Prose Clarity

---
name: prose-clarity
description: "Rewrite and self-check drafted prose against a machine-checkable clarity standard derived from ASD-STE100 (Simplified Technical English) — no marketing adjectives, no phrasal verbs, no semicolons, active voice, short sentences, and a required carve-out for the epistemic hedge markers the evidence ledger depends on. Use when drafting any package document's prose, before returning a draft, or when an independent verification pass evaluates editorial quality. This skill MUST be consulted because a banned-word list alone barely moves AI slop — the habits that produce it (hedge-stacking, nominalization) generate new slop the list never anticipated — and only rules a script can verify hold up under revision."
allowed-tools: Read, Write, Edit, Grep, Glob, Bash
context: fork
agent: general-purpose
---

# Prose Clarity

Owns the sentence- and paragraph-level form of drafted prose — never its truth, never its structure, never its claim state.

## Iron Law

**THE LINTER CATCHES FORM, NEVER TRUTH — AND A REQUIRED HEDGE ON AN `I`-STATE CLAIM IS NEVER A VIOLATION, IN ANY MODE.**

A sentence that must carry `[EV-####]`, keep `Inferred:`/`Unknown:`/`Recommendation:`, or preserve a limitation that keeps it true is correct at whatever length that takes. This skill fixes the form of slop. It cannot make a hollow paragraph true, and it never overrides an Iron Rule from another skill.

## Classification is by passage, not by document

Any line inside a numbered or bulleted step sequence, a runbook or procedure block, or a command caption resolves to **strict** mode: 20-word sentence cap, full rule set. Everything else — narrative rationale, executive summary, due-diligence prose — resolves to **STE-flavored**: 25-word cap, relaxed dictionary, same active-voice and no-phrasal-verb discipline. This is mechanically detectable from list markers and fenced/table syntax, and it handles a single document that mixes both, which most of them do.

## Hard categories — apply in every mode, always block

Banned marketing adjectives · banned phrasal verbs · semicolons · banned filler and hedge phrases (minus the carve-out below) · sentence-length cap for the resolved mode · paragraph cap (more than six sentences). Full lists: `references/prose-style-and-vocabulary.md`.

## Advisory categories — reported, never block

Passive voice, nominalization, em-dash count. Both grammar checks are heuristic and false-positive-prone — this plugin's own reference docs use em-dashes constitutively, so em-dash must never gate release. Advisory hits feed `dossier-pass-c-audience`'s Step 5 and the scorecard's Dimension 10 as context for whether a hard-category fix was genuine or superficial, not as findings on their own.

## The epistemic-hedging carve-out

`references/source-authority-and-claim-states.md` requires specific hedge phrases as structural markers for `Interpreted`-state claims: `"This suggests…"`, `"The most likely reading is…"`, `"Inferred from…"`, and the prose prefixes `Inferred:`, `Unknown:`, `Recommendation:`. These are the package's epistemic-honesty mechanism, not slop. A line opening with one of these markers is exempt from the hedge-phrase and sentence-length rules, in every mode, unconditionally. The full allow-list lives in `references/prose-style-and-vocabulary.md` — check it before ever touching a flagged hedge.

## Precedence

Iron Rules from other skills always outrank this one. Where this skill and `disclosure-gating` appear to disagree on a public document, `disclosure-gating` wins — truth and legal exposure over style. Cross-document synonym rotation stays `project-modeling`'s and `dossier-pass-c-audience`'s job; this skill only concerns the document currently being drafted.

## Self-lint procedure

Run `bin/dossier-prose-lint.sh --file <path> --json`. Fix hard-category hits with `Edit`. Re-run, capped at two revision passes — the same per-document budget discipline `dossier-doc-drafter` already applies to context. Return the draft regardless of outcome after the cap, and if a violation remains, say so plainly and name the Iron Rule that made the fix impossible (a required citation, a required hedge, an owner name that cannot shrink).

## Output Format

```markdown
PROSE_LINT=clean
```
or
```markdown
PROSE_LINT={n} hard violation(s) after revision
| Category | Count | Location |
|---|---|---|
```

## Rationalization Prevention

| Excuse | Response |
|--------|----------|
| "This adjective is technically accurate" | Concreteness is the axis, not accuracy. Replace it with the fact it stands in for. |
| "Hedging shows appropriate humility" | Only for an `I`-state claim carrying the required marker. Elsewhere a hedge is a claim you will not make — write `Unknown:`, or make the claim. |
| "Passive voice reads more objective here" | It hides the actor a reader needs. Name them, or state that the actor is unknown. |
| "The sentence is long because the fact is complex" | Split it. Two 15-word sentences carry the same fact and cost a reader less. |
| "The linter flagged my required hedge" | Check the allow-list in `references/prose-style-and-vocabulary.md` first. This is very likely a linter defect, not a prose defect. |
| "It's just one em-dash" | Em-dash is advisory, not blocking. Do not spend a revision pass chasing it. |

## Integration

Loaded by `dossier-doc-drafter` for every document; it self-lints before returning a draft. Invoked independently by `dossier-pass-c-audience` in Step 5 — it re-runs the script itself rather than reading the drafter's self-report, preserving the Independence Protocol. Invoked by `bin/dossier-gate.sh` for release-gate condition `G18`, across the whole package. Referenced by Dimension 10 of `references/scorecard-rubric.md` for the advisory-category judgment context.

References: `references/prose-style-and-vocabulary.md`, `references/source-authority-and-claim-states.md`, `references/release-gate-conditions.md`.