenAI Codex / Cursor / Windsurf / Aider: the root AGENTS.md carries the core discipline in the convention those tools read. Point your agent file at it, or paste skills/orwell-prose/SKILL.md into your project rules.
• Anything else (xAI/Grok, a custom agent): load SKILL.md as a system prompt and the two agent files as named personas. The reference files under skills/orwell-prose/references/ are the lookup tables they draw on.

Write this way by default (optional)

The skill and agents act on text you hand them. If you also want Claude's own prose to follow Orwell's rules by default — in this and every project — add one line to your CLAUDE.md (~/.claude/CLAUDE.md for all projects, or a project-level CLAUDE.md for just one):

Write prose in plain English: short, concrete words; the active voice; no dead metaphors, no
padding, no jargon where an everyday word works. But this is a preference, not a mechanical rule —
keep a passive when the actor is unknown or beside the point, keep a long word or a term of art
when no short word carries the same meaning, and keep clarity over brevity when they conflict.
Never strip a sentence into something stilted, ambiguous, or wrong just to make it plainer.

The second half matters: Orwell's Rule 6 — "break any of these rules sooner than say anything outright barbarous" — is what keeps a plain-English instruction from turning into a mechanical strip-the-passives pass that produces prose as dead as what it replaced. This is a personal config choice, not something the plugin enforces.

Repository layout

windowpane/
├── .claude-plugin/
│   ├── plugin.json          # plugin manifest
│   └── marketplace.json     # makes this repo installable as a marketplace
├── agents/
│   ├── prose-critic.md      # read-only critic + detector
│   └── prose-editor.md      # file-editing copy-editor
├── commands/
│   ├── proofread.md         # /proofread → editor mode
│   └── unspin.md            # /unspin → detector mode
├── skills/orwell-prose/
│   ├── SKILL.md
│   ├── references/          # fault-taxonomy, false-limbs, dying-metaphors, detector-mode
│   └── examples/before-after.md
├── evals/                   # runnable calibration regression gate (run.sh + cases.jsonl)
├── assets/logo.svg
├── docs/DESIGN.md           # why it's built this way + validation
├── AGENTS.md                # cross-harness entry point
└── LICENSE

Credits

Built on George Orwell's "Politics and the English Language" (1946), which is in the public domain in many jurisdictions. Windowpane is an independent tool and is not affiliated with the Orwell Estate or the Orwell Foundation. Plugin code and prompts © 2026 Brian Flanagan, MIT licensed — see LICENSE.