Dw Unslop

dw-unslop — the last pass before a human reads it

The only skill here that edits wording. Everything else in the lane writes for an agent, where density is the point; this one runs over the few things a person actually reads, where the patterns that mark text as machine-written spend the reader's attention before the content gets any.

Two failures, not one. The obvious one is slop — puffery, hedging, the vocabulary. The other is over-correction: prose sanded down until it has no opinion and no rhythm, which reads as generated just as fast. Step 3 is there for the second one.

What it reads and writes

Reads whatever the argument names: empty takes the open change's PR body and the commit bodies on its branch, a path takes that file, pasted text takes itself. Writes the rewritten text — into a tracked file only after you have seen the diff — and no .ai/ artifact of its own.

It refuses, by name: any SKILL.md, AGENTS.md / CLAUDE.md, docs/agents/*, docs/decisions/*, CONTEXT.md, CHANGE.md, HANDOFF.md, .ai/backlog/*, and any ## Gotchas block. Every one of those is written for an agent rather than a person — some loaded every session, some only when a router points at them — so a sentence trimmed out of one is an instruction a later session no longer gets, and nothing is watching to notice. Name the file, give the reason in one line, stop — never counter-offer a lighter touch on the same file.

Workflow

1. Scan

Read the text against references/patterns.md: 31 numbered rules in seven groups, each written as an edit rather than a flag. Note which numbers fire — step 4 reports them.

2. Rewrite

Preserve the meaning and match the register the text was aiming at. A PR body is not an essay: the fix for puffery there is the fact the puffery stood in for, not livelier puffery. Where a rule fires but the rewrite would change what the sentence claims, leave it and say so. The claim is the author's.

3. Don't flatten it

Cutting patterns is half the job. Sterile is its own tell, so the pass has to put something back:

• Keep the opinion. A sentence that judges beats a sentence that lists. "The generator is not worth two skills" carries more than "there are trade-offs here."
• Vary the rhythm. One short sentence. Then one that takes its time, because the thought it is carrying actually needs the room. Uniform sentence length is the most reliable machine tell there is.
• Keep a specific detail. A number, a filename, a file:line. Specificity is what slop cannot fake.
• Let some structure stay uneven. Perfect parallelism looks generated — three bullets over material that has four things in it is rule 10 firing.

This step is what stops the skill becoming a length optimiser. An agent told to tighten prose optimises for length, because length is the one property it can measure — and it pays for the shorter file out of whatever it could not: a flag dropped from a command that still reads like a command, the "why" cut from under a rule that kept its "what". No format check sees that bill. Shorter is not the goal. Less slop is.

4. Report what fired

Show the rewritten text and name the rules by number, one line each. On a tracked file, stop for approval before writing — the diff is the part worth reading, and a wording pass applied unseen is how a claim quietly changes owner.

The two tests

The catalog is patterns to recognise. These two are the algorithm, and they catch the sentences no list will — apply them first, because both are deletions and a deleted sentence needs no rewrite:

• The restatement test. Try to restate the sentence as a concrete instruction, a fact or a number. If you cannot, it names a feeling rather than a thing, so cut it. "The database stays close at hand" fails; ".toSQL() returns the exact string sent to the database" passes.
• The portability test. If the sentence would read identically in another project's docs, it says nothing about this one. Cut it. This is the test that kills the well-formed paragraph a reader nods along to and learns nothing from.

References

• references/patterns.md — the numbered catalog, kept at its upstream numbers so "rule 16" means the same thing in a review here and in the plugin it came from. Two rules are deliberately softened there, each with the reason on its own line.

Next: dw-land, which owns the PR body; a commit body about to be written follows ## Git conventions.

$ARGUMENTS