Made for Claude Code

Convert AI text into human-friendly prose

Prose Craft gives your agent two hard rules, a short list of measured habits, and a filter that sends every violation back until the text is clean.

Watch it edit Install in two commands
Typical draft

Setup takes one command — no daemon, no account.

Crafted

Setup takes one command, with no daemon and no account.

Intake to output

Prose Craft in Action

Each draft starts as typical AI-written prose. The slider applies Prose Craft one sentence at a time while the real filter report on the right counts down to zero. These samples are illustrative, written by hand, and checked with the actual filter.

Intake

>

Output
0% rewritten
typical AI draftwith Prose Craft
$ python3 scripts/prose-filter.py check draft.md
exit 1
Sentence rhythm and openers

Each bar is one sentence, sized by word count. Striped red means the filter flags it, green means rewritten, dark means it never needed a fix.

original
current
Habit · sentence length

Join choppy sentences so the writing flows.

AI-written fiction chops its ideas into short sentences. A model's sentences ran six words or fewer 44% of the time in my measurements, against 8–19% for the authors I compared. Drag the slider to join a run of short sentences into one that reasons its way through, and keep a short sentence only where you want emphasis.

Output
0% varied
monotonevaried
Sentence rhythm

Each bar is one sentence, sized by word count.

original
current
Part 1 · hard rules

Two rules a script can check.

RULE 1

Subject first.

No sentence opens with an adverbial. That covers prepositional phrases, infinitive purpose phrases, Because, When, and If clauses, participial openers, and -ly adverbs. Move the phrase after the verb.

In parallel, I ran an ablation.
I ran an ablation in parallel.

Fine as openers: Furthermore, However, Therefore, Thus, Also, Then, Still, Finally, Meanwhile, Instead.

RULE 2

No em dashes.

Never write the long dash, or a spaced hyphen standing in for one. Reach for a period, a comma, a colon, or parentheses. An en dash inside a number range is fine.

The test passed — barely.
The test passed, barely.

Em dashes were flagged in nine of thirteen blind-test pairs.

Would the filter flag it?

“”

Part 2 · measured habits

Habits that fix the way AI writes.

The skill tells the agent what to write, not only what to avoid, because a rule against something leaves room to argue. These habits guide the draft as the agent writes it. The script does not check them.

LENGTH

Vary the sentences.

Let a long sentence carry the reasoning. Use a short one only where it earns the stress.

COMPLETENESS

Give the full account.

A PR body or email gets what changed, why, the caveats, and the next step. Readers preferred the thorough version every time.

STRUCTURE

Match the register.

Headings and lists for reference material. Plain paragraphs for a chat message, an email, or a story.

SPECIFICS

Lead with the vivid fact.

Slide titles state the conclusion, not the topic. No Challenge, Solution, Results template.

FICTION

Pick details you have not used.

A model left to itself writes the same story: a lighthouse keeper, the same few names, the same opening. The skill steers it toward fresh details and a story that builds to a turn.

CLOSINGS

End on a concrete line.

A landing page ends on its best fact, such as “Setup takes one command.” It skips “Contact us today to learn more!”

The filter

Flags problems for the agent to fix

The script needs only Python 3. It skips fenced and inline code, prints each violation with its line, and exits 0 when the text is clean. Nothing runs on its own, because the plugin installs no hooks.

ROUND 1

Check

exit 1 · violations found

The report lists each flagged sentence and a correction message.

ROUND 2

Regenerate

message on stdin

Your agent receives the correction message and prints corrected text.

ROUND 3 OR SOONER

Clean

exit 0 · no violations

Exit 1 if it still fails.
Exit 2 if your agent fails.

# bring your own agent
python3 scripts/prose-filter.py run --agent "claude -p" draft.md
python3 scripts/prose-filter.py run --agent "codex exec -" draft.md
python3 scripts/prose-filter.py run --agent "ollama run MODEL" draft.md
--max-rounds N sets the attempts (default 3).
Your copy

Try Prose Craft Live

This is the same check the plugin runs, ported to run in your browser. It reads two rules: no em dash, and no sentence that opens with an adverbial. This checker keeps your text in your browser. The workbench goes one step further and has Claude rewrite each flagged sentence.

Filter report
exit 1
Craft it in the workbench → The workbench sends your text to Claude on your own account. It re-checks the result and asks Claude to fix anything still flagged, for up to three rounds.
Evidence

What the measurements show.

Em dash rate

An open-weights model fine-tuned against em dashes, on a held-out set of prompts.

base
80%
tuned
9%
Sentences of six words or fewer

Share of sentences in AI-written fiction versus the authors I compared.

model
44%
masters
8–19%
Not built to beat detectors

Prose Craft does not try to pass AI detection. The goal is prose that reads well to a person, and detector scores do not measure that.

No facts added

Prose Craft changes how a draft is written, not what it says. It uses only the facts you supply and adds no dates, numbers, people or claims of its own.

Other harnesses

Works in any harness.

That includes Codex and OpenCode. Prose Craft is a standard SKILL.md folder, so any harness that reads Agent Skills can load it. Drop prose-craft into its skills directory, then ask for it by name or let the description trigger it. The filter is plain Python 3 and takes any agent command.

# Codex: built-in skill installer
python3 ~/.codex/skills/.system/skill-installer/scripts/install-skill-from-github.py --repo nhangen/prose-craft --path skills/prose-craft
# OpenCode: clone, then copy the folder
git clone https://github.com/nhangen/prose-craft
cp -R prose-craft/skills/prose-craft ~/.config/opencode/skills/
Then
  • Ask the agent to write with prose-craft, or let it load on its own when you draft prose.
  • Run the filter from inside the installed skill folder: python3 scripts/prose-filter.py check draft.md.
  • Pass any agent to run --agent to loop until clean.
Install

Installs in two commands.

It runs whenever you write copy. Claude Code loads it on its own for prose a person will read, or you can call it by name with /prose-craft.

Prose Craft ships in nhangen/claude-plugins, a marketplace of useful Claude Code plugins.

Browse the marketplace Read the source
# add the marketplace
claude plugin marketplace add nhangen/claude-plugins
# install the plugin
claude plugin install prose-craft@nhangen-tools
# check a draft
python3 scripts/prose-filter.py check draft.md