Deslop (prose)

Review gate for text, not code. deslop handles code; this handles anything a human will read as prose.

Run it on the draft, before writing to disk and before showing the user.

Step 0: load the destination's own rules

If the target path is inside a tree that has a CLAUDE.md at its root (for example docs/CLAUDE.md), read that file first. It outranks everything below. Those files carry the project's own substitution tables, terminology bans, and glossary thresholds, and they are usually the only place a house rule is written down.

Skipping this step is the single most common way this gate fails. Do not skip it because the rules "look generic".

Step 1: mechanical checks

Grep the draft. Any hit is a defect.

grep -nP '\x{2014}|\s-\s' <file>      # em dashes and spaced dashes
grep -niE 'robust|seamless|abrangente|poderoso|simplesmente|essencialmente|basicamente|obviamente|vale ressaltar|é importante notar|In this change'

Then run whatever verification greps the destination CLAUDE.md defines.

Step 2: calques and jargon (pt-BR drafts only)

A literal translation that is grammatically valid can still be wrong. Check the destination's substitution table first, then these.

Rule of thumb: if the team says the English word out loud in a code review, keep the English word and italicize it. If nobody says it, translate the concept.

Check the corpus before inventing a term. Grep the destination tree for both candidates and follow what is already there, unless what is there is provably a repeated mistake.

Step 3: gloss the jargon that is not an acronym

Glossary rules usually trigger on acronym count and therefore miss the terms that actually confuse people: library names, product names, and internal concept names.

Every library, product and named concept gets a one-line gloss at its first occurrence, or a row in the glossary table. Ask of each one: would a competent dev outside this repo know what this is?

Worked examples of failures this catches:

• "undici": a reader cannot tell it is the HTTP library built into Node that implements fetch.
• "agente interno": reads like an HTTP client. It is an https.Agent, a connection pool plus TLS options.

Glossing the term is not enough when the term is the crux of the document. If the whole point turns on why two things do not fit together, spend a short paragraph on the mechanism.

Step 4: shape

• Parallel structure with 3 or more items becomes a table. Repeated ### file headings with one paragraph each are parallel structure.
• No section the calling skill did not authorize. If a section feels necessary and is not on the list, ask instead of adding it.
• Respect the caller's line budget. Count the body, excluding frontmatter, and report the number.
• Every sentence carries a fact, a decision or a constraint. Delete the rest.

Step 5: report

State what changed and the before/after line and word count. If the draft is over budget, say by how much and why, rather than cutting content the user explicitly asked for.

Red Flags: STOP and fix

• Reviewing a doc bound for a tree with its own CLAUDE.md without having read it.
• Translating a technical term the team says in English.
• A library or product name appearing with no gloss.
• Three or more sibling sections with identical shape, left as headings.
• A section that the calling skill never listed.
• Claiming the review passed without running the greps.

Rationalization Table